Merge pull request #140 from WebAssembly/alignment

Latest LLVM + http://reviews.llvm.org/D16534

64 files changed, 4604 insertions, 2812 deletions

diff --git a/src/s2wasm.h b/src/s2wasm.h
index 8b34d6b03..13f424f7c 100644
--- a/src/s2wasm.h
+++ b/src/s2wasm.h
@@ -159,6 +159,7 @@ private:
     return cashew::IString(str.c_str(), false);
   }
 
+  // get an int
   int32_t getInt() {
     const char* loc = s;
     uint32_t value = 0;
@@ -192,6 +193,15 @@ private:
     return value;
   }
 
+  // get an int from an arbitrary string, with our full error handling
+  int32_t getInt(const char *from) {
+    const char *before = s;
+    s = from;
+    auto ret = getInt();
+    s = before;
+    return ret;
+  }
+
   // gets a constant, which may be a relocation for later.
   // returns whether this is a relocation
   bool getConst(uint32_t* target) {
@@ -492,6 +502,10 @@ private:
           inputs[i] = curr;
         }
         if (*s == ')') s++; // tolerate 0(argument) syntax, where we started at the 'a'
+        if (*s == ':') { // tolerate :attribute=value syntax (see getAttributes)
+          s++;
+          skipToSep();
+        }
         if (i < num - 1) skipComma();
       }
       for (int i = num-1; i >= 0; i--) {
@@ -515,6 +529,24 @@ private:
         addToBlock(set);
       }
     };
+    auto getAttributes = [&](int num) {
+      const char *before = s;
+      std::vector<const char*> attributes; // TODO: optimize (if .s format doesn't change)
+      attributes.resize(num);
+      for (int i = 0; i < num; i++) {
+        skipToSep();
+        if (*s == ')') s++; // tolerate 0(argument) syntax, where we started at the 'a'
+        if (*s == ':') {
+          attributes[i] = s + 1;
+        } else {
+          attributes[i] = nullptr;
+        }
+        if (i < num - 1) skipComma();
+      }
+      s = before;
+      return attributes;
+    };
+    //
     auto makeBinary = [&](BinaryOp op, WasmType type) {
       Name assign = getAssign();
       skipComma();
@@ -559,9 +591,14 @@ private:
       match("_u");
       Name assign = getAssign();
       getConst(&curr->offset);
-      curr->align = curr->bytes; // XXX
       mustMatch("(");
+      auto attributes = getAttributes(1);
       curr->ptr = getInput();
+      curr->align = curr->bytes;
+      if (attributes[0]) {
+        assert(strncmp(attributes[0], "p2align=", 8) == 0);
+        curr->align = pow(2, getInt(attributes[0] + 8));
+      }
       setOutput(curr, assign);
     };
     auto makeStore = [&](WasmType type) {
@@ -570,12 +607,17 @@ private:
       curr->type = type;
       int32_t bytes = getInt();
       curr->bytes = bytes > 0 ? bytes : getWasmTypeSize(type);
-      curr->align = curr->bytes; // XXX
       Name assign = getAssign();
       getConst(&curr->offset);
       mustMatch("(");
+      auto attributes = getAttributes(2);
       auto inputs = getInputs(2);
       curr->ptr = inputs[0];
+      curr->align = curr->bytes;
+      if (attributes[0]) {
+        assert(strncmp(attributes[0], "p2align=", 8) == 0);
+        curr->align = pow(2, getInt(attributes[0] + 8));
+      }
       curr->value = inputs[1];
       setOutput(curr, assign);
     };
diff --git a/test/llvm_autogenerated/call.s b/test/llvm_autogenerated/call.s
index 7081c43a8..52f99b2cf 100644
--- a/test/llvm_autogenerated/call.s
+++ b/test/llvm_autogenerated/call.s
@@ -6,6 +6,7 @@ call_i32_nullary:
 	.result 	i32
 	i32.call	$push0=, i32_nullary@FUNCTION
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	call_i32_nullary, .Lfunc_end0-call_i32_nullary
 
@@ -15,6 +16,7 @@ call_i64_nullary:
 	.result 	i64
 	i64.call	$push0=, i64_nullary@FUNCTION
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	call_i64_nullary, .Lfunc_end1-call_i64_nullary
 
@@ -24,6 +26,7 @@ call_float_nullary:
 	.result 	f32
 	f32.call	$push0=, float_nullary@FUNCTION
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	call_float_nullary, .Lfunc_end2-call_float_nullary
 
@@ -33,6 +36,7 @@ call_double_nullary:
 	.result 	f64
 	f64.call	$push0=, double_nullary@FUNCTION
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	call_double_nullary, .Lfunc_end3-call_double_nullary
 
@@ -41,6 +45,7 @@ call_double_nullary:
 call_void_nullary:
 	call    	void_nullary@FUNCTION
 	return
+	.endfunc
 .Lfunc_end4:
 	.size	call_void_nullary, .Lfunc_end4-call_void_nullary
 
@@ -51,6 +56,7 @@ call_i32_unary:
 	.result 	i32
 	i32.call	$push0=, i32_unary@FUNCTION, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	call_i32_unary, .Lfunc_end5-call_i32_unary
 
@@ -61,6 +67,7 @@ call_i32_binary:
 	.result 	i32
 	i32.call	$push0=, i32_binary@FUNCTION, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	call_i32_binary, .Lfunc_end6-call_i32_binary
 
@@ -70,6 +77,7 @@ call_indirect_void:
 	.param  	i32
 	call_indirect	$0
 	return
+	.endfunc
 .Lfunc_end7:
 	.size	call_indirect_void, .Lfunc_end7-call_indirect_void
 
@@ -80,6 +88,7 @@ call_indirect_i32:
 	.result 	i32
 	i32.call_indirect	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	call_indirect_i32, .Lfunc_end8-call_indirect_i32
 
@@ -88,6 +97,7 @@ call_indirect_i32:
 tail_call_void_nullary:
 	call    	void_nullary@FUNCTION
 	return
+	.endfunc
 .Lfunc_end9:
 	.size	tail_call_void_nullary, .Lfunc_end9-tail_call_void_nullary
 
@@ -96,6 +106,7 @@ tail_call_void_nullary:
 fastcc_tail_call_void_nullary:
 	call    	void_nullary@FUNCTION
 	return
+	.endfunc
 .Lfunc_end10:
 	.size	fastcc_tail_call_void_nullary, .Lfunc_end10-fastcc_tail_call_void_nullary
 
@@ -104,8 +115,8 @@ fastcc_tail_call_void_nullary:
 coldcc_tail_call_void_nullary:
 	call    	void_nullary@FUNCTION
 	return
+	.endfunc
 .Lfunc_end11:
 	.size	coldcc_tail_call_void_nullary, .Lfunc_end11-coldcc_tail_call_void_nullary
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/cfg-stackify.s b/test/llvm_autogenerated/cfg-stackify.s
index fcb9652d9..584c02adc 100644
--- a/test/llvm_autogenerated/cfg-stackify.s
+++ b/test/llvm_autogenerated/cfg-stackify.s
@@ -8,15 +8,16 @@ test0:
 	i32.const	$1=, 0
 .LBB0_1:
 	loop
-	i32.const	$push0=, 1
-	i32.add 	$1=, $1, $pop0
-	i32.ge_s	$push1=, $1, $0
-	br_if   	$pop1, 1
+	i32.const	$push1=, 1
+	i32.add 	$1=, $1, $pop1
+	i32.ge_s	$push0=, $1, $0
+	br_if   	$pop0, 1
 	call    	something@FUNCTION
 	br      	0
 .LBB0_3:
 	end_loop
 	return
+	.endfunc
 .Lfunc_end0:
 	.size	test0, .Lfunc_end0-test0
 
@@ -28,15 +29,16 @@ test1:
 	i32.const	$1=, 0
 .LBB1_1:
 	loop
-	i32.const	$push0=, 1
-	i32.add 	$1=, $1, $pop0
-	i32.ge_s	$push1=, $1, $0
-	br_if   	$pop1, 1
+	i32.const	$push1=, 1
+	i32.add 	$1=, $1, $pop1
+	i32.ge_s	$push0=, $1, $0
+	br_if   	$pop0, 1
 	call    	something@FUNCTION
 	br      	0
 .LBB1_3:
 	end_loop
 	return
+	.endfunc
 .Lfunc_end1:
 	.size	test1, .Lfunc_end1-test1
 
@@ -63,6 +65,7 @@ test2:
 	end_loop
 	end_block
 	return
+	.endfunc
 .Lfunc_end2:
 	.size	test2, .Lfunc_end2-test2
 
@@ -71,11 +74,10 @@ test2:
 doublediamond:
 	.param  	i32, i32, i32
 	.result 	i32
-	.local  	i32
 	block
 	block
 	i32.const	$push0=, 0
-	i32.store	$3=, 0($2), $pop0
+	i32.store	$discard=, 0($2), $pop0
 	br_if   	$0, 0
 	i32.const	$push4=, 1
 	i32.store	$discard=, 0($2), $pop4
@@ -97,7 +99,9 @@ doublediamond:
 	end_block
 	i32.const	$push5=, 5
 	i32.store	$discard=, 0($2), $pop5
-	return  	$3
+	i32.const	$push6=, 0
+	return  	$pop6
+	.endfunc
 .Lfunc_end3:
 	.size	doublediamond, .Lfunc_end3-doublediamond
 
@@ -118,6 +122,7 @@ triangle:
 	i32.const	$push2=, 2
 	i32.store	$discard=, 0($0), $pop2
 	return  	$2
+	.endfunc
 .Lfunc_end4:
 	.size	triangle, .Lfunc_end4-triangle
 
@@ -126,11 +131,10 @@ triangle:
 diamond:
 	.param  	i32, i32
 	.result 	i32
-	.local  	i32
 	block
 	block
 	i32.const	$push0=, 0
-	i32.store	$2=, 0($0), $pop0
+	i32.store	$discard=, 0($0), $pop0
 	br_if   	$1, 0
 	i32.const	$push2=, 1
 	i32.store	$discard=, 0($0), $pop2
@@ -143,7 +147,9 @@ diamond:
 	end_block
 	i32.const	$push3=, 3
 	i32.store	$discard=, 0($0), $pop3
-	return  	$2
+	i32.const	$push4=, 0
+	return  	$pop4
+	.endfunc
 .Lfunc_end5:
 	.size	diamond, .Lfunc_end5-diamond
 
@@ -155,6 +161,7 @@ single_block:
 	i32.const	$push0=, 0
 	i32.store	$push1=, 0($0), $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end6:
 	.size	single_block, .Lfunc_end6-single_block
 
@@ -172,6 +179,7 @@ minimal_loop:
 	br      	0
 .LBB7_2:
 	end_loop
+	.endfunc
 .Lfunc_end7:
 	.size	minimal_loop, .Lfunc_end7-minimal_loop
 
@@ -180,20 +188,21 @@ minimal_loop:
 simple_loop:
 	.param  	i32, i32
 	.result 	i32
-	.local  	i32
 	i32.const	$push0=, 0
-	i32.store	$2=, 0($0), $pop0
+	i32.store	$discard=, 0($0), $pop0
 .LBB8_1:
 	loop
-	i32.const	$push1=, 1
-	i32.store	$discard=, 0($0), $pop1
-	i32.const	$push3=, 0
-	i32.eq  	$push4=, $1, $pop3
-	br_if   	$pop4, 0
+	i32.const	$push3=, 1
+	i32.store	$discard=, 0($0), $pop3
+	i32.const	$push4=, 0
+	i32.eq  	$push5=, $1, $pop4
+	br_if   	$pop5, 0
 	end_loop
-	i32.const	$push2=, 2
-	i32.store	$discard=, 0($0), $pop2
-	return  	$2
+	i32.const	$push1=, 2
+	i32.store	$discard=, 0($0), $pop1
+	i32.const	$push2=, 0
+	return  	$pop2
+	.endfunc
 .Lfunc_end8:
 	.size	simple_loop, .Lfunc_end8-simple_loop
 
@@ -222,6 +231,7 @@ doubletriangle:
 	i32.const	$push4=, 5
 	i32.store	$discard=, 0($2), $pop4
 	return  	$3
+	.endfunc
 .Lfunc_end9:
 	.size	doubletriangle, .Lfunc_end9-doubletriangle
 
@@ -230,11 +240,10 @@ doubletriangle:
 ifelse_earlyexits:
 	.param  	i32, i32, i32
 	.result 	i32
-	.local  	i32
 	block
 	block
 	i32.const	$push0=, 0
-	i32.store	$3=, 0($2), $pop0
+	i32.store	$discard=, 0($2), $pop0
 	br_if   	$0, 0
 	i32.const	$push3=, 1
 	i32.store	$discard=, 0($2), $pop3
@@ -250,7 +259,9 @@ ifelse_earlyexits:
 	end_block
 	i32.const	$push4=, 4
 	i32.store	$discard=, 0($2), $pop4
-	return  	$3
+	i32.const	$push5=, 0
+	return  	$pop5
+	.endfunc
 .Lfunc_end10:
 	.size	ifelse_earlyexits, .Lfunc_end10-ifelse_earlyexits
 
@@ -266,22 +277,22 @@ doublediamond_in_a_loop:
 	i32.const	$push0=, 0
 	i32.store	$discard=, 0($2), $pop0
 	br_if   	$0, 0
-	i32.const	$push4=, 1
-	i32.store	$discard=, 0($2), $pop4
+	i32.const	$push1=, 1
+	i32.store	$discard=, 0($2), $pop1
 	br      	1
 .LBB11_3:
 	end_block
 	block
-	i32.const	$push1=, 2
-	i32.store	$discard=, 0($2), $pop1
+	i32.const	$push2=, 2
+	i32.store	$discard=, 0($2), $pop2
 	br_if   	$1, 0
 	i32.const	$push3=, 3
 	i32.store	$discard=, 0($2), $pop3
 	br      	1
 .LBB11_5:
 	end_block
-	i32.const	$push2=, 4
-	i32.store	$discard=, 0($2), $pop2
+	i32.const	$push4=, 4
+	i32.store	$discard=, 0($2), $pop4
 .LBB11_6:
 	end_block
 	i32.const	$push5=, 5
@@ -289,6 +300,7 @@ doublediamond_in_a_loop:
 	br      	0
 .LBB11_7:
 	end_loop
+	.endfunc
 .Lfunc_end11:
 	.size	doublediamond_in_a_loop, .Lfunc_end11-doublediamond_in_a_loop
 
@@ -315,6 +327,7 @@ test3:
 .LBB12_5:
 	end_block
 	return
+	.endfunc
 .Lfunc_end12:
 	.size	test3, .Lfunc_end12-test3
 
@@ -353,6 +366,7 @@ test4:
 .LBB13_8:
 	end_block
 	return
+	.endfunc
 .Lfunc_end13:
 	.size	test4, .Lfunc_end13-test4
 
@@ -360,29 +374,35 @@ test4:
 	.type	test5,@function
 test5:
 	.param  	i32, i32
-	.local  	i32, i32
+	.local  	i32
+	i32.const	$push5=, 1
+	i32.and 	$0=, $0, $pop5
+	i32.const	$push4=, 1
+	i32.and 	$2=, $1, $pop4
 .LBB14_1:
 	block
 	loop
-	i32.const	$2=, 0
-	i32.store	$3=, 0($2), $2
-	i32.const	$2=, 1
-	i32.and 	$push0=, $0, $2
-	i32.const	$push5=, 0
-	i32.eq  	$push6=, $pop0, $pop5
-	br_if   	$pop6, 2
-	i32.store	$push2=, 0($3), $2
-	i32.and 	$push3=, $1, $pop2
-	br_if   	$pop3, 0
+	i32.const	$push7=, 0
+	i32.const	$push6=, 0
+	i32.store	$1=, 0($pop7), $pop6
+	i32.const	$push9=, 0
+	i32.eq  	$push10=, $0, $pop9
+	br_if   	$pop10, 2
+	i32.const	$push8=, 1
+	i32.store	$discard=, 0($1), $pop8
+	br_if   	$2, 0
 	end_loop
-	i32.const	$push4=, 3
-	i32.store	$discard=, 0($3), $pop4
+	i32.const	$push2=, 0
+	i32.const	$push3=, 3
+	i32.store	$discard=, 0($pop2), $pop3
 	return
 .LBB14_4:
 	end_block
+	i32.const	$push0=, 0
 	i32.const	$push1=, 2
-	i32.store	$discard=, 0($3), $pop1
+	i32.store	$discard=, 0($pop0), $pop1
 	return
+	.endfunc
 .Lfunc_end14:
 	.size	test5, .Lfunc_end14-test5
 
@@ -391,37 +411,43 @@ test5:
 test6:
 	.param  	i32, i32
 	.local  	i32, i32, i32
+	i32.const	$push6=, 1
+	i32.and 	$2=, $0, $pop6
 .LBB15_1:
 	block
 	block
 	loop
-	i32.const	$2=, 0
-	i32.store	$discard=, 0($2), $2
-	i32.const	$3=, 1
-	i32.and 	$push0=, $0, $3
-	i32.const	$push4=, 0
-	i32.eq  	$push5=, $pop0, $pop4
-	br_if   	$pop5, 3
-	i32.store	$discard=, 0($2), $3
+	i32.const	$push8=, 0
+	i32.const	$push7=, 0
+	i32.store	$0=, 0($pop8), $pop7
+	i32.const	$push10=, 0
+	i32.eq  	$push11=, $2, $pop10
+	br_if   	$pop11, 3
+	i32.const	$push9=, 1
+	i32.store	$3=, 0($0), $pop9
 	i32.and 	$4=, $1, $3
-	i32.const	$push6=, 0
-	i32.eq  	$push7=, $4, $pop6
-	br_if   	$pop7, 2
-	i32.store	$discard=, 0($2), $3
+	i32.const	$push12=, 0
+	i32.eq  	$push13=, $4, $pop12
+	br_if   	$pop13, 2
+	i32.store	$discard=, 0($0), $3
 	br_if   	$4, 0
 	end_loop
-	i32.const	$push3=, 2
-	i32.store	$discard=, 0($2), $pop3
+	i32.const	$push4=, 0
+	i32.const	$push5=, 2
+	i32.store	$discard=, 0($pop4), $pop5
 	return
 .LBB15_5:
 	end_block
+	i32.const	$push0=, 0
 	i32.const	$push1=, 3
-	i32.store	$discard=, 0($2), $pop1
+	i32.store	$discard=, 0($pop0), $pop1
 .LBB15_6:
 	end_block
-	i32.const	$push2=, 4
-	i32.store	$discard=, 0($2), $pop2
+	i32.const	$push2=, 0
+	i32.const	$push3=, 4
+	i32.store	$discard=, 0($pop2), $pop3
 	return
+	.endfunc
 .Lfunc_end15:
 	.size	test6, .Lfunc_end15-test6
 
@@ -430,32 +456,37 @@ test6:
 test7:
 	.param  	i32, i32
 	.local  	i32, i32
-	i32.const	$3=, 0
-	i32.store	$2=, 0($3), $3
+	i32.const	$push0=, 0
+	i32.const	$push8=, 0
+	i32.store	$2=, 0($pop0), $pop8
+	i32.const	$push7=, 1
+	i32.and 	$3=, $0, $pop7
 .LBB16_1:
 	loop
 	block
-	i32.const	$push0=, 1
-	i32.store	$3=, 0($2), $pop0
-	i32.and 	$push1=, $0, $3
-	br_if   	$pop1, 0
-	i32.const	$push2=, 2
-	i32.store	$discard=, 0($2), $pop2
-	i32.and 	$push3=, $1, $3
-	br_if   	$pop3, 1
-	i32.const	$push4=, 4
-	i32.store	$discard=, 0($2), $pop4
+	i32.const	$push9=, 1
+	i32.store	$0=, 0($2), $pop9
+	br_if   	$3, 0
+	i32.const	$push10=, 2
+	i32.store	$discard=, 0($2), $pop10
+	i32.and 	$push1=, $1, $0
+	br_if   	$pop1, 1
+	i32.const	$push2=, 0
+	i32.const	$push3=, 4
+	i32.store	$discard=, 0($pop2), $pop3
 	unreachable
 .LBB16_4:
 	end_block
-	i32.const	$push5=, 3
-	i32.store	$discard=, 0($2), $pop5
-	i32.and 	$push6=, $1, $3
-	br_if   	$pop6, 0
+	i32.const	$push11=, 3
+	i32.store	$discard=, 0($2), $pop11
+	i32.and 	$push4=, $1, $0
+	br_if   	$pop4, 0
 	end_loop
-	i32.const	$push7=, 5
-	i32.store	$discard=, 0($2), $pop7
+	i32.const	$push5=, 0
+	i32.const	$push6=, 5
+	i32.store	$discard=, 0($pop5), $pop6
 	unreachable
+	.endfunc
 .Lfunc_end16:
 	.size	test7, .Lfunc_end16-test7
 
@@ -463,25 +494,27 @@ test7:
 	.type	test8,@function
 test8:
 	.result 	i32
-	.local  	i32
-	i32.const	$0=, 0
 .LBB17_1:
 	loop
 	block
 	i32.const	$push0=, 0
-	i32.eq  	$push1=, $0, $pop0
-	br_if   	$pop1, 0
+	i32.const	$push3=, 0
+	i32.eq  	$push4=, $pop0, $pop3
+	br_if   	$pop4, 0
 	i32.const	$push2=, 0
-	i32.eq  	$push3=, $0, $pop2
-	br_if   	$pop3, 1
+	i32.const	$push5=, 0
+	i32.eq  	$push6=, $pop2, $pop5
+	br_if   	$pop6, 1
 .LBB17_3:
 	end_block
 	loop
-	br_if   	$0, 0
+	i32.const	$push1=, 0
+	br_if   	$pop1, 0
 	br      	2
 .LBB17_4:
 	end_loop
 	end_loop
+	.endfunc
 .Lfunc_end17:
 	.size	test8, .Lfunc_end17-test8
 
@@ -489,47 +522,50 @@ test8:
 	.type	test9,@function
 test9:
 	.local  	i32, i32
-	i32.const	$1=, 0
-	i32.store	$0=, 0($1), $1
+	i32.const	$push0=, 0
+	i32.const	$push11=, 0
+	i32.store	$0=, 0($pop0), $pop11
 .LBB18_1:
 	loop
-	i32.const	$push0=, 1
-	i32.store	$1=, 0($0), $pop0
+	i32.const	$push12=, 1
+	i32.store	$1=, 0($0), $pop12
 	i32.call	$push1=, a@FUNCTION
 	i32.and 	$push2=, $pop1, $1
-	i32.const	$push13=, 0
-	i32.eq  	$push14=, $pop2, $pop13
-	br_if   	$pop14, 1
+	i32.const	$push16=, 0
+	i32.eq  	$push17=, $pop2, $pop16
+	br_if   	$pop17, 1
 .LBB18_2:
 	loop
 	block
-	i32.const	$push4=, 2
-	i32.store	$discard=, 0($0), $pop4
+	i32.const	$push13=, 2
+	i32.store	$discard=, 0($0), $pop13
 	i32.call	$push5=, a@FUNCTION
 	i32.and 	$push6=, $pop5, $1
-	i32.const	$push15=, 0
-	i32.eq  	$push16=, $pop6, $pop15
-	br_if   	$pop16, 0
-	i32.const	$push10=, 3
-	i32.store	$discard=, 0($0), $pop10
-	i32.call	$push11=, a@FUNCTION
-	i32.and 	$push12=, $pop11, $1
-	br_if   	$pop12, 1
+	i32.const	$push18=, 0
+	i32.eq  	$push19=, $pop6, $pop18
+	br_if   	$pop19, 0
+	i32.const	$push14=, 3
+	i32.store	$discard=, 0($0), $pop14
+	i32.call	$push9=, a@FUNCTION
+	i32.and 	$push10=, $pop9, $1
+	br_if   	$pop10, 1
 	br      	3
 .LBB18_4:
 	end_block
-	i32.const	$push7=, 4
-	i32.store	$discard=, 0($0), $pop7
-	i32.call	$push8=, a@FUNCTION
-	i32.and 	$push9=, $pop8, $1
-	br_if   	$pop9, 0
+	i32.const	$push15=, 4
+	i32.store	$discard=, 0($0), $pop15
+	i32.call	$push7=, a@FUNCTION
+	i32.and 	$push8=, $pop7, $1
+	br_if   	$pop8, 0
 	br      	2
 .LBB18_5:
 	end_loop
 	end_loop
-	i32.const	$push3=, 5
-	i32.store	$discard=, 0($0), $pop3
+	i32.const	$push3=, 0
+	i32.const	$push4=, 5
+	i32.store	$discard=, 0($pop3), $pop4
 	return
+	.endfunc
 .Lfunc_end18:
 	.size	test9, .Lfunc_end18-test9
 
@@ -554,9 +590,9 @@ test10:
 .LBB19_3:
 	loop
 	copy_local	$2=, $4
-	i32.const	$push0=, 4
-	i32.gt_u	$push1=, $2, $pop0
-	br_if   	$pop1, 5
+	i32.const	$push1=, 4
+	i32.gt_u	$push0=, $2, $pop1
+	br_if   	$pop0, 5
 	copy_local	$4=, $3
 	tableswitch	$2, 0, 0, 1, 5, 2, 4
 .LBB19_5:
@@ -569,6 +605,7 @@ test10:
 	br      	0
 .LBB19_7:
 	end_loop
+	.endfunc
 .Lfunc_end19:
 	.size	test10, .Lfunc_end19-test10
 
@@ -576,50 +613,60 @@ test10:
 	.type	test11,@function
 test11:
 	.local  	i32
-	i32.const	$0=, 0
-	i32.store	$discard=, 0($0), $0
 	block
 	block
 	block
 	block
+	i32.const	$push0=, 0
+	i32.const	$push13=, 0
+	i32.store	$0=, 0($pop0), $pop13
 	br_if   	$0, 0
 	block
-	i32.const	$push4=, 1
-	i32.store	$discard=, 0($0), $pop4
+	i32.const	$push6=, 1
+	i32.store	$discard=, 0($0), $pop6
 	br_if   	$0, 0
-	i32.const	$push5=, 2
-	i32.store	$discard=, 0($0), $pop5
-	br_if   	$0, 2
+	i32.const	$push8=, 0
+	i32.const	$push7=, 2
+	i32.store	$discard=, 0($pop8), $pop7
+	i32.const	$push14=, 0
+	br_if   	$pop14, 2
 .LBB20_3:
 	end_block
-	i32.const	$push7=, 3
-	i32.store	$discard=, 0($0), $pop7
+	i32.const	$push11=, 0
+	i32.const	$push12=, 3
+	i32.store	$discard=, 0($pop11), $pop12
 	return
 .LBB20_4:
 	end_block
-	i32.const	$push0=, 4
-	i32.store	$discard=, 0($0), $pop0
-	br_if   	$0, 2
-	i32.const	$push1=, 5
+	i32.const	$push1=, 4
 	i32.store	$discard=, 0($0), $pop1
-	i32.const	$push8=, 0
-	i32.eq  	$push9=, $0, $pop8
-	br_if   	$pop9, 1
+	br_if   	$0, 2
+	i32.const	$push16=, 0
+	i32.const	$push2=, 5
+	i32.store	$discard=, 0($pop16), $pop2
+	i32.const	$push15=, 0
+	i32.const	$push18=, 0
+	i32.eq  	$push19=, $pop15, $pop18
+	br_if   	$pop19, 1
 .LBB20_6:
 	end_block
-	i32.const	$push6=, 7
-	i32.store	$discard=, 0($0), $pop6
+	i32.const	$push9=, 0
+	i32.const	$push10=, 7
+	i32.store	$discard=, 0($pop9), $pop10
 	return
 .LBB20_7:
 	end_block
-	i32.const	$push2=, 6
-	i32.store	$discard=, 0($0), $pop2
+	i32.const	$push17=, 0
+	i32.const	$push3=, 6
+	i32.store	$discard=, 0($pop17), $pop3
 	return
 .LBB20_8:
 	end_block
-	i32.const	$push3=, 8
-	i32.store	$discard=, 0($0), $pop3
+	i32.const	$push4=, 0
+	i32.const	$push5=, 8
+	i32.store	$discard=, 0($pop4), $pop5
 	return
+	.endfunc
 .Lfunc_end20:
 	.size	test11, .Lfunc_end20-test11
 
@@ -634,24 +681,24 @@ test12:
 	block
 	block
 	block
-	i32.const	$push0=, 103
-	i32.gt_s	$push1=, $1, $pop0
-	br_if   	$pop1, 0
+	i32.const	$push5=, 103
+	i32.gt_s	$push0=, $1, $pop5
+	br_if   	$pop0, 0
 	i32.const	$push6=, 42
-	i32.eq  	$push7=, $1, $pop6
-	br_if   	$pop7, 2
-	i32.const	$push8=, 76
-	i32.eq  	$push9=, $1, $pop8
-	br_if   	$pop9, 2
+	i32.eq  	$push3=, $1, $pop6
+	br_if   	$pop3, 2
+	i32.const	$push7=, 76
+	i32.eq  	$push4=, $1, $pop7
+	br_if   	$pop4, 2
 	br      	1
 .LBB21_4:
 	end_block
-	i32.const	$push2=, 108
-	i32.eq  	$push3=, $1, $pop2
-	br_if   	$pop3, 1
-	i32.const	$push4=, 104
-	i32.eq  	$push5=, $1, $pop4
-	br_if   	$pop5, 1
+	i32.const	$push8=, 108
+	i32.eq  	$push1=, $1, $pop8
+	br_if   	$pop1, 1
+	i32.const	$push9=, 104
+	i32.eq  	$push2=, $1, $pop9
+	br_if   	$pop2, 1
 .LBB21_6:
 	end_block
 	return
@@ -662,6 +709,7 @@ test12:
 	br      	0
 .LBB21_8:
 	end_loop
+	.endfunc
 .Lfunc_end21:
 	.size	test12, .Lfunc_end21-test12
 
@@ -691,25 +739,27 @@ test13:
 	br_if   	$pop6, 0
 	end_block
 	unreachable
+	.endfunc
 .Lfunc_end22:
 	.size	test13, .Lfunc_end22-test13
 
 	.globl	test14
 	.type	test14,@function
 test14:
-	.local  	i32
-	i32.const	$0=, 0
 .LBB23_1:
 	loop
-	br_if   	$0, 0
+	i32.const	$push0=, 0
+	br_if   	$pop0, 0
 .LBB23_2:
 	end_loop
 	loop
-	br_if   	$0, 0
+	i32.const	$discard=, 0
+	i32.const	$push1=, 0
+	br_if   	$pop1, 0
 	end_loop
 	return
+	.endfunc
 .Lfunc_end23:
 	.size	test14, .Lfunc_end23-test14
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/cfg-stackify.wast b/test/llvm_autogenerated/cfg-stackify.wast
index b74e242d5..c3b2cef82 100644
--- a/test/llvm_autogenerated/cfg-stackify.wast
+++ b/test/llvm_autogenerated/cfg-stackify.wast
@@ -129,16 +129,13 @@
     )
   )
   (func $doublediamond (param $$0 i32) (param $$1 i32) (param $$2 i32) (result i32)
-    (local $$3 i32)
     (block $fake_return_waka123
       (block
         (block $label$0
           (block $label$1
-            (set_local $$3
-              (i32.store align=4
-                (get_local $$2)
-                (i32.const 0)
-              )
+            (i32.store align=4
+              (get_local $$2)
+              (i32.const 0)
             )
             (br_if
               (get_local $$0)
@@ -175,7 +172,7 @@
           (i32.const 5)
         )
         (br $fake_return_waka123
-          (get_local $$3)
+          (i32.const 0)
         )
       )
     )
@@ -211,16 +208,13 @@
     )
   )
   (func $diamond (param $$0 i32) (param $$1 i32) (result i32)
-    (local $$2 i32)
     (block $fake_return_waka123
       (block
         (block $label$0
           (block $label$1
-            (set_local $$2
-              (i32.store align=4
-                (get_local $$0)
-                (i32.const 0)
-              )
+            (i32.store align=4
+              (get_local $$0)
+              (i32.const 0)
             )
             (br_if
               (get_local $$1)
@@ -242,7 +236,7 @@
           (i32.const 3)
         )
         (br $fake_return_waka123
-          (get_local $$2)
+          (i32.const 0)
         )
       )
     )
@@ -273,14 +267,11 @@
     )
   )
   (func $simple_loop (param $$0 i32) (param $$1 i32) (result i32)
-    (local $$2 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$2
-          (i32.store align=4
-            (get_local $$0)
-            (i32.const 0)
-          )
+        (i32.store align=4
+          (get_local $$0)
+          (i32.const 0)
         )
         (loop $label$1 $label$0
           (i32.store align=4
@@ -300,7 +291,7 @@
           (i32.const 2)
         )
         (br $fake_return_waka123
-          (get_local $$2)
+          (i32.const 0)
         )
       )
     )
@@ -350,16 +341,13 @@
     )
   )
   (func $ifelse_earlyexits (param $$0 i32) (param $$1 i32) (param $$2 i32) (result i32)
-    (local $$3 i32)
     (block $fake_return_waka123
       (block
         (block $label$0
           (block $label$1
-            (set_local $$3
-              (i32.store align=4
-                (get_local $$2)
-                (i32.const 0)
-              )
+            (i32.store align=4
+              (get_local $$2)
+              (i32.const 0)
             )
             (br_if
               (get_local $$0)
@@ -389,7 +377,7 @@
           (i32.const 4)
         )
         (br $fake_return_waka123
-          (get_local $$3)
+          (i32.const 0)
         )
       )
     )
@@ -525,52 +513,52 @@
   )
   (func $test5 (param $$0 i32) (param $$1 i32)
     (local $$2 i32)
-    (local $$3 i32)
     (block $fake_return_waka123
       (block
+        (set_local $$0
+          (i32.and
+            (get_local $$0)
+            (i32.const 1)
+          )
+        )
+        (set_local $$2
+          (i32.and
+            (get_local $$1)
+            (i32.const 1)
+          )
+        )
         (block $label$0
           (loop $label$2 $label$1
-            (set_local $$2
-              (i32.const 0)
-            )
-            (set_local $$3
+            (set_local $$1
               (i32.store align=4
-                (get_local $$2)
-                (get_local $$2)
+                (i32.const 0)
+                (i32.const 0)
               )
             )
-            (set_local $$2
-              (i32.const 1)
-            )
             (br_if
               (i32.eq
-                (i32.and
-                  (get_local $$0)
-                  (get_local $$2)
-                )
+                (get_local $$0)
                 (i32.const 0)
               )
               $label$0
             )
+            (i32.store align=4
+              (get_local $$1)
+              (i32.const 1)
+            )
             (br_if
-              (i32.and
-                (get_local $$1)
-                (i32.store align=4
-                  (get_local $$3)
-                  (get_local $$2)
-                )
-              )
+              (get_local $$2)
               $label$1
             )
           )
           (i32.store align=4
-            (get_local $$3)
+            (i32.const 0)
             (i32.const 3)
           )
           (br $fake_return_waka123)
         )
         (i32.store align=4
-          (get_local $$3)
+          (i32.const 0)
           (i32.const 2)
         )
         (br $fake_return_waka123)
@@ -583,32 +571,33 @@
     (local $$4 i32)
     (block $fake_return_waka123
       (block
+        (set_local $$2
+          (i32.and
+            (get_local $$0)
+            (i32.const 1)
+          )
+        )
         (block $label$0
           (block $label$1
             (loop $label$3 $label$2
-              (set_local $$2
-                (i32.const 0)
-              )
-              (i32.store align=4
-                (get_local $$2)
-                (get_local $$2)
-              )
-              (set_local $$3
-                (i32.const 1)
+              (set_local $$0
+                (i32.store align=4
+                  (i32.const 0)
+                  (i32.const 0)
+                )
               )
               (br_if
                 (i32.eq
-                  (i32.and
-                    (get_local $$0)
-                    (get_local $$3)
-                  )
+                  (get_local $$2)
                   (i32.const 0)
                 )
                 $label$0
               )
-              (i32.store align=4
-                (get_local $$2)
-                (get_local $$3)
+              (set_local $$3
+                (i32.store align=4
+                  (get_local $$0)
+                  (i32.const 1)
+                )
               )
               (set_local $$4
                 (i32.and
@@ -624,7 +613,7 @@
                 $label$1
               )
               (i32.store align=4
-                (get_local $$2)
+                (get_local $$0)
                 (get_local $$3)
               )
               (br_if
@@ -633,18 +622,18 @@
               )
             )
             (i32.store align=4
-              (get_local $$2)
+              (i32.const 0)
               (i32.const 2)
             )
             (br $fake_return_waka123)
           )
           (i32.store align=4
-            (get_local $$2)
+            (i32.const 0)
             (i32.const 3)
           )
         )
         (i32.store align=4
-          (get_local $$2)
+          (i32.const 0)
           (i32.const 4)
         )
         (br $fake_return_waka123)
@@ -654,28 +643,28 @@
   (func $test7 (param $$0 i32) (param $$1 i32)
     (local $$2 i32)
     (local $$3 i32)
-    (set_local $$3
-      (i32.const 0)
-    )
     (set_local $$2
       (i32.store align=4
-        (get_local $$3)
-        (get_local $$3)
+        (i32.const 0)
+        (i32.const 0)
+      )
+    )
+    (set_local $$3
+      (i32.and
+        (get_local $$0)
+        (i32.const 1)
       )
     )
     (loop $label$1 $label$0
       (block $label$2
-        (set_local $$3
+        (set_local $$0
           (i32.store align=4
             (get_local $$2)
             (i32.const 1)
           )
         )
         (br_if
-          (i32.and
-            (get_local $$0)
-            (get_local $$3)
-          )
+          (get_local $$3)
           $label$2
         )
         (i32.store align=4
@@ -685,12 +674,12 @@
         (br_if
           (i32.and
             (get_local $$1)
-            (get_local $$3)
+            (get_local $$0)
           )
           $label$0
         )
         (i32.store align=4
-          (get_local $$2)
+          (i32.const 0)
           (i32.const 4)
         )
         (unreachable)
@@ -702,34 +691,30 @@
       (br_if
         (i32.and
           (get_local $$1)
-          (get_local $$3)
+          (get_local $$0)
         )
         $label$0
       )
     )
     (i32.store align=4
-      (get_local $$2)
+      (i32.const 0)
       (i32.const 5)
     )
     (unreachable)
   )
   (func $test8 (result i32)
-    (local $$0 i32)
-    (set_local $$0
-      (i32.const 0)
-    )
     (loop $label$1 $label$0
       (block $label$2
         (br_if
           (i32.eq
-            (get_local $$0)
+            (i32.const 0)
             (i32.const 0)
           )
           $label$2
         )
         (br_if
           (i32.eq
-            (get_local $$0)
+            (i32.const 0)
             (i32.const 0)
           )
           $label$0
@@ -737,7 +722,7 @@
       )
       (loop $label$4 $label$3
         (br_if
-          (get_local $$0)
+          (i32.const 0)
           $label$3
         )
         (br $label$0)
@@ -749,13 +734,10 @@
     (local $$1 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$1
-          (i32.const 0)
-        )
         (set_local $$0
           (i32.store align=4
-            (get_local $$1)
-            (get_local $$1)
+            (i32.const 0)
+            (i32.const 0)
           )
         )
         (loop $label$1 $label$0
@@ -819,7 +801,7 @@
           )
         )
         (i32.store align=4
-          (get_local $$0)
+          (i32.const 0)
           (i32.const 5)
         )
         (br $fake_return_waka123)
@@ -899,17 +881,16 @@
     (local $$0 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$0
-          (i32.const 0)
-        )
-        (i32.store align=4
-          (get_local $$0)
-          (get_local $$0)
-        )
         (block $label$0
           (block $label$1
             (block $label$2
               (block $label$3
+                (set_local $$0
+                  (i32.store align=4
+                    (i32.const 0)
+                    (i32.const 0)
+                  )
+                )
                 (br_if
                   (get_local $$0)
                   $label$3
@@ -924,16 +905,16 @@
                     $label$4
                   )
                   (i32.store align=4
-                    (get_local $$0)
+                    (i32.const 0)
                     (i32.const 2)
                   )
                   (br_if
-                    (get_local $$0)
+                    (i32.const 0)
                     $label$2
                   )
                 )
                 (i32.store align=4
-                  (get_local $$0)
+                  (i32.const 0)
                   (i32.const 3)
                 )
                 (br $fake_return_waka123)
@@ -947,31 +928,31 @@
                 $label$0
               )
               (i32.store align=4
-                (get_local $$0)
+                (i32.const 0)
                 (i32.const 5)
               )
               (br_if
                 (i32.eq
-                  (get_local $$0)
+                  (i32.const 0)
                   (i32.const 0)
                 )
                 $label$1
               )
             )
             (i32.store align=4
-              (get_local $$0)
+              (i32.const 0)
               (i32.const 7)
             )
             (br $fake_return_waka123)
           )
           (i32.store align=4
-            (get_local $$0)
+            (i32.const 0)
             (i32.const 6)
           )
           (br $fake_return_waka123)
         )
         (i32.store align=4
-          (get_local $$0)
+          (i32.const 0)
           (i32.const 8)
         )
         (br $fake_return_waka123)
@@ -1085,21 +1066,18 @@
     )
   )
   (func $test14
-    (local $$0 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$0
-          (i32.const 0)
-        )
         (loop $label$1 $label$0
           (br_if
-            (get_local $$0)
+            (i32.const 0)
             $label$0
           )
         )
         (loop $label$3 $label$2
+          (i32.const 0)
           (br_if
-            (get_local $$0)
+            (i32.const 0)
             $label$2
           )
         )
diff --git a/test/llvm_autogenerated/comparisons_f32.s b/test/llvm_autogenerated/comparisons_f32.s
index a527faa3d..b14d0d8bd 100644
--- a/test/llvm_autogenerated/comparisons_f32.s
+++ b/test/llvm_autogenerated/comparisons_f32.s
@@ -9,6 +9,7 @@ ord_f32:
 	f32.eq  	$push0=, $1, $1
 	i32.and 	$push2=, $pop1, $pop0
 	return  	$pop2
+	.endfunc
 .Lfunc_end0:
 	.size	ord_f32, .Lfunc_end0-ord_f32
 
@@ -21,6 +22,7 @@ uno_f32:
 	f32.ne  	$push0=, $1, $1
 	i32.or  	$push2=, $pop1, $pop0
 	return  	$pop2
+	.endfunc
 .Lfunc_end1:
 	.size	uno_f32, .Lfunc_end1-uno_f32
 
@@ -31,6 +33,7 @@ oeq_f32:
 	.result 	i32
 	f32.eq  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	oeq_f32, .Lfunc_end2-oeq_f32
 
@@ -41,6 +44,7 @@ une_f32:
 	.result 	i32
 	f32.ne  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	une_f32, .Lfunc_end3-une_f32
 
@@ -51,6 +55,7 @@ olt_f32:
 	.result 	i32
 	f32.lt  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	olt_f32, .Lfunc_end4-olt_f32
 
@@ -61,6 +66,7 @@ ole_f32:
 	.result 	i32
 	f32.le  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	ole_f32, .Lfunc_end5-ole_f32
 
@@ -71,6 +77,7 @@ ogt_f32:
 	.result 	i32
 	f32.gt  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	ogt_f32, .Lfunc_end6-ogt_f32
 
@@ -81,6 +88,7 @@ oge_f32:
 	.result 	i32
 	f32.ge  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	oge_f32, .Lfunc_end7-oge_f32
 
@@ -95,6 +103,7 @@ ueq_f32:
 	i32.or  	$push3=, $pop2, $pop1
 	i32.or  	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end8:
 	.size	ueq_f32, .Lfunc_end8-ueq_f32
 
@@ -109,6 +118,7 @@ one_f32:
 	i32.and 	$push3=, $pop2, $pop1
 	i32.and 	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end9:
 	.size	one_f32, .Lfunc_end9-one_f32
 
@@ -123,6 +133,7 @@ ult_f32:
 	i32.or  	$push3=, $pop2, $pop1
 	i32.or  	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end10:
 	.size	ult_f32, .Lfunc_end10-ult_f32
 
@@ -137,6 +148,7 @@ ule_f32:
 	i32.or  	$push3=, $pop2, $pop1
 	i32.or  	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end11:
 	.size	ule_f32, .Lfunc_end11-ule_f32
 
@@ -151,6 +163,7 @@ ugt_f32:
 	i32.or  	$push3=, $pop2, $pop1
 	i32.or  	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end12:
 	.size	ugt_f32, .Lfunc_end12-ugt_f32
 
@@ -165,8 +178,8 @@ uge_f32:
 	i32.or  	$push3=, $pop2, $pop1
 	i32.or  	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end13:
 	.size	uge_f32, .Lfunc_end13-uge_f32
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/comparisons_f64.s b/test/llvm_autogenerated/comparisons_f64.s
index ab72bd46e..8bb75bcb6 100644
--- a/test/llvm_autogenerated/comparisons_f64.s
+++ b/test/llvm_autogenerated/comparisons_f64.s
@@ -9,6 +9,7 @@ ord_f64:
 	f64.eq  	$push0=, $1, $1
 	i32.and 	$push2=, $pop1, $pop0
 	return  	$pop2
+	.endfunc
 .Lfunc_end0:
 	.size	ord_f64, .Lfunc_end0-ord_f64
 
@@ -21,6 +22,7 @@ uno_f64:
 	f64.ne  	$push0=, $1, $1
 	i32.or  	$push2=, $pop1, $pop0
 	return  	$pop2
+	.endfunc
 .Lfunc_end1:
 	.size	uno_f64, .Lfunc_end1-uno_f64
 
@@ -31,6 +33,7 @@ oeq_f64:
 	.result 	i32
 	f64.eq  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	oeq_f64, .Lfunc_end2-oeq_f64
 
@@ -41,6 +44,7 @@ une_f64:
 	.result 	i32
 	f64.ne  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	une_f64, .Lfunc_end3-une_f64
 
@@ -51,6 +55,7 @@ olt_f64:
 	.result 	i32
 	f64.lt  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	olt_f64, .Lfunc_end4-olt_f64
 
@@ -61,6 +66,7 @@ ole_f64:
 	.result 	i32
 	f64.le  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	ole_f64, .Lfunc_end5-ole_f64
 
@@ -71,6 +77,7 @@ ogt_f64:
 	.result 	i32
 	f64.gt  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	ogt_f64, .Lfunc_end6-ogt_f64
 
@@ -81,6 +88,7 @@ oge_f64:
 	.result 	i32
 	f64.ge  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	oge_f64, .Lfunc_end7-oge_f64
 
@@ -95,6 +103,7 @@ ueq_f64:
 	i32.or  	$push3=, $pop2, $pop1
 	i32.or  	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end8:
 	.size	ueq_f64, .Lfunc_end8-ueq_f64
 
@@ -109,6 +118,7 @@ one_f64:
 	i32.and 	$push3=, $pop2, $pop1
 	i32.and 	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end9:
 	.size	one_f64, .Lfunc_end9-one_f64
 
@@ -123,6 +133,7 @@ ult_f64:
 	i32.or  	$push3=, $pop2, $pop1
 	i32.or  	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end10:
 	.size	ult_f64, .Lfunc_end10-ult_f64
 
@@ -137,6 +148,7 @@ ule_f64:
 	i32.or  	$push3=, $pop2, $pop1
 	i32.or  	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end11:
 	.size	ule_f64, .Lfunc_end11-ule_f64
 
@@ -151,6 +163,7 @@ ugt_f64:
 	i32.or  	$push3=, $pop2, $pop1
 	i32.or  	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end12:
 	.size	ugt_f64, .Lfunc_end12-ugt_f64
 
@@ -165,8 +178,8 @@ uge_f64:
 	i32.or  	$push3=, $pop2, $pop1
 	i32.or  	$push4=, $pop0, $pop3
 	return  	$pop4
+	.endfunc
 .Lfunc_end13:
 	.size	uge_f64, .Lfunc_end13-uge_f64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/comparisons_i32.s b/test/llvm_autogenerated/comparisons_i32.s
index a38bd77bd..340f6ae96 100644
--- a/test/llvm_autogenerated/comparisons_i32.s
+++ b/test/llvm_autogenerated/comparisons_i32.s
@@ -7,6 +7,7 @@ eq_i32:
 	.result 	i32
 	i32.eq  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	eq_i32, .Lfunc_end0-eq_i32
 
@@ -17,6 +18,7 @@ ne_i32:
 	.result 	i32
 	i32.ne  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	ne_i32, .Lfunc_end1-ne_i32
 
@@ -27,6 +29,7 @@ slt_i32:
 	.result 	i32
 	i32.lt_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	slt_i32, .Lfunc_end2-slt_i32
 
@@ -37,6 +40,7 @@ sle_i32:
 	.result 	i32
 	i32.le_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	sle_i32, .Lfunc_end3-sle_i32
 
@@ -47,6 +51,7 @@ ult_i32:
 	.result 	i32
 	i32.lt_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	ult_i32, .Lfunc_end4-ult_i32
 
@@ -57,6 +62,7 @@ ule_i32:
 	.result 	i32
 	i32.le_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	ule_i32, .Lfunc_end5-ule_i32
 
@@ -67,6 +73,7 @@ sgt_i32:
 	.result 	i32
 	i32.gt_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	sgt_i32, .Lfunc_end6-sgt_i32
 
@@ -77,6 +84,7 @@ sge_i32:
 	.result 	i32
 	i32.ge_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	sge_i32, .Lfunc_end7-sge_i32
 
@@ -87,6 +95,7 @@ ugt_i32:
 	.result 	i32
 	i32.gt_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	ugt_i32, .Lfunc_end8-ugt_i32
 
@@ -97,8 +106,8 @@ uge_i32:
 	.result 	i32
 	i32.ge_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end9:
 	.size	uge_i32, .Lfunc_end9-uge_i32
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/comparisons_i64.s b/test/llvm_autogenerated/comparisons_i64.s
index 1b151c7bd..9a9fd6a3b 100644
--- a/test/llvm_autogenerated/comparisons_i64.s
+++ b/test/llvm_autogenerated/comparisons_i64.s
@@ -7,6 +7,7 @@ eq_i64:
 	.result 	i32
 	i64.eq  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	eq_i64, .Lfunc_end0-eq_i64
 
@@ -17,6 +18,7 @@ ne_i64:
 	.result 	i32
 	i64.ne  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	ne_i64, .Lfunc_end1-ne_i64
 
@@ -27,6 +29,7 @@ slt_i64:
 	.result 	i32
 	i64.lt_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	slt_i64, .Lfunc_end2-slt_i64
 
@@ -37,6 +40,7 @@ sle_i64:
 	.result 	i32
 	i64.le_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	sle_i64, .Lfunc_end3-sle_i64
 
@@ -47,6 +51,7 @@ ult_i64:
 	.result 	i32
 	i64.lt_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	ult_i64, .Lfunc_end4-ult_i64
 
@@ -57,6 +62,7 @@ ule_i64:
 	.result 	i32
 	i64.le_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	ule_i64, .Lfunc_end5-ule_i64
 
@@ -67,6 +73,7 @@ sgt_i64:
 	.result 	i32
 	i64.gt_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	sgt_i64, .Lfunc_end6-sgt_i64
 
@@ -77,6 +84,7 @@ sge_i64:
 	.result 	i32
 	i64.ge_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	sge_i64, .Lfunc_end7-sge_i64
 
@@ -87,6 +95,7 @@ ugt_i64:
 	.result 	i32
 	i64.gt_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	ugt_i64, .Lfunc_end8-ugt_i64
 
@@ -97,8 +106,8 @@ uge_i64:
 	.result 	i32
 	i64.ge_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end9:
 	.size	uge_i64, .Lfunc_end9-uge_i64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/conv.s b/test/llvm_autogenerated/conv.s
index 9bb08c906..339051b1c 100644
--- a/test/llvm_autogenerated/conv.s
+++ b/test/llvm_autogenerated/conv.s
@@ -7,6 +7,7 @@ i32_wrap_i64:
 	.result 	i32
 	i32.wrap/i64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	i32_wrap_i64, .Lfunc_end0-i32_wrap_i64
 
@@ -17,6 +18,7 @@ i64_extend_s_i32:
 	.result 	i64
 	i64.extend_s/i32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	i64_extend_s_i32, .Lfunc_end1-i64_extend_s_i32
 
@@ -27,6 +29,7 @@ i64_extend_u_i32:
 	.result 	i64
 	i64.extend_u/i32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	i64_extend_u_i32, .Lfunc_end2-i64_extend_u_i32
 
@@ -37,6 +40,7 @@ i32_trunc_s_f32:
 	.result 	i32
 	i32.trunc_s/f32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	i32_trunc_s_f32, .Lfunc_end3-i32_trunc_s_f32
 
@@ -47,6 +51,7 @@ i32_trunc_u_f32:
 	.result 	i32
 	i32.trunc_u/f32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	i32_trunc_u_f32, .Lfunc_end4-i32_trunc_u_f32
 
@@ -57,6 +62,7 @@ i32_trunc_s_f64:
 	.result 	i32
 	i32.trunc_s/f64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	i32_trunc_s_f64, .Lfunc_end5-i32_trunc_s_f64
 
@@ -67,6 +73,7 @@ i32_trunc_u_f64:
 	.result 	i32
 	i32.trunc_u/f64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	i32_trunc_u_f64, .Lfunc_end6-i32_trunc_u_f64
 
@@ -77,6 +84,7 @@ i64_trunc_s_f32:
 	.result 	i64
 	i64.trunc_s/f32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	i64_trunc_s_f32, .Lfunc_end7-i64_trunc_s_f32
 
@@ -87,6 +95,7 @@ i64_trunc_u_f32:
 	.result 	i64
 	i64.trunc_u/f32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	i64_trunc_u_f32, .Lfunc_end8-i64_trunc_u_f32
 
@@ -97,6 +106,7 @@ i64_trunc_s_f64:
 	.result 	i64
 	i64.trunc_s/f64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end9:
 	.size	i64_trunc_s_f64, .Lfunc_end9-i64_trunc_s_f64
 
@@ -107,6 +117,7 @@ i64_trunc_u_f64:
 	.result 	i64
 	i64.trunc_u/f64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end10:
 	.size	i64_trunc_u_f64, .Lfunc_end10-i64_trunc_u_f64
 
@@ -117,6 +128,7 @@ f32_convert_s_i32:
 	.result 	f32
 	f32.convert_s/i32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end11:
 	.size	f32_convert_s_i32, .Lfunc_end11-f32_convert_s_i32
 
@@ -127,6 +139,7 @@ f32_convert_u_i32:
 	.result 	f32
 	f32.convert_u/i32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end12:
 	.size	f32_convert_u_i32, .Lfunc_end12-f32_convert_u_i32
 
@@ -137,6 +150,7 @@ f64_convert_s_i32:
 	.result 	f64
 	f64.convert_s/i32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end13:
 	.size	f64_convert_s_i32, .Lfunc_end13-f64_convert_s_i32
 
@@ -147,6 +161,7 @@ f64_convert_u_i32:
 	.result 	f64
 	f64.convert_u/i32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end14:
 	.size	f64_convert_u_i32, .Lfunc_end14-f64_convert_u_i32
 
@@ -157,6 +172,7 @@ f32_convert_s_i64:
 	.result 	f32
 	f32.convert_s/i64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end15:
 	.size	f32_convert_s_i64, .Lfunc_end15-f32_convert_s_i64
 
@@ -167,6 +183,7 @@ f32_convert_u_i64:
 	.result 	f32
 	f32.convert_u/i64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end16:
 	.size	f32_convert_u_i64, .Lfunc_end16-f32_convert_u_i64
 
@@ -177,6 +194,7 @@ f64_convert_s_i64:
 	.result 	f64
 	f64.convert_s/i64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end17:
 	.size	f64_convert_s_i64, .Lfunc_end17-f64_convert_s_i64
 
@@ -187,6 +205,7 @@ f64_convert_u_i64:
 	.result 	f64
 	f64.convert_u/i64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end18:
 	.size	f64_convert_u_i64, .Lfunc_end18-f64_convert_u_i64
 
@@ -197,6 +216,7 @@ f64_promote_f32:
 	.result 	f64
 	f64.promote/f32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end19:
 	.size	f64_promote_f32, .Lfunc_end19-f64_promote_f32
 
@@ -207,6 +227,7 @@ f32_demote_f64:
 	.result 	f32
 	f32.demote/f64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end20:
 	.size	f32_demote_f64, .Lfunc_end20-f32_demote_f64
 
@@ -219,6 +240,7 @@ anyext:
 	i64.const	$push1=, 32
 	i64.shl 	$push2=, $pop0, $pop1
 	return  	$pop2
+	.endfunc
 .Lfunc_end21:
 	.size	anyext, .Lfunc_end21-anyext
 
@@ -229,6 +251,7 @@ bitcast_i32_to_float:
 	.result 	f32
 	f32.reinterpret/i32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end22:
 	.size	bitcast_i32_to_float, .Lfunc_end22-bitcast_i32_to_float
 
@@ -239,6 +262,7 @@ bitcast_float_to_i32:
 	.result 	i32
 	i32.reinterpret/f32	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end23:
 	.size	bitcast_float_to_i32, .Lfunc_end23-bitcast_float_to_i32
 
@@ -249,6 +273,7 @@ bitcast_i64_to_double:
 	.result 	f64
 	f64.reinterpret/i64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end24:
 	.size	bitcast_i64_to_double, .Lfunc_end24-bitcast_i64_to_double
 
@@ -259,8 +284,8 @@ bitcast_double_to_i64:
 	.result 	i64
 	i64.reinterpret/f64	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end25:
 	.size	bitcast_double_to_i64, .Lfunc_end25-bitcast_double_to_i64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/copysign-casts.s b/test/llvm_autogenerated/copysign-casts.s
index 64e7e354e..1b3db084c 100644
--- a/test/llvm_autogenerated/copysign-casts.s
+++ b/test/llvm_autogenerated/copysign-casts.s
@@ -8,6 +8,7 @@ fold_promote:
 	f64.promote/f32	$push0=, $1
 	f64.copysign	$push1=, $0, $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end0:
 	.size	fold_promote, .Lfunc_end0-fold_promote
 
@@ -19,8 +20,8 @@ fold_demote:
 	f32.demote/f64	$push0=, $1
 	f32.copysign	$push1=, $0, $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end1:
 	.size	fold_demote, .Lfunc_end1-fold_demote
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/cpus.s b/test/llvm_autogenerated/cpus.s
index 9e1d701d2..9ec96b578 100644
--- a/test/llvm_autogenerated/cpus.s
+++ b/test/llvm_autogenerated/cpus.s
@@ -6,8 +6,8 @@ f:
 	.param  	i32
 	.result 	i32
 	return  	$0
+	.endfunc
 .Lfunc_end0:
 	.size	f, .Lfunc_end0-f
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/dead-vreg.s b/test/llvm_autogenerated/dead-vreg.s
index 83c484f36..8839c3011 100644
--- a/test/llvm_autogenerated/dead-vreg.s
+++ b/test/llvm_autogenerated/dead-vreg.s
@@ -4,45 +4,46 @@
 	.type	foo,@function
 foo:
 	.param  	i32, i32, i32
-	.local  	i32, i32, i32, i32, i32, i32, i32
-	i32.const	$4=, 1
+	.local  	i32, i32, i32, i32, i32, i32
 	block
-	i32.lt_s	$push0=, $2, $4
+	i32.const	$push3=, 1
+	i32.lt_s	$push0=, $2, $pop3
 	br_if   	$pop0, 0
-	i32.const	$5=, 0
 	i32.const	$push1=, 2
 	i32.shl 	$3=, $1, $pop1
-	copy_local	$6=, $5
+	i32.const	$5=, 0
+	i32.const	$push4=, 1
+	i32.lt_s	$4=, $1, $pop4
 .LBB0_2:
 	loop
-	copy_local	$7=, $5
-	copy_local	$8=, $0
-	copy_local	$9=, $1
+	i32.const	$6=, 0
+	copy_local	$7=, $0
+	copy_local	$8=, $1
 	block
-	i32.lt_s	$push2=, $1, $4
-	br_if   	$pop2, 0
+	br_if   	$4, 0
 .LBB0_3:
 	loop
-	i32.const	$push3=, -1
-	i32.add 	$9=, $9, $pop3
-	i32.store	$discard=, 0($8), $7
-	i32.const	$push4=, 4
-	i32.add 	$8=, $8, $pop4
-	i32.add 	$7=, $7, $6
-	br_if   	$9, 0
+	i32.store	$discard=, 0($7), $6
+	i32.const	$push6=, -1
+	i32.add 	$8=, $8, $pop6
+	i32.const	$push5=, 4
+	i32.add 	$7=, $7, $pop5
+	i32.add 	$6=, $6, $5
+	br_if   	$8, 0
 .LBB0_4:
 	end_loop
 	end_block
-	i32.add 	$6=, $6, $4
+	i32.const	$push7=, 1
+	i32.add 	$5=, $5, $pop7
 	i32.add 	$0=, $0, $3
-	i32.ne  	$push5=, $6, $2
-	br_if   	$pop5, 0
+	i32.ne  	$push2=, $5, $2
+	br_if   	$pop2, 0
 .LBB0_5:
 	end_loop
 	end_block
 	return
+	.endfunc
 .Lfunc_end0:
 	.size	foo, .Lfunc_end0-foo
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/dead-vreg.wast b/test/llvm_autogenerated/dead-vreg.wast
index 3eb4dd029..282cd8fe0 100644
--- a/test/llvm_autogenerated/dead-vreg.wast
+++ b/test/llvm_autogenerated/dead-vreg.wast
@@ -8,83 +8,79 @@
     (local $$6 i32)
     (local $$7 i32)
     (local $$8 i32)
-    (local $$9 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$4
-          (i32.const 1)
-        )
         (block $label$0
           (br_if
             (i32.lt_s
               (get_local $$2)
-              (get_local $$4)
+              (i32.const 1)
             )
             $label$0
           )
-          (set_local $$5
-            (i32.const 0)
-          )
           (set_local $$3
             (i32.shl
               (get_local $$1)
               (i32.const 2)
             )
           )
-          (set_local $$6
-            (get_local $$5)
+          (set_local $$5
+            (i32.const 0)
+          )
+          (set_local $$4
+            (i32.lt_s
+              (get_local $$1)
+              (i32.const 1)
+            )
           )
           (loop $label$2 $label$1
-            (set_local $$7
-              (get_local $$5)
+            (set_local $$6
+              (i32.const 0)
             )
-            (set_local $$8
+            (set_local $$7
               (get_local $$0)
             )
-            (set_local $$9
+            (set_local $$8
               (get_local $$1)
             )
             (block $label$3
               (br_if
-                (i32.lt_s
-                  (get_local $$1)
-                  (get_local $$4)
-                )
+                (get_local $$4)
                 $label$3
               )
               (loop $label$5 $label$4
-                (set_local $$9
-                  (i32.add
-                    (get_local $$9)
-                    (i32.const -1)
-                  )
-                )
                 (i32.store align=4
-                  (get_local $$8)
                   (get_local $$7)
+                  (get_local $$6)
                 )
                 (set_local $$8
                   (i32.add
                     (get_local $$8)
-                    (i32.const 4)
+                    (i32.const -1)
                   )
                 )
                 (set_local $$7
                   (i32.add
                     (get_local $$7)
+                    (i32.const 4)
+                  )
+                )
+                (set_local $$6
+                  (i32.add
                     (get_local $$6)
+                    (get_local $$5)
                   )
                 )
                 (br_if
-                  (get_local $$9)
+                  (get_local $$8)
                   $label$4
                 )
               )
             )
-            (set_local $$6
+            (set_local $$5
               (i32.add
-                (get_local $$6)
-                (get_local $$4)
+                (get_local $$5)
+                (i32.const 1)
               )
             )
             (set_local $$0
@@ -95,7 +91,7 @@
             )
             (br_if
               (i32.ne
-                (get_local $$6)
+                (get_local $$5)
                 (get_local $$2)
               )
               $label$1
diff --git a/test/llvm_autogenerated/f32.s b/test/llvm_autogenerated/f32.s
index 229f3f7dc..6eff47f4e 100644
--- a/test/llvm_autogenerated/f32.s
+++ b/test/llvm_autogenerated/f32.s
@@ -7,6 +7,7 @@ fadd32:
 	.result 	f32
 	f32.add 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	fadd32, .Lfunc_end0-fadd32
 
@@ -17,6 +18,7 @@ fsub32:
 	.result 	f32
 	f32.sub 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	fsub32, .Lfunc_end1-fsub32
 
@@ -27,6 +29,7 @@ fmul32:
 	.result 	f32
 	f32.mul 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	fmul32, .Lfunc_end2-fmul32
 
@@ -37,6 +40,7 @@ fdiv32:
 	.result 	f32
 	f32.div 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	fdiv32, .Lfunc_end3-fdiv32
 
@@ -47,6 +51,7 @@ fabs32:
 	.result 	f32
 	f32.abs 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	fabs32, .Lfunc_end4-fabs32
 
@@ -57,6 +62,7 @@ fneg32:
 	.result 	f32
 	f32.neg 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	fneg32, .Lfunc_end5-fneg32
 
@@ -67,6 +73,7 @@ copysign32:
 	.result 	f32
 	f32.copysign	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	copysign32, .Lfunc_end6-copysign32
 
@@ -77,6 +84,7 @@ sqrt32:
 	.result 	f32
 	f32.sqrt	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	sqrt32, .Lfunc_end7-sqrt32
 
@@ -87,6 +95,7 @@ ceil32:
 	.result 	f32
 	f32.ceil	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	ceil32, .Lfunc_end8-ceil32
 
@@ -97,6 +106,7 @@ floor32:
 	.result 	f32
 	f32.floor	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end9:
 	.size	floor32, .Lfunc_end9-floor32
 
@@ -107,6 +117,7 @@ trunc32:
 	.result 	f32
 	f32.trunc	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end10:
 	.size	trunc32, .Lfunc_end10-trunc32
 
@@ -117,6 +128,7 @@ nearest32:
 	.result 	f32
 	f32.nearest	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end11:
 	.size	nearest32, .Lfunc_end11-nearest32
 
@@ -127,6 +139,7 @@ nearest32_via_rint:
 	.result 	f32
 	f32.nearest	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end12:
 	.size	nearest32_via_rint, .Lfunc_end12-nearest32_via_rint
 
@@ -138,6 +151,7 @@ fmin32:
 	f32.const	$push0=, 0x0p0
 	f32.min 	$push1=, $0, $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end13:
 	.size	fmin32, .Lfunc_end13-fmin32
 
@@ -149,6 +163,7 @@ fmax32:
 	f32.const	$push0=, 0x0p0
 	f32.max 	$push1=, $0, $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end14:
 	.size	fmax32, .Lfunc_end14-fmax32
 
@@ -159,8 +174,8 @@ fma32:
 	.result 	f32
 	f32.call	$push0=, fmaf@FUNCTION, $0, $1, $2
 	return  	$pop0
+	.endfunc
 .Lfunc_end15:
 	.size	fma32, .Lfunc_end15-fma32
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/f64.s b/test/llvm_autogenerated/f64.s
index 520a1a68d..70b233ceb 100644
--- a/test/llvm_autogenerated/f64.s
+++ b/test/llvm_autogenerated/f64.s
@@ -7,6 +7,7 @@ fadd64:
 	.result 	f64
 	f64.add 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	fadd64, .Lfunc_end0-fadd64
 
@@ -17,6 +18,7 @@ fsub64:
 	.result 	f64
 	f64.sub 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	fsub64, .Lfunc_end1-fsub64
 
@@ -27,6 +29,7 @@ fmul64:
 	.result 	f64
 	f64.mul 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	fmul64, .Lfunc_end2-fmul64
 
@@ -37,6 +40,7 @@ fdiv64:
 	.result 	f64
 	f64.div 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	fdiv64, .Lfunc_end3-fdiv64
 
@@ -47,6 +51,7 @@ fabs64:
 	.result 	f64
 	f64.abs 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	fabs64, .Lfunc_end4-fabs64
 
@@ -57,6 +62,7 @@ fneg64:
 	.result 	f64
 	f64.neg 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	fneg64, .Lfunc_end5-fneg64
 
@@ -67,6 +73,7 @@ copysign64:
 	.result 	f64
 	f64.copysign	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	copysign64, .Lfunc_end6-copysign64
 
@@ -77,6 +84,7 @@ sqrt64:
 	.result 	f64
 	f64.sqrt	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	sqrt64, .Lfunc_end7-sqrt64
 
@@ -87,6 +95,7 @@ ceil64:
 	.result 	f64
 	f64.ceil	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	ceil64, .Lfunc_end8-ceil64
 
@@ -97,6 +106,7 @@ floor64:
 	.result 	f64
 	f64.floor	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end9:
 	.size	floor64, .Lfunc_end9-floor64
 
@@ -107,6 +117,7 @@ trunc64:
 	.result 	f64
 	f64.trunc	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end10:
 	.size	trunc64, .Lfunc_end10-trunc64
 
@@ -117,6 +128,7 @@ nearest64:
 	.result 	f64
 	f64.nearest	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end11:
 	.size	nearest64, .Lfunc_end11-nearest64
 
@@ -127,6 +139,7 @@ nearest64_via_rint:
 	.result 	f64
 	f64.nearest	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end12:
 	.size	nearest64_via_rint, .Lfunc_end12-nearest64_via_rint
 
@@ -138,6 +151,7 @@ fmin64:
 	f64.const	$push0=, 0x0p0
 	f64.min 	$push1=, $0, $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end13:
 	.size	fmin64, .Lfunc_end13-fmin64
 
@@ -149,6 +163,7 @@ fmax64:
 	f64.const	$push0=, 0x0p0
 	f64.max 	$push1=, $0, $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end14:
 	.size	fmax64, .Lfunc_end14-fmax64
 
@@ -159,8 +174,8 @@ fma64:
 	.result 	f64
 	f64.call	$push0=, fma@FUNCTION, $0, $1, $2
 	return  	$pop0
+	.endfunc
 .Lfunc_end15:
 	.size	fma64, .Lfunc_end15-fma64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/fast-isel.s b/test/llvm_autogenerated/fast-isel.s
index e511c33a1..63a2718a3 100644
--- a/test/llvm_autogenerated/fast-isel.s
+++ b/test/llvm_autogenerated/fast-isel.s
@@ -6,6 +6,7 @@ immediate_f32:
 	.result 	f32
 	f32.const	$push0=, 0x1.4p1
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	immediate_f32, .Lfunc_end0-immediate_f32
 
@@ -15,8 +16,8 @@ immediate_f64:
 	.result 	f64
 	f64.const	$push0=, 0x1.4p1
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	immediate_f64, .Lfunc_end1-immediate_f64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/frem.s b/test/llvm_autogenerated/frem.s
index af01a9c2b..6ef9559ab 100644
--- a/test/llvm_autogenerated/frem.s
+++ b/test/llvm_autogenerated/frem.s
@@ -7,6 +7,7 @@ frem32:
 	.result 	f32
 	f32.call	$push0=, fmodf@FUNCTION, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	frem32, .Lfunc_end0-frem32
 
@@ -17,8 +18,8 @@ frem64:
 	.result 	f64
 	f64.call	$push0=, fmod@FUNCTION, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	frem64, .Lfunc_end1-frem64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/func.s b/test/llvm_autogenerated/func.s
index 49425e63c..fbe92c359 100644
--- a/test/llvm_autogenerated/func.s
+++ b/test/llvm_autogenerated/func.s
@@ -4,6 +4,7 @@
 	.type	f0,@function
 f0:
 	return
+	.endfunc
 .Lfunc_end0:
 	.size	f0, .Lfunc_end0-f0
 
@@ -13,6 +14,7 @@ f1:
 	.result 	i32
 	i32.const	$push0=, 0
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	f1, .Lfunc_end1-f1
 
@@ -23,6 +25,7 @@ f2:
 	.result 	i32
 	i32.const	$push0=, 0
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	f2, .Lfunc_end2-f2
 
@@ -31,6 +34,7 @@ f2:
 f3:
 	.param  	i32, f32
 	return
+	.endfunc
 .Lfunc_end3:
 	.size	f3, .Lfunc_end3-f3
 
@@ -39,18 +43,19 @@ f3:
 f4:
 	.param  	i32
 	.result 	i32
-	.local  	i32
-	i32.const	$1=, 1
 	block
-	i32.and 	$push0=, $0, $1
-	i32.const	$push2=, 0
-	i32.eq  	$push3=, $pop0, $pop2
-	br_if   	$pop3, 0
+	i32.const	$push2=, 1
+	i32.and 	$push0=, $0, $pop2
+	i32.const	$push4=, 0
+	i32.eq  	$push5=, $pop0, $pop4
+	br_if   	$pop5, 0
 	i32.const	$push1=, 0
 	return  	$pop1
 .LBB4_2:
 	end_block
-	return  	$1
+	i32.const	$push3=, 1
+	return  	$pop3
+	.endfunc
 .Lfunc_end4:
 	.size	f4, .Lfunc_end4-f4
 
@@ -59,8 +64,8 @@ f4:
 f5:
 	.result 	f32
 	unreachable
+	.endfunc
 .Lfunc_end5:
 	.size	f5, .Lfunc_end5-f5
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/func.wast b/test/llvm_autogenerated/func.wast
index ae815c2e5..4ee2497f7 100644
--- a/test/llvm_autogenerated/func.wast
+++ b/test/llvm_autogenerated/func.wast
@@ -39,18 +39,14 @@
     )
   )
   (func $f4 (param $$0 i32) (result i32)
-    (local $$1 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$1
-          (i32.const 1)
-        )
         (block $label$0
           (br_if
             (i32.eq
               (i32.and
                 (get_local $$0)
-                (get_local $$1)
+                (i32.const 1)
               )
               (i32.const 0)
             )
@@ -61,7 +57,7 @@
           )
         )
         (br $fake_return_waka123
-          (get_local $$1)
+          (i32.const 1)
         )
       )
     )
diff --git a/test/llvm_autogenerated/global.s b/test/llvm_autogenerated/global.s
index 3b2432998..0eaf572ec 100644
--- a/test/llvm_autogenerated/global.s
+++ b/test/llvm_autogenerated/global.s
@@ -7,6 +7,7 @@ foo:
 	i32.const	$push0=, 0
 	i32.load	$push1=, answer($pop0)
 	return  	$pop1
+	.endfunc
 .Lfunc_end0:
 	.size	foo, .Lfunc_end0-foo
 
@@ -15,8 +16,9 @@ foo:
 call_memcpy:
 	.param  	i32, i32, i32
 	.result 	i32
-	call    	memcpy@FUNCTION, $0, $1, $2
+	i32.call	$discard=, memcpy@FUNCTION, $0, $1, $2
 	return  	$0
+	.endfunc
 .Lfunc_end1:
 	.size	call_memcpy, .Lfunc_end1-call_memcpy
 
@@ -145,5 +147,17 @@ rom:
 	.skip	512
 	.size	rom, 512
 
+	.type	array,@object
+array:
+	.skip	8
+	.size	array, 8
+
+	.type	pointer_to_array,@object
+	.section	.data.rel.ro,"aw",@progbits
+	.globl	pointer_to_array
+	.align	2
+pointer_to_array:
+	.int32	array+4
+	.size	pointer_to_array, 4
+
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/global.wast b/test/llvm_autogenerated/global.wast
index 6569b8376..6a9a6293d 100644
--- a/test/llvm_autogenerated/global.wast
+++ b/test/llvm_autogenerated/global.wast
@@ -1,7 +1,7 @@
 (module
-  (memory 1184 4294967295 (segment 8 "9\05\00\00") (segment 24 "\01\00\00\00") (segment 28 "*\00\00\00") (segment 32 "\ff\ff\ff\ff") (segment 64 "\00\00\00\00\01\00\00\00") (segment 72 "\ff\ff\ff\ff\ff\ff\ff\ff") (segment 92 "\00\00\00\80") (segment 96 "\00\00\00@") (segment 128 "\00\00\00\00\00\00\00\80") (segment 136 "\00\00\00\00\00\00\00@") (segment 656 "\e0\00\00\00"))
-  (type $FUNCSIG$viii (func (param i32 i32 i32)))
-  (import $memcpy "env" "memcpy" (param i32 i32 i32))
+  (memory 1196 4294967295 (segment 8 "9\05\00\00") (segment 24 "\01\00\00\00") (segment 28 "*\00\00\00") (segment 32 "\ff\ff\ff\ff") (segment 64 "\00\00\00\00\01\00\00\00") (segment 72 "\ff\ff\ff\ff\ff\ff\ff\ff") (segment 92 "\00\00\00\80") (segment 96 "\00\00\00@") (segment 128 "\00\00\00\00\00\00\00\80") (segment 136 "\00\00\00\00\00\00\00@") (segment 656 "\e0\00\00\00") (segment 1192 "\a4\04\00\00"))
+  (type $FUNCSIG$iiii (func (param i32 i32 i32) (result i32)))
+  (import $memcpy "env" "memcpy" (param i32 i32 i32) (result i32))
   (export "foo" $foo)
   (export "call_memcpy" $call_memcpy)
   (func $foo (result i32)
@@ -30,4 +30,4 @@
     )
   )
 )
-;; METADATA: { "asmConsts": {},"staticBump": 1183 }
+;; METADATA: { "asmConsts": {},"staticBump": 1195 }
diff --git a/test/llvm_autogenerated/globl.s b/test/llvm_autogenerated/globl.s
index 0453076f7..894ec0bf5 100644
--- a/test/llvm_autogenerated/globl.s
+++ b/test/llvm_autogenerated/globl.s
@@ -4,8 +4,8 @@
 	.type	foo,@function
 foo:
 	return
+	.endfunc
 .Lfunc_end0:
 	.size	foo, .Lfunc_end0-foo
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/i32-load-store-alignment.s b/test/llvm_autogenerated/i32-load-store-alignment.s
new file mode 100644
index 000000000..9f02ec6bb
--- /dev/null
+++ b/test/llvm_autogenerated/i32-load-store-alignment.s
@@ -0,0 +1,213 @@
+	.text
+	.file	"/s/llvm/llvm/test/CodeGen/WebAssembly/i32-load-store-alignment.ll"
+	.globl	ldi32_a1
+	.type	ldi32_a1,@function
+ldi32_a1:
+	.param  	i32
+	.result 	i32
+	i32.load	$push0=, 0($0):p2align=0
+	return  	$pop0
+	.endfunc
+.Lfunc_end0:
+	.size	ldi32_a1, .Lfunc_end0-ldi32_a1
+
+	.globl	ldi32_a2
+	.type	ldi32_a2,@function
+ldi32_a2:
+	.param  	i32
+	.result 	i32
+	i32.load	$push0=, 0($0):p2align=1
+	return  	$pop0
+	.endfunc
+.Lfunc_end1:
+	.size	ldi32_a2, .Lfunc_end1-ldi32_a2
+
+	.globl	ldi32_a4
+	.type	ldi32_a4,@function
+ldi32_a4:
+	.param  	i32
+	.result 	i32
+	i32.load	$push0=, 0($0)
+	return  	$pop0
+	.endfunc
+.Lfunc_end2:
+	.size	ldi32_a4, .Lfunc_end2-ldi32_a4
+
+	.globl	ldi32
+	.type	ldi32,@function
+ldi32:
+	.param  	i32
+	.result 	i32
+	i32.load	$push0=, 0($0)
+	return  	$pop0
+	.endfunc
+.Lfunc_end3:
+	.size	ldi32, .Lfunc_end3-ldi32
+
+	.globl	ldi32_a8
+	.type	ldi32_a8,@function
+ldi32_a8:
+	.param  	i32
+	.result 	i32
+	i32.load	$push0=, 0($0):p2align=3
+	return  	$pop0
+	.endfunc
+.Lfunc_end4:
+	.size	ldi32_a8, .Lfunc_end4-ldi32_a8
+
+	.globl	ldi8_a1
+	.type	ldi8_a1,@function
+ldi8_a1:
+	.param  	i32
+	.result 	i32
+	i32.load8_u	$push0=, 0($0)
+	return  	$pop0
+	.endfunc
+.Lfunc_end5:
+	.size	ldi8_a1, .Lfunc_end5-ldi8_a1
+
+	.globl	ldi8_a2
+	.type	ldi8_a2,@function
+ldi8_a2:
+	.param  	i32
+	.result 	i32
+	i32.load8_u	$push0=, 0($0):p2align=1
+	return  	$pop0
+	.endfunc
+.Lfunc_end6:
+	.size	ldi8_a2, .Lfunc_end6-ldi8_a2
+
+	.globl	ldi16_a1
+	.type	ldi16_a1,@function
+ldi16_a1:
+	.param  	i32
+	.result 	i32
+	i32.load16_u	$push0=, 0($0):p2align=0
+	return  	$pop0
+	.endfunc
+.Lfunc_end7:
+	.size	ldi16_a1, .Lfunc_end7-ldi16_a1
+
+	.globl	ldi16_a2
+	.type	ldi16_a2,@function
+ldi16_a2:
+	.param  	i32
+	.result 	i32
+	i32.load16_u	$push0=, 0($0)
+	return  	$pop0
+	.endfunc
+.Lfunc_end8:
+	.size	ldi16_a2, .Lfunc_end8-ldi16_a2
+
+	.globl	ldi16_a4
+	.type	ldi16_a4,@function
+ldi16_a4:
+	.param  	i32
+	.result 	i32
+	i32.load16_u	$push0=, 0($0):p2align=2
+	return  	$pop0
+	.endfunc
+.Lfunc_end9:
+	.size	ldi16_a4, .Lfunc_end9-ldi16_a4
+
+	.globl	sti32_a1
+	.type	sti32_a1,@function
+sti32_a1:
+	.param  	i32, i32
+	i32.store	$discard=, 0($0):p2align=0, $1
+	return
+	.endfunc
+.Lfunc_end10:
+	.size	sti32_a1, .Lfunc_end10-sti32_a1
+
+	.globl	sti32_a2
+	.type	sti32_a2,@function
+sti32_a2:
+	.param  	i32, i32
+	i32.store	$discard=, 0($0):p2align=1, $1
+	return
+	.endfunc
+.Lfunc_end11:
+	.size	sti32_a2, .Lfunc_end11-sti32_a2
+
+	.globl	sti32_a4
+	.type	sti32_a4,@function
+sti32_a4:
+	.param  	i32, i32
+	i32.store	$discard=, 0($0), $1
+	return
+	.endfunc
+.Lfunc_end12:
+	.size	sti32_a4, .Lfunc_end12-sti32_a4
+
+	.globl	sti32
+	.type	sti32,@function
+sti32:
+	.param  	i32, i32
+	i32.store	$discard=, 0($0), $1
+	return
+	.endfunc
+.Lfunc_end13:
+	.size	sti32, .Lfunc_end13-sti32
+
+	.globl	sti32_a8
+	.type	sti32_a8,@function
+sti32_a8:
+	.param  	i32, i32
+	i32.store	$discard=, 0($0):p2align=3, $1
+	return
+	.endfunc
+.Lfunc_end14:
+	.size	sti32_a8, .Lfunc_end14-sti32_a8
+
+	.globl	sti8_a1
+	.type	sti8_a1,@function
+sti8_a1:
+	.param  	i32, i32
+	i32.store8	$discard=, 0($0), $1
+	return
+	.endfunc
+.Lfunc_end15:
+	.size	sti8_a1, .Lfunc_end15-sti8_a1
+
+	.globl	sti8_a2
+	.type	sti8_a2,@function
+sti8_a2:
+	.param  	i32, i32
+	i32.store8	$discard=, 0($0):p2align=1, $1
+	return
+	.endfunc
+.Lfunc_end16:
+	.size	sti8_a2, .Lfunc_end16-sti8_a2
+
+	.globl	sti16_a1
+	.type	sti16_a1,@function
+sti16_a1:
+	.param  	i32, i32
+	i32.store16	$discard=, 0($0):p2align=0, $1
+	return
+	.endfunc
+.Lfunc_end17:
+	.size	sti16_a1, .Lfunc_end17-sti16_a1
+
+	.globl	sti16_a2
+	.type	sti16_a2,@function
+sti16_a2:
+	.param  	i32, i32
+	i32.store16	$discard=, 0($0), $1
+	return
+	.endfunc
+.Lfunc_end18:
+	.size	sti16_a2, .Lfunc_end18-sti16_a2
+
+	.globl	sti16_a4
+	.type	sti16_a4,@function
+sti16_a4:
+	.param  	i32, i32
+	i32.store16	$discard=, 0($0):p2align=2, $1
+	return
+	.endfunc
+.Lfunc_end19:
+	.size	sti16_a4, .Lfunc_end19-sti16_a4
+
+
diff --git a/test/llvm_autogenerated/i32-load-store-alignment.wast b/test/llvm_autogenerated/i32-load-store-alignment.wast
new file mode 100644
index 000000000..b0ed4951a
--- /dev/null
+++ b/test/llvm_autogenerated/i32-load-store-alignment.wast
@@ -0,0 +1,244 @@
+(module
+  (memory 0 4294967295)
+  (export "ldi32_a1" $ldi32_a1)
+  (export "ldi32_a2" $ldi32_a2)
+  (export "ldi32_a4" $ldi32_a4)
+  (export "ldi32" $ldi32)
+  (export "ldi32_a8" $ldi32_a8)
+  (export "ldi8_a1" $ldi8_a1)
+  (export "ldi8_a2" $ldi8_a2)
+  (export "ldi16_a1" $ldi16_a1)
+  (export "ldi16_a2" $ldi16_a2)
+  (export "ldi16_a4" $ldi16_a4)
+  (export "sti32_a1" $sti32_a1)
+  (export "sti32_a2" $sti32_a2)
+  (export "sti32_a4" $sti32_a4)
+  (export "sti32" $sti32)
+  (export "sti32_a8" $sti32_a8)
+  (export "sti8_a1" $sti8_a1)
+  (export "sti8_a2" $sti8_a2)
+  (export "sti16_a1" $sti16_a1)
+  (export "sti16_a2" $sti16_a2)
+  (export "sti16_a4" $sti16_a4)
+  (func $ldi32_a1 (param $$0 i32) (result i32)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i32.load align=1
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi32_a2 (param $$0 i32) (result i32)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i32.load align=2
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi32_a4 (param $$0 i32) (result i32)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i32.load align=4
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi32 (param $$0 i32) (result i32)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i32.load align=4
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi32_a8 (param $$0 i32) (result i32)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i32.load align=8
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi8_a1 (param $$0 i32) (result i32)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i32.load8_u align=1
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi8_a2 (param $$0 i32) (result i32)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i32.load8_u align=2
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi16_a1 (param $$0 i32) (result i32)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i32.load16_u align=1
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi16_a2 (param $$0 i32) (result i32)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i32.load16_u align=2
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi16_a4 (param $$0 i32) (result i32)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i32.load16_u align=4
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $sti32_a1 (param $$0 i32) (param $$1 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store align=1
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti32_a2 (param $$0 i32) (param $$1 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store align=2
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti32_a4 (param $$0 i32) (param $$1 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store align=4
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti32 (param $$0 i32) (param $$1 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store align=4
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti32_a8 (param $$0 i32) (param $$1 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store align=8
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti8_a1 (param $$0 i32) (param $$1 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store align=8
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti8_a2 (param $$0 i32) (param $$1 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store align=2
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti16_a1 (param $$0 i32) (param $$1 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store align=1
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti16_a2 (param $$0 i32) (param $$1 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store align=16
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti16_a4 (param $$0 i32) (param $$1 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store align=4
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+)
+;; METADATA: { "asmConsts": {},"staticBump": 4 }
diff --git a/test/llvm_autogenerated/i32.s b/test/llvm_autogenerated/i32.s
index c214cac8f..66128cd2e 100644
--- a/test/llvm_autogenerated/i32.s
+++ b/test/llvm_autogenerated/i32.s
@@ -7,6 +7,7 @@ add32:
 	.result 	i32
 	i32.add 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	add32, .Lfunc_end0-add32
 
@@ -17,6 +18,7 @@ sub32:
 	.result 	i32
 	i32.sub 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	sub32, .Lfunc_end1-sub32
 
@@ -27,6 +29,7 @@ mul32:
 	.result 	i32
 	i32.mul 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	mul32, .Lfunc_end2-mul32
 
@@ -37,6 +40,7 @@ sdiv32:
 	.result 	i32
 	i32.div_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	sdiv32, .Lfunc_end3-sdiv32
 
@@ -47,6 +51,7 @@ udiv32:
 	.result 	i32
 	i32.div_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	udiv32, .Lfunc_end4-udiv32
 
@@ -57,6 +62,7 @@ srem32:
 	.result 	i32
 	i32.rem_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	srem32, .Lfunc_end5-srem32
 
@@ -67,6 +73,7 @@ urem32:
 	.result 	i32
 	i32.rem_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	urem32, .Lfunc_end6-urem32
 
@@ -77,6 +84,7 @@ and32:
 	.result 	i32
 	i32.and 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	and32, .Lfunc_end7-and32
 
@@ -87,6 +95,7 @@ or32:
 	.result 	i32
 	i32.or  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	or32, .Lfunc_end8-or32
 
@@ -97,6 +106,7 @@ xor32:
 	.result 	i32
 	i32.xor 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end9:
 	.size	xor32, .Lfunc_end9-xor32
 
@@ -107,6 +117,7 @@ shl32:
 	.result 	i32
 	i32.shl 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end10:
 	.size	shl32, .Lfunc_end10-shl32
 
@@ -117,6 +128,7 @@ shr32:
 	.result 	i32
 	i32.shr_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end11:
 	.size	shr32, .Lfunc_end11-shr32
 
@@ -127,6 +139,7 @@ sar32:
 	.result 	i32
 	i32.shr_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end12:
 	.size	sar32, .Lfunc_end12-sar32
 
@@ -137,6 +150,7 @@ clz32:
 	.result 	i32
 	i32.clz 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end13:
 	.size	clz32, .Lfunc_end13-clz32
 
@@ -147,6 +161,7 @@ clz32_zero_undef:
 	.result 	i32
 	i32.clz 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end14:
 	.size	clz32_zero_undef, .Lfunc_end14-clz32_zero_undef
 
@@ -157,6 +172,7 @@ ctz32:
 	.result 	i32
 	i32.ctz 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end15:
 	.size	ctz32, .Lfunc_end15-ctz32
 
@@ -167,6 +183,7 @@ ctz32_zero_undef:
 	.result 	i32
 	i32.ctz 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end16:
 	.size	ctz32_zero_undef, .Lfunc_end16-ctz32_zero_undef
 
@@ -177,8 +194,8 @@ popcnt32:
 	.result 	i32
 	i32.popcnt	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end17:
 	.size	popcnt32, .Lfunc_end17-popcnt32
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/i64-load-store-alignment.s b/test/llvm_autogenerated/i64-load-store-alignment.s
new file mode 100644
index 000000000..a61aa2965
--- /dev/null
+++ b/test/llvm_autogenerated/i64-load-store-alignment.s
@@ -0,0 +1,318 @@
+	.text
+	.file	"/s/llvm/llvm/test/CodeGen/WebAssembly/i64-load-store-alignment.ll"
+	.globl	ldi64_a1
+	.type	ldi64_a1,@function
+ldi64_a1:
+	.param  	i32
+	.result 	i64
+	i64.load	$push0=, 0($0):p2align=0
+	return  	$pop0
+	.endfunc
+.Lfunc_end0:
+	.size	ldi64_a1, .Lfunc_end0-ldi64_a1
+
+	.globl	ldi64_a2
+	.type	ldi64_a2,@function
+ldi64_a2:
+	.param  	i32
+	.result 	i64
+	i64.load	$push0=, 0($0):p2align=1
+	return  	$pop0
+	.endfunc
+.Lfunc_end1:
+	.size	ldi64_a2, .Lfunc_end1-ldi64_a2
+
+	.globl	ldi64_a4
+	.type	ldi64_a4,@function
+ldi64_a4:
+	.param  	i32
+	.result 	i64
+	i64.load	$push0=, 0($0):p2align=2
+	return  	$pop0
+	.endfunc
+.Lfunc_end2:
+	.size	ldi64_a4, .Lfunc_end2-ldi64_a4
+
+	.globl	ldi64_a8
+	.type	ldi64_a8,@function
+ldi64_a8:
+	.param  	i32
+	.result 	i64
+	i64.load	$push0=, 0($0)
+	return  	$pop0
+	.endfunc
+.Lfunc_end3:
+	.size	ldi64_a8, .Lfunc_end3-ldi64_a8
+
+	.globl	ldi64
+	.type	ldi64,@function
+ldi64:
+	.param  	i32
+	.result 	i64
+	i64.load	$push0=, 0($0)
+	return  	$pop0
+	.endfunc
+.Lfunc_end4:
+	.size	ldi64, .Lfunc_end4-ldi64
+
+	.globl	ldi64_a16
+	.type	ldi64_a16,@function
+ldi64_a16:
+	.param  	i32
+	.result 	i64
+	i64.load	$push0=, 0($0):p2align=4
+	return  	$pop0
+	.endfunc
+.Lfunc_end5:
+	.size	ldi64_a16, .Lfunc_end5-ldi64_a16
+
+	.globl	ldi8_a1
+	.type	ldi8_a1,@function
+ldi8_a1:
+	.param  	i32
+	.result 	i64
+	i64.load8_u	$push0=, 0($0)
+	return  	$pop0
+	.endfunc
+.Lfunc_end6:
+	.size	ldi8_a1, .Lfunc_end6-ldi8_a1
+
+	.globl	ldi8_a2
+	.type	ldi8_a2,@function
+ldi8_a2:
+	.param  	i32
+	.result 	i64
+	i64.load8_u	$push0=, 0($0):p2align=1
+	return  	$pop0
+	.endfunc
+.Lfunc_end7:
+	.size	ldi8_a2, .Lfunc_end7-ldi8_a2
+
+	.globl	ldi16_a1
+	.type	ldi16_a1,@function
+ldi16_a1:
+	.param  	i32
+	.result 	i64
+	i64.load16_u	$push0=, 0($0):p2align=0
+	return  	$pop0
+	.endfunc
+.Lfunc_end8:
+	.size	ldi16_a1, .Lfunc_end8-ldi16_a1
+
+	.globl	ldi16_a2
+	.type	ldi16_a2,@function
+ldi16_a2:
+	.param  	i32
+	.result 	i64
+	i64.load16_u	$push0=, 0($0)
+	return  	$pop0
+	.endfunc
+.Lfunc_end9:
+	.size	ldi16_a2, .Lfunc_end9-ldi16_a2
+
+	.globl	ldi16_a4
+	.type	ldi16_a4,@function
+ldi16_a4:
+	.param  	i32
+	.result 	i64
+	i64.load16_u	$push0=, 0($0):p2align=2
+	return  	$pop0
+	.endfunc
+.Lfunc_end10:
+	.size	ldi16_a4, .Lfunc_end10-ldi16_a4
+
+	.globl	ldi32_a1
+	.type	ldi32_a1,@function
+ldi32_a1:
+	.param  	i32
+	.result 	i64
+	i64.load32_u	$push0=, 0($0):p2align=0
+	return  	$pop0
+	.endfunc
+.Lfunc_end11:
+	.size	ldi32_a1, .Lfunc_end11-ldi32_a1
+
+	.globl	ldi32_a2
+	.type	ldi32_a2,@function
+ldi32_a2:
+	.param  	i32
+	.result 	i64
+	i64.load32_u	$push0=, 0($0):p2align=1
+	return  	$pop0
+	.endfunc
+.Lfunc_end12:
+	.size	ldi32_a2, .Lfunc_end12-ldi32_a2
+
+	.globl	ldi32_a4
+	.type	ldi32_a4,@function
+ldi32_a4:
+	.param  	i32
+	.result 	i64
+	i64.load32_u	$push0=, 0($0)
+	return  	$pop0
+	.endfunc
+.Lfunc_end13:
+	.size	ldi32_a4, .Lfunc_end13-ldi32_a4
+
+	.globl	ldi32_a8
+	.type	ldi32_a8,@function
+ldi32_a8:
+	.param  	i32
+	.result 	i64
+	i64.load32_u	$push0=, 0($0):p2align=3
+	return  	$pop0
+	.endfunc
+.Lfunc_end14:
+	.size	ldi32_a8, .Lfunc_end14-ldi32_a8
+
+	.globl	sti64_a1
+	.type	sti64_a1,@function
+sti64_a1:
+	.param  	i32, i64
+	i64.store	$discard=, 0($0):p2align=0, $1
+	return
+	.endfunc
+.Lfunc_end15:
+	.size	sti64_a1, .Lfunc_end15-sti64_a1
+
+	.globl	sti64_a2
+	.type	sti64_a2,@function
+sti64_a2:
+	.param  	i32, i64
+	i64.store	$discard=, 0($0):p2align=1, $1
+	return
+	.endfunc
+.Lfunc_end16:
+	.size	sti64_a2, .Lfunc_end16-sti64_a2
+
+	.globl	sti64_a4
+	.type	sti64_a4,@function
+sti64_a4:
+	.param  	i32, i64
+	i64.store	$discard=, 0($0):p2align=2, $1
+	return
+	.endfunc
+.Lfunc_end17:
+	.size	sti64_a4, .Lfunc_end17-sti64_a4
+
+	.globl	sti64_a8
+	.type	sti64_a8,@function
+sti64_a8:
+	.param  	i32, i64
+	i64.store	$discard=, 0($0), $1
+	return
+	.endfunc
+.Lfunc_end18:
+	.size	sti64_a8, .Lfunc_end18-sti64_a8
+
+	.globl	sti64
+	.type	sti64,@function
+sti64:
+	.param  	i32, i64
+	i64.store	$discard=, 0($0), $1
+	return
+	.endfunc
+.Lfunc_end19:
+	.size	sti64, .Lfunc_end19-sti64
+
+	.globl	sti64_a16
+	.type	sti64_a16,@function
+sti64_a16:
+	.param  	i32, i64
+	i64.store	$discard=, 0($0):p2align=4, $1
+	return
+	.endfunc
+.Lfunc_end20:
+	.size	sti64_a16, .Lfunc_end20-sti64_a16
+
+	.globl	sti8_a1
+	.type	sti8_a1,@function
+sti8_a1:
+	.param  	i32, i64
+	i64.store8	$discard=, 0($0), $1
+	return
+	.endfunc
+.Lfunc_end21:
+	.size	sti8_a1, .Lfunc_end21-sti8_a1
+
+	.globl	sti8_a2
+	.type	sti8_a2,@function
+sti8_a2:
+	.param  	i32, i64
+	i64.store8	$discard=, 0($0):p2align=1, $1
+	return
+	.endfunc
+.Lfunc_end22:
+	.size	sti8_a2, .Lfunc_end22-sti8_a2
+
+	.globl	sti16_a1
+	.type	sti16_a1,@function
+sti16_a1:
+	.param  	i32, i64
+	i64.store16	$discard=, 0($0):p2align=0, $1
+	return
+	.endfunc
+.Lfunc_end23:
+	.size	sti16_a1, .Lfunc_end23-sti16_a1
+
+	.globl	sti16_a2
+	.type	sti16_a2,@function
+sti16_a2:
+	.param  	i32, i64
+	i64.store16	$discard=, 0($0), $1
+	return
+	.endfunc
+.Lfunc_end24:
+	.size	sti16_a2, .Lfunc_end24-sti16_a2
+
+	.globl	sti16_a4
+	.type	sti16_a4,@function
+sti16_a4:
+	.param  	i32, i64
+	i64.store16	$discard=, 0($0):p2align=2, $1
+	return
+	.endfunc
+.Lfunc_end25:
+	.size	sti16_a4, .Lfunc_end25-sti16_a4
+
+	.globl	sti32_a1
+	.type	sti32_a1,@function
+sti32_a1:
+	.param  	i32, i64
+	i64.store32	$discard=, 0($0):p2align=0, $1
+	return
+	.endfunc
+.Lfunc_end26:
+	.size	sti32_a1, .Lfunc_end26-sti32_a1
+
+	.globl	sti32_a2
+	.type	sti32_a2,@function
+sti32_a2:
+	.param  	i32, i64
+	i64.store32	$discard=, 0($0):p2align=1, $1
+	return
+	.endfunc
+.Lfunc_end27:
+	.size	sti32_a2, .Lfunc_end27-sti32_a2
+
+	.globl	sti32_a4
+	.type	sti32_a4,@function
+sti32_a4:
+	.param  	i32, i64
+	i64.store32	$discard=, 0($0), $1
+	return
+	.endfunc
+.Lfunc_end28:
+	.size	sti32_a4, .Lfunc_end28-sti32_a4
+
+	.globl	sti32_a8
+	.type	sti32_a8,@function
+sti32_a8:
+	.param  	i32, i64
+	i64.store32	$discard=, 0($0):p2align=3, $1
+	return
+	.endfunc
+.Lfunc_end29:
+	.size	sti32_a8, .Lfunc_end29-sti32_a8
+
+
diff --git a/test/llvm_autogenerated/i64-load-store-alignment.wast b/test/llvm_autogenerated/i64-load-store-alignment.wast
new file mode 100644
index 000000000..7f6903240
--- /dev/null
+++ b/test/llvm_autogenerated/i64-load-store-alignment.wast
@@ -0,0 +1,364 @@
+(module
+  (memory 0 4294967295)
+  (export "ldi64_a1" $ldi64_a1)
+  (export "ldi64_a2" $ldi64_a2)
+  (export "ldi64_a4" $ldi64_a4)
+  (export "ldi64_a8" $ldi64_a8)
+  (export "ldi64" $ldi64)
+  (export "ldi64_a16" $ldi64_a16)
+  (export "ldi8_a1" $ldi8_a1)
+  (export "ldi8_a2" $ldi8_a2)
+  (export "ldi16_a1" $ldi16_a1)
+  (export "ldi16_a2" $ldi16_a2)
+  (export "ldi16_a4" $ldi16_a4)
+  (export "ldi32_a1" $ldi32_a1)
+  (export "ldi32_a2" $ldi32_a2)
+  (export "ldi32_a4" $ldi32_a4)
+  (export "ldi32_a8" $ldi32_a8)
+  (export "sti64_a1" $sti64_a1)
+  (export "sti64_a2" $sti64_a2)
+  (export "sti64_a4" $sti64_a4)
+  (export "sti64_a8" $sti64_a8)
+  (export "sti64" $sti64)
+  (export "sti64_a16" $sti64_a16)
+  (export "sti8_a1" $sti8_a1)
+  (export "sti8_a2" $sti8_a2)
+  (export "sti16_a1" $sti16_a1)
+  (export "sti16_a2" $sti16_a2)
+  (export "sti16_a4" $sti16_a4)
+  (export "sti32_a1" $sti32_a1)
+  (export "sti32_a2" $sti32_a2)
+  (export "sti32_a4" $sti32_a4)
+  (export "sti32_a8" $sti32_a8)
+  (func $ldi64_a1 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load align=1
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi64_a2 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load align=2
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi64_a4 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load align=4
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi64_a8 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load align=8
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi64 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load align=8
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi64_a16 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load align=16
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi8_a1 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load8_u align=1
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi8_a2 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load8_u align=2
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi16_a1 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load16_u align=1
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi16_a2 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load16_u align=2
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi16_a4 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load16_u align=4
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi32_a1 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load align=1
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi32_a2 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load align=2
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi32_a4 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load align=4
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $ldi32_a8 (param $$0 i32) (result i64)
+    (block $fake_return_waka123
+      (block
+        (br $fake_return_waka123
+          (i64.load align=8
+            (get_local $$0)
+          )
+        )
+      )
+    )
+  )
+  (func $sti64_a1 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=1
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti64_a2 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=2
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti64_a4 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=4
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti64_a8 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=8
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti64 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=8
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti64_a16 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=16
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti8_a1 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=8
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti8_a2 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=2
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti16_a1 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=1
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti16_a2 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=16
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti16_a4 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=4
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti32_a1 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=1
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti32_a2 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=2
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti32_a4 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=32
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $sti32_a8 (param $$0 i32) (param $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (i64.store align=8
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+)
+;; METADATA: { "asmConsts": {},"staticBump": 4 }
diff --git a/test/llvm_autogenerated/i64.s b/test/llvm_autogenerated/i64.s
index a6a96e6f7..a095f0245 100644
--- a/test/llvm_autogenerated/i64.s
+++ b/test/llvm_autogenerated/i64.s
@@ -7,6 +7,7 @@ add64:
 	.result 	i64
 	i64.add 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	add64, .Lfunc_end0-add64
 
@@ -17,6 +18,7 @@ sub64:
 	.result 	i64
 	i64.sub 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	sub64, .Lfunc_end1-sub64
 
@@ -27,6 +29,7 @@ mul64:
 	.result 	i64
 	i64.mul 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	mul64, .Lfunc_end2-mul64
 
@@ -37,6 +40,7 @@ sdiv64:
 	.result 	i64
 	i64.div_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	sdiv64, .Lfunc_end3-sdiv64
 
@@ -47,6 +51,7 @@ udiv64:
 	.result 	i64
 	i64.div_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	udiv64, .Lfunc_end4-udiv64
 
@@ -57,6 +62,7 @@ srem64:
 	.result 	i64
 	i64.rem_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	srem64, .Lfunc_end5-srem64
 
@@ -67,6 +73,7 @@ urem64:
 	.result 	i64
 	i64.rem_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	urem64, .Lfunc_end6-urem64
 
@@ -77,6 +84,7 @@ and64:
 	.result 	i64
 	i64.and 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	and64, .Lfunc_end7-and64
 
@@ -87,6 +95,7 @@ or64:
 	.result 	i64
 	i64.or  	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	or64, .Lfunc_end8-or64
 
@@ -97,6 +106,7 @@ xor64:
 	.result 	i64
 	i64.xor 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end9:
 	.size	xor64, .Lfunc_end9-xor64
 
@@ -107,6 +117,7 @@ shl64:
 	.result 	i64
 	i64.shl 	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end10:
 	.size	shl64, .Lfunc_end10-shl64
 
@@ -117,6 +128,7 @@ shr64:
 	.result 	i64
 	i64.shr_u	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end11:
 	.size	shr64, .Lfunc_end11-shr64
 
@@ -127,6 +139,7 @@ sar64:
 	.result 	i64
 	i64.shr_s	$push0=, $0, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end12:
 	.size	sar64, .Lfunc_end12-sar64
 
@@ -137,6 +150,7 @@ clz64:
 	.result 	i64
 	i64.clz 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end13:
 	.size	clz64, .Lfunc_end13-clz64
 
@@ -147,6 +161,7 @@ clz64_zero_undef:
 	.result 	i64
 	i64.clz 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end14:
 	.size	clz64_zero_undef, .Lfunc_end14-clz64_zero_undef
 
@@ -157,6 +172,7 @@ ctz64:
 	.result 	i64
 	i64.ctz 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end15:
 	.size	ctz64, .Lfunc_end15-ctz64
 
@@ -167,6 +183,7 @@ ctz64_zero_undef:
 	.result 	i64
 	i64.ctz 	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end16:
 	.size	ctz64_zero_undef, .Lfunc_end16-ctz64_zero_undef
 
@@ -177,8 +194,8 @@ popcnt64:
 	.result 	i64
 	i64.popcnt	$push0=, $0
 	return  	$pop0
+	.endfunc
 .Lfunc_end17:
 	.size	popcnt64, .Lfunc_end17-popcnt64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/ident.s b/test/llvm_autogenerated/ident.s
index 46a08c7e7..782ec8812 100644
--- a/test/llvm_autogenerated/ident.s
+++ b/test/llvm_autogenerated/ident.s
@@ -2,4 +2,3 @@
 	.file	"/s/llvm/llvm/test/CodeGen/WebAssembly/ident.ll"
 
 	.ident	"hello world"
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/immediates.s b/test/llvm_autogenerated/immediates.s
index e1cf622a3..bba1d13e8 100644
--- a/test/llvm_autogenerated/immediates.s
+++ b/test/llvm_autogenerated/immediates.s
@@ -6,6 +6,7 @@ zero_i32:
 	.result 	i32
 	i32.const	$push0=, 0
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	zero_i32, .Lfunc_end0-zero_i32
 
@@ -15,6 +16,7 @@ one_i32:
 	.result 	i32
 	i32.const	$push0=, 1
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	one_i32, .Lfunc_end1-one_i32
 
@@ -24,6 +26,7 @@ max_i32:
 	.result 	i32
 	i32.const	$push0=, 2147483647
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	max_i32, .Lfunc_end2-max_i32
 
@@ -33,6 +36,7 @@ min_i32:
 	.result 	i32
 	i32.const	$push0=, -2147483648
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	min_i32, .Lfunc_end3-min_i32
 
@@ -42,6 +46,7 @@ zero_i64:
 	.result 	i64
 	i64.const	$push0=, 0
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	zero_i64, .Lfunc_end4-zero_i64
 
@@ -51,6 +56,7 @@ one_i64:
 	.result 	i64
 	i64.const	$push0=, 1
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	one_i64, .Lfunc_end5-one_i64
 
@@ -60,6 +66,7 @@ max_i64:
 	.result 	i64
 	i64.const	$push0=, 9223372036854775807
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	max_i64, .Lfunc_end6-max_i64
 
@@ -69,6 +76,7 @@ min_i64:
 	.result 	i64
 	i64.const	$push0=, -9223372036854775808
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	min_i64, .Lfunc_end7-min_i64
 
@@ -78,6 +86,7 @@ negzero_f32:
 	.result 	f32
 	f32.const	$push0=, -0x0p0
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	negzero_f32, .Lfunc_end8-negzero_f32
 
@@ -87,6 +96,7 @@ zero_f32:
 	.result 	f32
 	f32.const	$push0=, 0x0p0
 	return  	$pop0
+	.endfunc
 .Lfunc_end9:
 	.size	zero_f32, .Lfunc_end9-zero_f32
 
@@ -96,6 +106,7 @@ one_f32:
 	.result 	f32
 	f32.const	$push0=, 0x1p0
 	return  	$pop0
+	.endfunc
 .Lfunc_end10:
 	.size	one_f32, .Lfunc_end10-one_f32
 
@@ -105,6 +116,7 @@ two_f32:
 	.result 	f32
 	f32.const	$push0=, 0x1p1
 	return  	$pop0
+	.endfunc
 .Lfunc_end11:
 	.size	two_f32, .Lfunc_end11-two_f32
 
@@ -114,6 +126,7 @@ nan_f32:
 	.result 	f32
 	f32.const	$push0=, nan
 	return  	$pop0
+	.endfunc
 .Lfunc_end12:
 	.size	nan_f32, .Lfunc_end12-nan_f32
 
@@ -123,6 +136,7 @@ negnan_f32:
 	.result 	f32
 	f32.const	$push0=, -nan
 	return  	$pop0
+	.endfunc
 .Lfunc_end13:
 	.size	negnan_f32, .Lfunc_end13-negnan_f32
 
@@ -132,6 +146,7 @@ inf_f32:
 	.result 	f32
 	f32.const	$push0=, infinity
 	return  	$pop0
+	.endfunc
 .Lfunc_end14:
 	.size	inf_f32, .Lfunc_end14-inf_f32
 
@@ -141,6 +156,7 @@ neginf_f32:
 	.result 	f32
 	f32.const	$push0=, -infinity
 	return  	$pop0
+	.endfunc
 .Lfunc_end15:
 	.size	neginf_f32, .Lfunc_end15-neginf_f32
 
@@ -150,6 +166,7 @@ negzero_f64:
 	.result 	f64
 	f64.const	$push0=, -0x0p0
 	return  	$pop0
+	.endfunc
 .Lfunc_end16:
 	.size	negzero_f64, .Lfunc_end16-negzero_f64
 
@@ -159,6 +176,7 @@ zero_f64:
 	.result 	f64
 	f64.const	$push0=, 0x0p0
 	return  	$pop0
+	.endfunc
 .Lfunc_end17:
 	.size	zero_f64, .Lfunc_end17-zero_f64
 
@@ -168,6 +186,7 @@ one_f64:
 	.result 	f64
 	f64.const	$push0=, 0x1p0
 	return  	$pop0
+	.endfunc
 .Lfunc_end18:
 	.size	one_f64, .Lfunc_end18-one_f64
 
@@ -177,6 +196,7 @@ two_f64:
 	.result 	f64
 	f64.const	$push0=, 0x1p1
 	return  	$pop0
+	.endfunc
 .Lfunc_end19:
 	.size	two_f64, .Lfunc_end19-two_f64
 
@@ -186,6 +206,7 @@ nan_f64:
 	.result 	f64
 	f64.const	$push0=, nan
 	return  	$pop0
+	.endfunc
 .Lfunc_end20:
 	.size	nan_f64, .Lfunc_end20-nan_f64
 
@@ -195,6 +216,7 @@ negnan_f64:
 	.result 	f64
 	f64.const	$push0=, -nan
 	return  	$pop0
+	.endfunc
 .Lfunc_end21:
 	.size	negnan_f64, .Lfunc_end21-negnan_f64
 
@@ -204,6 +226,7 @@ inf_f64:
 	.result 	f64
 	f64.const	$push0=, infinity
 	return  	$pop0
+	.endfunc
 .Lfunc_end22:
 	.size	inf_f64, .Lfunc_end22-inf_f64
 
@@ -213,8 +236,8 @@ neginf_f64:
 	.result 	f64
 	f64.const	$push0=, -infinity
 	return  	$pop0
+	.endfunc
 .Lfunc_end23:
 	.size	neginf_f64, .Lfunc_end23-neginf_f64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/legalize.s b/test/llvm_autogenerated/legalize.s
index 53bd6c1ed..79855c4df 100644
--- a/test/llvm_autogenerated/legalize.s
+++ b/test/llvm_autogenerated/legalize.s
@@ -9,6 +9,7 @@ shl_i3:
 	i32.and 	$push1=, $1, $pop0
 	i32.shl 	$push2=, $0, $pop1
 	return  	$pop2
+	.endfunc
 .Lfunc_end0:
 	.size	shl_i3, .Lfunc_end0-shl_i3
 
@@ -21,6 +22,7 @@ shl_i53:
 	i64.and 	$push1=, $1, $pop0
 	i64.shl 	$push2=, $0, $pop1
 	return  	$pop2
+	.endfunc
 .Lfunc_end1:
 	.size	shl_i53, .Lfunc_end1-shl_i53
 
@@ -29,11 +31,12 @@ shl_i53:
 sext_in_reg_i32_i64:
 	.param  	i64
 	.result 	i64
-	.local  	i64
-	i64.const	$1=, 32
-	i64.shl 	$push0=, $0, $1
-	i64.shr_s	$push1=, $pop0, $1
-	return  	$pop1
+	i64.const	$push0=, 32
+	i64.shl 	$push1=, $0, $pop0
+	i64.const	$push3=, 32
+	i64.shr_s	$push2=, $pop1, $pop3
+	return  	$pop2
+	.endfunc
 .Lfunc_end2:
 	.size	sext_in_reg_i32_i64, .Lfunc_end2-sext_in_reg_i32_i64
 
@@ -45,6 +48,7 @@ fpext_f32_f64:
 	f32.load	$push0=, 0($0)
 	f64.promote/f32	$push1=, $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end3:
 	.size	fpext_f32_f64, .Lfunc_end3-fpext_f32_f64
 
@@ -56,6 +60,7 @@ fpconv_f64_f32:
 	f64.load	$push0=, 0($0)
 	f32.demote/f64	$push1=, $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end4:
 	.size	fpconv_f64_f32, .Lfunc_end4-fpconv_f64_f32
 
@@ -63,857 +68,1002 @@ fpconv_f64_f32:
 	.type	bigshift,@function
 bigshift:
 	.param  	i32, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64
-	.local  	i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i64, i32, i32, i64, i64, i64, i32, i32, i64, i32, i32, i64, i64, i64, i32, i32, i64, i64, i64, i32, i32, i64, i32, i32, i64, i32, i32, i32, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i64, i64, i64, i32, i64, i64, i32, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32
-	i32.const	$181=, __stack_pointer
-	i32.load	$181=, 0($181)
-	i32.const	$182=, 1024
-	i32.sub 	$279=, $181, $182
-	i32.const	$182=, __stack_pointer
-	i32.store	$279=, 0($182), $279
+	.local  	i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i32, i32, i64, i64, i64, i32, i32, i64, i32, i32, i64, i64, i64, i32, i32, i64, i64, i64, i32, i32, i64, i32, i32, i64, i32, i32, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i64, i64, i64, i32, i64, i64, i32, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32
+	i32.const	$179=, __stack_pointer
+	i32.load	$179=, 0($179)
+	i32.const	$180=, 1024
+	i32.sub 	$310=, $179, $180
+	i32.const	$180=, __stack_pointer
+	i32.store	$310=, 0($180), $310
 	i64.const	$push0=, 896
 	i64.sub 	$push1=, $pop0, $17
-	i32.const	$184=, 480
-	i32.add 	$184=, $279, $184
-	call    	__lshrti3@FUNCTION, $184, $1, $2, $pop1
+	i32.const	$182=, 480
+	i32.add 	$182=, $310, $182
+	call    	__lshrti3@FUNCTION, $182, $1, $2, $pop1
 	i64.const	$push2=, -768
 	i64.add 	$33=, $17, $pop2
-	i32.const	$185=, 464
-	i32.add 	$185=, $279, $185
-	call    	__ashlti3@FUNCTION, $185, $3, $4, $33
+	i32.const	$183=, 464
+	i32.add 	$183=, $310, $183
+	call    	__ashlti3@FUNCTION, $183, $3, $4, $33
 	i64.const	$push3=, -896
 	i64.add 	$push4=, $17, $pop3
-	i32.const	$186=, 496
-	i32.add 	$186=, $279, $186
-	call    	__ashlti3@FUNCTION, $186, $1, $2, $pop4
+	i32.const	$184=, 496
+	i32.add 	$184=, $310, $184
+	call    	__ashlti3@FUNCTION, $184, $1, $2, $pop4
 	i64.const	$push5=, 640
 	i64.sub 	$34=, $pop5, $17
-	i32.const	$187=, 352
-	i32.add 	$187=, $279, $187
-	call    	__lshrti3@FUNCTION, $187, $5, $6, $34
+	i32.const	$185=, 352
+	i32.add 	$185=, $310, $185
+	call    	__lshrti3@FUNCTION, $185, $5, $6, $34
 	i64.const	$push6=, -512
 	i64.add 	$35=, $17, $pop6
-	i32.const	$188=, 336
-	i32.add 	$188=, $279, $188
-	call    	__ashlti3@FUNCTION, $188, $7, $8, $35
+	i32.const	$186=, 336
+	i32.add 	$186=, $310, $186
+	call    	__ashlti3@FUNCTION, $186, $7, $8, $35
 	i64.const	$push7=, -640
 	i64.add 	$36=, $17, $pop7
-	i32.const	$189=, 368
-	i32.add 	$189=, $279, $189
-	call    	__ashlti3@FUNCTION, $189, $5, $6, $36
+	i32.const	$187=, 368
+	i32.add 	$187=, $310, $187
+	call    	__ashlti3@FUNCTION, $187, $5, $6, $36
 	i64.const	$push8=, 768
 	i64.sub 	$37=, $pop8, $17
-	i32.const	$190=, 432
-	i32.add 	$190=, $279, $190
-	call    	__lshrti3@FUNCTION, $190, $3, $4, $37
-	i64.const	$38=, 384
-	i64.sub 	$39=, $38, $17
-	i32.const	$191=, 864
-	i32.add 	$191=, $279, $191
-	call    	__lshrti3@FUNCTION, $191, $9, $10, $39
-	i64.const	$push9=, -256
-	i64.add 	$40=, $17, $pop9
-	i32.const	$192=, 848
-	i32.add 	$192=, $279, $192
-	call    	__ashlti3@FUNCTION, $192, $11, $12, $40
-	i64.const	$push10=, -384
-	i64.add 	$41=, $17, $pop10
-	i32.const	$193=, 880
-	i32.add 	$193=, $279, $193
-	call    	__ashlti3@FUNCTION, $193, $9, $10, $41
-	i32.const	$194=, 1008
-	i32.add 	$194=, $279, $194
-	call    	__ashlti3@FUNCTION, $194, $15, $16, $17
-	i64.const	$42=, 128
-	i64.sub 	$51=, $42, $17
-	i32.const	$195=, 960
-	i32.add 	$195=, $279, $195
-	call    	__lshrti3@FUNCTION, $195, $13, $14, $51
-	i64.const	$push11=, -128
-	i64.add 	$43=, $17, $pop11
-	i32.const	$196=, 976
-	i32.add 	$196=, $279, $196
-	call    	__ashlti3@FUNCTION, $196, $13, $14, $43
-	i64.const	$44=, 256
-	i64.sub 	$45=, $44, $17
-	i32.const	$197=, 816
-	i32.add 	$197=, $279, $197
-	call    	__lshrti3@FUNCTION, $197, $11, $12, $45
-	i64.const	$46=, 512
-	i64.sub 	$47=, $46, $17
-	i32.const	$198=, 240
-	i32.add 	$198=, $279, $198
-	call    	__lshrti3@FUNCTION, $198, $7, $8, $47
-	i32.const	$199=, 912
-	i32.add 	$199=, $279, $199
-	call    	__ashlti3@FUNCTION, $199, $11, $12, $17
-	i32.const	$200=, 928
-	i32.add 	$200=, $279, $200
-	call    	__lshrti3@FUNCTION, $200, $9, $10, $51
-	i32.const	$201=, 944
-	i32.add 	$201=, $279, $201
-	call    	__ashlti3@FUNCTION, $201, $9, $10, $43
-	i64.sub 	$48=, $44, $47
-	i32.const	$202=, 80
-	i32.add 	$202=, $279, $202
-	call    	__ashlti3@FUNCTION, $202, $7, $8, $48
-	i64.sub 	$push12=, $42, $48
-	i32.const	$203=, 96
-	i32.add 	$203=, $279, $203
-	call    	__lshrti3@FUNCTION, $203, $5, $6, $pop12
-	i64.sub 	$49=, $42, $47
-	i32.const	$204=, 112
-	i32.add 	$204=, $279, $204
-	call    	__ashlti3@FUNCTION, $204, $5, $6, $49
-	i32.const	$205=, 48
-	i32.add 	$205=, $279, $205
-	call    	__lshrti3@FUNCTION, $205, $3, $4, $47
-	i32.const	$206=, 176
-	i32.add 	$206=, $279, $206
-	call    	__lshrti3@FUNCTION, $206, $7, $8, $45
-	i32.const	$207=, 288
-	i32.add 	$207=, $279, $207
-	call    	__lshrti3@FUNCTION, $207, $1, $2, $34
-	i32.const	$208=, 272
-	i32.add 	$208=, $279, $208
-	call    	__ashlti3@FUNCTION, $208, $3, $4, $35
-	i32.const	$209=, 304
-	i32.add 	$209=, $279, $209
-	call    	__ashlti3@FUNCTION, $209, $1, $2, $36
-	i32.const	$210=, 128
-	i32.add 	$210=, $279, $210
-	call    	__lshrti3@FUNCTION, $210, $5, $6, $45
-	i64.sub 	$push13=, $38, $47
-	i32.const	$211=, 144
-	i32.add 	$211=, $279, $211
-	call    	__ashlti3@FUNCTION, $211, $7, $8, $pop13
-	i32.const	$212=, 160
-	i32.add 	$212=, $279, $212
-	call    	__lshrti3@FUNCTION, $212, $7, $8, $51
-	i32.const	$213=, 0
-	i32.add 	$213=, $279, $213
-	call    	__lshrti3@FUNCTION, $213, $1, $2, $47
-	i32.const	$214=, 16
-	i32.add 	$214=, $279, $214
-	call    	__ashlti3@FUNCTION, $214, $3, $4, $49
-	i32.const	$215=, 32
-	i32.add 	$215=, $279, $215
-	call    	__lshrti3@FUNCTION, $215, $3, $4, $39
-	i32.const	$216=, 64
-	i32.add 	$216=, $279, $216
-	call    	__ashlti3@FUNCTION, $216, $5, $6, $48
-	i32.const	$217=, 896
-	i32.add 	$217=, $279, $217
-	call    	__ashlti3@FUNCTION, $217, $9, $10, $17
-	i32.const	$218=, 256
-	i32.add 	$218=, $279, $218
-	call    	__ashlti3@FUNCTION, $218, $1, $2, $35
-	i32.const	$219=, 192
-	i32.add 	$219=, $279, $219
-	call    	__lshrti3@FUNCTION, $219, $5, $6, $47
-	i32.const	$220=, 208
-	i32.add 	$220=, $279, $220
-	call    	__ashlti3@FUNCTION, $220, $7, $8, $49
-	i32.const	$221=, 224
-	i32.add 	$221=, $279, $221
-	call    	__lshrti3@FUNCTION, $221, $7, $8, $39
-	i32.const	$222=, 768
-	i32.add 	$222=, $279, $222
-	call    	__lshrti3@FUNCTION, $222, $9, $10, $45
-	i64.sub 	$49=, $42, $45
-	i32.const	$223=, 784
-	i32.add 	$223=, $279, $223
-	call    	__ashlti3@FUNCTION, $223, $11, $12, $49
-	i32.const	$224=, 800
-	i32.add 	$224=, $279, $224
-	call    	__lshrti3@FUNCTION, $224, $11, $12, $51
-	i32.const	$225=, 992
-	i32.add 	$225=, $279, $225
-	call    	__ashlti3@FUNCTION, $225, $13, $14, $17
-	i32.const	$226=, 832
-	i32.add 	$226=, $279, $226
-	call    	__ashlti3@FUNCTION, $226, $9, $10, $40
-	i32.const	$227=, 384
-	i32.add 	$227=, $279, $227
-	call    	__lshrti3@FUNCTION, $227, $1, $2, $37
-	i64.sub 	$push14=, $42, $37
-	i32.const	$228=, 400
-	i32.add 	$228=, $279, $228
-	call    	__ashlti3@FUNCTION, $228, $3, $4, $pop14
-	i32.const	$229=, 416
-	i32.add 	$229=, $279, $229
-	call    	__lshrti3@FUNCTION, $229, $3, $4, $34
-	i32.const	$230=, 320
-	i32.add 	$230=, $279, $230
-	call    	__ashlti3@FUNCTION, $230, $5, $6, $35
-	i32.const	$231=, 448
-	i32.add 	$231=, $279, $231
-	call    	__ashlti3@FUNCTION, $231, $1, $2, $33
-	i32.const	$232=, 736
-	i32.add 	$232=, $279, $232
-	call    	__lshrti3@FUNCTION, $232, $1, $2, $39
-	i32.const	$233=, 720
-	i32.add 	$233=, $279, $233
-	call    	__ashlti3@FUNCTION, $233, $3, $4, $40
-	i32.const	$234=, 752
-	i32.add 	$234=, $279, $234
-	call    	__ashlti3@FUNCTION, $234, $1, $2, $41
-	i32.const	$235=, 592
-	i32.add 	$235=, $279, $235
-	call    	__ashlti3@FUNCTION, $235, $7, $8, $17
-	i32.const	$236=, 608
-	i32.add 	$236=, $279, $236
-	call    	__lshrti3@FUNCTION, $236, $5, $6, $51
-	i32.const	$237=, 624
-	i32.add 	$237=, $279, $237
-	call    	__ashlti3@FUNCTION, $237, $5, $6, $43
-	i32.const	$238=, 688
-	i32.add 	$238=, $279, $238
-	call    	__lshrti3@FUNCTION, $238, $3, $4, $45
-	i32.const	$239=, 640
-	i32.add 	$239=, $279, $239
-	call    	__lshrti3@FUNCTION, $239, $1, $2, $45
-	i32.const	$240=, 656
-	i32.add 	$240=, $279, $240
-	call    	__ashlti3@FUNCTION, $240, $3, $4, $49
-	i32.const	$241=, 672
-	i32.add 	$241=, $279, $241
-	call    	__lshrti3@FUNCTION, $241, $3, $4, $51
-	i32.const	$242=, 576
-	i32.add 	$242=, $279, $242
-	call    	__ashlti3@FUNCTION, $242, $5, $6, $17
-	i32.const	$243=, 704
-	i32.add 	$243=, $279, $243
-	call    	__ashlti3@FUNCTION, $243, $1, $2, $40
-	i32.const	$244=, 528
-	i32.add 	$244=, $279, $244
-	call    	__ashlti3@FUNCTION, $244, $3, $4, $17
-	i32.const	$245=, 544
-	i32.add 	$245=, $279, $245
-	call    	__lshrti3@FUNCTION, $245, $1, $2, $51
-	i32.const	$246=, 560
-	i32.add 	$246=, $279, $246
-	call    	__ashlti3@FUNCTION, $246, $1, $2, $43
-	i32.const	$247=, 512
-	i32.add 	$247=, $279, $247
-	call    	__ashlti3@FUNCTION, $247, $1, $2, $17
-	i32.const	$78=, 8
-	i32.const	$248=, 480
-	i32.add 	$248=, $279, $248
-	i32.add 	$push40=, $248, $78
-	i64.load	$39=, 0($pop40)
-	i32.const	$249=, 464
-	i32.add 	$249=, $279, $249
-	i32.add 	$push41=, $249, $78
-	i64.load	$43=, 0($pop41)
-	i64.load	$34=, 480($279)
-	i64.load	$49=, 464($279)
-	i64.load	$36=, 496($279)
-	i32.const	$250=, 496
-	i32.add 	$250=, $279, $250
-	i32.add 	$push43=, $250, $78
-	i64.load	$38=, 0($pop43)
-	i64.lt_u	$50=, $33, $42
-	i64.const	$51=, 0
-	i64.eq  	$52=, $33, $51
-	i32.const	$251=, 352
-	i32.add 	$251=, $279, $251
-	i32.add 	$push46=, $251, $78
-	i64.load	$33=, 0($pop46)
-	i32.const	$252=, 336
-	i32.add 	$252=, $279, $252
-	i32.add 	$push47=, $252, $78
-	i64.load	$41=, 0($pop47)
-	i32.const	$253=, 368
-	i32.add 	$253=, $279, $253
-	i32.add 	$push49=, $253, $78
-	i64.load	$79=, 0($pop49)
-	i32.const	$254=, 432
-	i32.add 	$254=, $279, $254
-	i32.add 	$push52=, $254, $78
-	i64.load	$80=, 0($pop52)
-	i32.const	$255=, 864
-	i32.add 	$255=, $279, $255
-	i32.add 	$push57=, $255, $78
-	i64.load	$81=, 0($pop57)
-	i32.const	$256=, 848
-	i32.add 	$256=, $279, $256
-	i32.add 	$push58=, $256, $78
-	i64.load	$82=, 0($pop58)
-	i32.const	$257=, 880
-	i32.add 	$257=, $279, $257
-	i32.add 	$push60=, $257, $78
-	i64.load	$83=, 0($pop60)
-	i32.const	$258=, 1008
-	i32.add 	$258=, $279, $258
-	i32.add 	$push63=, $258, $78
-	i64.load	$84=, 0($pop63)
-	i32.const	$259=, 960
-	i32.add 	$259=, $279, $259
-	i32.add 	$push64=, $259, $78
-	i64.load	$85=, 0($pop64)
-	i32.const	$260=, 976
-	i32.add 	$260=, $279, $260
-	i32.add 	$push66=, $260, $78
-	i64.load	$86=, 0($pop66)
-	i32.const	$261=, 816
-	i32.add 	$261=, $279, $261
-	i32.add 	$push69=, $261, $78
-	i64.load	$87=, 0($pop69)
-	i32.const	$262=, 240
-	i32.add 	$262=, $279, $262
-	i32.add 	$push74=, $262, $78
-	i64.load	$88=, 0($pop74)
-	i32.const	$263=, 912
-	i32.add 	$263=, $279, $263
-	i32.add 	$push99=, $263, $78
-	i64.load	$100=, 0($pop99)
-	i32.const	$264=, 928
-	i32.add 	$264=, $279, $264
-	i32.add 	$push100=, $264, $78
-	i64.load	$101=, 0($pop100)
-	i32.const	$265=, 944
-	i32.add 	$265=, $279, $265
-	i32.add 	$push102=, $265, $78
-	i64.load	$102=, 0($pop102)
-	i32.const	$266=, 80
-	i32.add 	$266=, $279, $266
-	i32.add 	$push106=, $266, $78
-	i64.load	$103=, 0($pop106)
-	i32.const	$267=, 96
-	i32.add 	$267=, $279, $267
-	i32.add 	$push107=, $267, $78
-	i64.load	$104=, 0($pop107)
-	i64.load	$90=, 80($279)
-	i64.load	$91=, 96($279)
-	i64.load	$92=, 112($279)
-	i32.const	$268=, 112
-	i32.add 	$268=, $279, $268
-	i32.add 	$push109=, $268, $78
-	i64.load	$105=, 0($pop109)
-	i64.lt_u	$89=, $48, $42
-	i64.eq  	$93=, $48, $51
-	i32.const	$269=, 48
-	i32.add 	$269=, $279, $269
-	i32.add 	$push112=, $269, $78
-	i64.load	$48=, 0($pop112)
-	i32.const	$270=, 176
-	i32.add 	$270=, $279, $270
-	i32.add 	$push115=, $270, $78
-	i64.load	$106=, 0($pop115)
-	i32.const	$271=, 288
-	i32.add 	$271=, $279, $271
-	i32.add 	$push120=, $271, $78
-	i64.load	$107=, 0($pop120)
-	i32.const	$272=, 272
-	i32.add 	$272=, $279, $272
-	i32.add 	$push121=, $272, $78
-	i64.load	$108=, 0($pop121)
-	i32.const	$273=, 304
-	i32.add 	$273=, $279, $273
-	i32.add 	$push123=, $273, $78
-	i64.load	$109=, 0($pop123)
-	i32.const	$274=, 128
-	i32.add 	$274=, $279, $274
-	i32.add 	$push146=, $274, $78
-	i64.load	$119=, 0($pop146)
-	i32.const	$275=, 144
-	i32.add 	$275=, $279, $275
-	i32.add 	$push147=, $275, $78
-	i64.load	$120=, 0($pop147)
-	i32.const	$276=, 160
-	i32.add 	$276=, $279, $276
-	i32.add 	$push149=, $276, $78
-	i64.load	$121=, 0($pop149)
-	i32.const	$277=, 0
-	i32.add 	$277=, $279, $277
-	i32.add 	$push152=, $277, $78
-	i64.load	$122=, 0($pop152)
-	i32.const	$278=, 16
-	i32.add 	$278=, $279, $278
-	i32.add 	$push153=, $278, $78
-	i64.load	$123=, 0($pop153)
-	i64.load	$117=, 64($279)
-	i32.const	$279=, 64
-	i32.add 	$279=, $279, $279
-	i32.add 	$push158=, $279, $78
-	i64.load	$125=, 0($pop158)
-	i32.const	$280=, 32
-	i32.add 	$280=, $279, $280
-	i32.add 	$push155=, $280, $78
-	i64.load	$124=, 0($pop155)
-	i32.const	$281=, 896
-	i32.add 	$281=, $279, $281
-	i32.add 	$push163=, $281, $78
-	i64.load	$126=, 0($pop163)
-	i32.const	$282=, 256
-	i32.add 	$282=, $279, $282
-	i32.add 	$push167=, $282, $78
-	i64.load	$127=, 0($pop167)
-	i32.const	$283=, 192
-	i32.add 	$283=, $279, $283
-	i32.add 	$push195=, $283, $78
-	i64.load	$142=, 0($pop195)
-	i32.const	$284=, 208
-	i32.add 	$284=, $279, $284
-	i32.add 	$push196=, $284, $78
-	i64.load	$143=, 0($pop196)
-	i64.load	$114=, 0($279)
-	i64.load	$115=, 16($279)
-	i64.load	$116=, 32($279)
-	i32.const	$285=, 224
-	i32.add 	$285=, $279, $285
-	i32.add 	$push198=, $285, $78
-	i64.load	$144=, 0($pop198)
-	i64.lt_u	$74=, $47, $42
-	i64.load	$128=, 192($279)
-	i64.load	$129=, 208($279)
-	i64.load	$130=, 224($279)
-	i64.load	$75=, 240($279)
-	i64.load	$94=, 48($279)
-	i64.load	$95=, 176($279)
-	i64.lt_u	$71=, $45, $42
-	i64.lt_u	$76=, $47, $44
-	i64.load	$110=, 128($279)
-	i64.load	$111=, 144($279)
-	i64.load	$112=, 160($279)
-	i64.eq  	$96=, $47, $51
-	i64.eq  	$113=, $45, $51
-	i32.const	$286=, 768
-	i32.add 	$286=, $279, $286
-	i32.add 	$push202=, $286, $78
-	i64.load	$47=, 0($pop202)
-	i32.const	$287=, 784
-	i32.add 	$287=, $279, $287
-	i32.add 	$push203=, $287, $78
-	i64.load	$45=, 0($pop203)
-	i32.const	$288=, 800
-	i32.add 	$288=, $279, $288
-	i32.add 	$push205=, $288, $78
-	i64.load	$145=, 0($pop205)
-	i32.const	$289=, 992
-	i32.add 	$289=, $279, $289
-	i32.add 	$push208=, $289, $78
-	i64.load	$146=, 0($pop208)
-	i32.const	$290=, 832
-	i32.add 	$290=, $279, $290
-	i32.add 	$push211=, $290, $78
-	i64.load	$147=, 0($pop211)
-	i32.const	$291=, 384
-	i32.add 	$291=, $279, $291
-	i32.add 	$push216=, $291, $78
-	i64.load	$148=, 0($pop216)
-	i32.const	$292=, 400
-	i32.add 	$292=, $279, $292
-	i32.add 	$push217=, $292, $78
-	i64.load	$149=, 0($pop217)
-	i64.load	$136=, 384($279)
-	i64.load	$137=, 400($279)
-	i32.const	$293=, 416
-	i32.add 	$293=, $279, $293
-	i32.add 	$push219=, $293, $78
-	i64.load	$150=, 0($pop219)
-	i64.load	$138=, 416($279)
-	i64.load	$59=, 432($279)
-	i64.lt_u	$58=, $37, $42
-	i64.eq  	$139=, $37, $51
-	i32.const	$294=, 320
-	i32.add 	$294=, $279, $294
-	i32.add 	$push222=, $294, $78
-	i64.load	$37=, 0($pop222)
-	i64.load	$54=, 352($279)
-	i64.load	$55=, 336($279)
-	i64.load	$56=, 368($279)
-	i32.const	$295=, 448
-	i32.add 	$295=, $279, $295
-	i32.add 	$push225=, $295, $78
-	i64.load	$151=, 0($pop225)
-	i64.lt_u	$53=, $35, $42
-	i64.load	$97=, 288($279)
-	i64.load	$98=, 272($279)
-	i64.load	$99=, 304($279)
-	i64.load	$118=, 256($279)
-	i64.load	$140=, 320($279)
-	i64.load	$141=, 448($279)
-	i64.eq  	$57=, $35, $51
-	i64.lt_u	$60=, $35, $44
-	i32.const	$296=, 736
-	i32.add 	$296=, $279, $296
-	i32.add 	$push242=, $296, $78
-	i64.load	$35=, 0($pop242)
-	i32.const	$297=, 720
-	i32.add 	$297=, $279, $297
-	i32.add 	$push243=, $297, $78
-	i64.load	$159=, 0($pop243)
-	i64.load	$62=, 864($279)
-	i64.load	$63=, 848($279)
-	i64.load	$64=, 880($279)
-	i32.const	$298=, 752
-	i32.add 	$298=, $279, $298
-	i32.add 	$push245=, $298, $78
-	i64.load	$160=, 0($pop245)
-	i64.lt_u	$61=, $40, $42
-	i64.load	$152=, 736($279)
-	i64.load	$153=, 720($279)
-	i64.load	$154=, 752($279)
-	i64.eq  	$65=, $40, $51
-	i32.const	$299=, 592
-	i32.add 	$299=, $279, $299
-	i32.add 	$push248=, $299, $78
-	i64.load	$40=, 0($pop248)
-	i32.const	$300=, 608
-	i32.add 	$300=, $279, $300
-	i32.add 	$push249=, $300, $78
-	i64.load	$161=, 0($pop249)
-	i32.const	$301=, 624
-	i32.add 	$301=, $279, $301
-	i32.add 	$push251=, $301, $78
-	i64.load	$162=, 0($pop251)
-	i32.const	$302=, 688
-	i32.add 	$302=, $279, $302
-	i32.add 	$push254=, $302, $78
-	i64.load	$163=, 0($pop254)
-	i32.const	$303=, 640
-	i32.add 	$303=, $279, $303
-	i32.add 	$push269=, $303, $78
-	i64.load	$169=, 0($pop269)
-	i32.const	$304=, 656
-	i32.add 	$304=, $279, $304
-	i32.add 	$push270=, $304, $78
-	i64.load	$170=, 0($pop270)
-	i64.load	$131=, 768($279)
-	i64.load	$132=, 784($279)
-	i64.load	$133=, 800($279)
-	i32.const	$305=, 672
-	i32.add 	$305=, $279, $305
-	i32.add 	$push272=, $305, $78
-	i64.load	$171=, 0($pop272)
-	i64.load	$164=, 640($279)
-	i64.load	$165=, 656($279)
-	i64.load	$166=, 672($279)
-	i64.load	$72=, 816($279)
-	i64.load	$158=, 688($279)
-	i32.const	$306=, 576
-	i32.add 	$306=, $279, $306
-	i32.add 	$push275=, $306, $78
-	i64.load	$172=, 0($pop275)
-	i64.load	$135=, 832($279)
-	i64.load	$168=, 704($279)
-	i32.const	$307=, 704
-	i32.add 	$307=, $279, $307
-	i32.add 	$push278=, $307, $78
-	i64.load	$173=, 0($pop278)
-	i32.const	$308=, 528
-	i32.add 	$308=, $279, $308
-	i32.add 	$push288=, $308, $78
-	i64.load	$177=, 0($pop288)
-	i32.const	$309=, 544
-	i32.add 	$309=, $279, $309
-	i32.add 	$push289=, $309, $78
-	i64.load	$178=, 0($pop289)
-	i32.const	$310=, 560
-	i32.add 	$310=, $279, $310
-	i32.add 	$push291=, $310, $78
-	i64.load	$179=, 0($pop291)
-	i64.load	$67=, 1008($279)
-	i64.load	$68=, 960($279)
-	i64.load	$69=, 976($279)
-	i64.lt_u	$66=, $17, $42
-	i64.eq  	$70=, $17, $51
-	i64.lt_u	$73=, $17, $44
-	i64.lt_u	$77=, $17, $46
-	i64.load	$17=, 912($279)
-	i64.load	$42=, 928($279)
-	i64.load	$44=, 944($279)
-	i64.load	$46=, 896($279)
-	i64.load	$134=, 992($279)
-	i64.load	$155=, 592($279)
-	i64.load	$156=, 608($279)
-	i64.load	$157=, 624($279)
-	i64.load	$167=, 576($279)
-	i64.load	$174=, 528($279)
-	i64.load	$175=, 544($279)
-	i64.load	$176=, 560($279)
-	i64.load	$180=, 512($279)
-	i32.add 	$push304=, $0, $78
-	i32.const	$311=, 512
-	i32.add 	$311=, $279, $311
-	i32.add 	$push299=, $311, $78
-	i64.load	$push300=, 0($pop299)
-	i64.select	$push301=, $66, $pop300, $51
-	i64.select	$push302=, $73, $pop301, $51
-	i64.select	$push303=, $77, $pop302, $51
-	i64.store	$discard=, 0($pop304), $pop303
-	i64.select	$push296=, $66, $180, $51
-	i64.select	$push297=, $73, $pop296, $51
-	i64.select	$push298=, $77, $pop297, $51
-	i64.store	$discard=, 0($0), $pop298
-	i32.const	$push305=, 24
-	i32.add 	$push306=, $0, $pop305
-	i64.or  	$push290=, $177, $178
-	i64.select	$push292=, $66, $pop290, $179
-	i64.select	$push293=, $70, $4, $pop292
-	i64.select	$push294=, $73, $pop293, $51
-	i64.select	$push295=, $77, $pop294, $51
-	i64.store	$discard=, 0($pop306), $pop295
-	i32.const	$push307=, 16
-	i32.add 	$push308=, $0, $pop307
-	i64.or  	$push283=, $174, $175
-	i64.select	$push284=, $66, $pop283, $176
-	i64.select	$push285=, $70, $3, $pop284
-	i64.select	$push286=, $73, $pop285, $51
-	i64.select	$push287=, $77, $pop286, $51
-	i64.store	$discard=, 0($pop308), $pop287
-	i32.const	$push309=, 56
-	i32.add 	$push310=, $0, $pop309
-	i64.or  	$push250=, $40, $161
-	i64.select	$push252=, $66, $pop250, $162
-	i64.select	$push253=, $70, $8, $pop252
-	i64.select	$push255=, $71, $163, $51
-	i64.or  	$push256=, $pop253, $pop255
-	i64.or  	$push244=, $159, $35
-	i64.select	$push246=, $61, $pop244, $160
-	i64.select	$push247=, $65, $4, $pop246
-	i64.select	$push257=, $73, $pop256, $pop247
-	i64.select	$push258=, $70, $8, $pop257
-	i64.select	$push259=, $77, $pop258, $51
-	i64.store	$discard=, 0($pop310), $pop259
-	i32.const	$push311=, 48
+	i32.const	$188=, 432
+	i32.add 	$188=, $310, $188
+	call    	__lshrti3@FUNCTION, $188, $3, $4, $37
+	i64.const	$push9=, 384
+	i64.sub 	$38=, $pop9, $17
+	i32.const	$189=, 864
+	i32.add 	$189=, $310, $189
+	call    	__lshrti3@FUNCTION, $189, $9, $10, $38
+	i64.const	$push10=, -256
+	i64.add 	$39=, $17, $pop10
+	i32.const	$190=, 848
+	i32.add 	$190=, $310, $190
+	call    	__ashlti3@FUNCTION, $190, $11, $12, $39
+	i64.const	$push11=, -384
+	i64.add 	$40=, $17, $pop11
+	i32.const	$191=, 880
+	i32.add 	$191=, $310, $191
+	call    	__ashlti3@FUNCTION, $191, $9, $10, $40
+	i32.const	$192=, 1008
+	i32.add 	$192=, $310, $192
+	call    	__ashlti3@FUNCTION, $192, $15, $16, $17
+	i64.const	$push12=, 128
+	i64.sub 	$41=, $pop12, $17
+	i32.const	$193=, 960
+	i32.add 	$193=, $310, $193
+	call    	__lshrti3@FUNCTION, $193, $13, $14, $41
+	i64.const	$push13=, -128
+	i64.add 	$42=, $17, $pop13
+	i32.const	$194=, 976
+	i32.add 	$194=, $310, $194
+	call    	__ashlti3@FUNCTION, $194, $13, $14, $42
+	i64.const	$push14=, 256
+	i64.sub 	$43=, $pop14, $17
+	i32.const	$195=, 816
+	i32.add 	$195=, $310, $195
+	call    	__lshrti3@FUNCTION, $195, $11, $12, $43
+	i64.const	$push15=, 512
+	i64.sub 	$44=, $pop15, $17
+	i32.const	$196=, 240
+	i32.add 	$196=, $310, $196
+	call    	__lshrti3@FUNCTION, $196, $7, $8, $44
+	i32.const	$197=, 912
+	i32.add 	$197=, $310, $197
+	call    	__ashlti3@FUNCTION, $197, $11, $12, $17
+	i32.const	$198=, 928
+	i32.add 	$198=, $310, $198
+	call    	__lshrti3@FUNCTION, $198, $9, $10, $41
+	i32.const	$199=, 944
+	i32.add 	$199=, $310, $199
+	call    	__ashlti3@FUNCTION, $199, $9, $10, $42
+	i64.const	$push483=, 256
+	i64.sub 	$45=, $pop483, $44
+	i32.const	$200=, 80
+	i32.add 	$200=, $310, $200
+	call    	__ashlti3@FUNCTION, $200, $7, $8, $45
+	i64.const	$push482=, 128
+	i64.sub 	$push16=, $pop482, $45
+	i32.const	$201=, 96
+	i32.add 	$201=, $310, $201
+	call    	__lshrti3@FUNCTION, $201, $5, $6, $pop16
+	i64.const	$push481=, 128
+	i64.sub 	$46=, $pop481, $44
+	i32.const	$202=, 112
+	i32.add 	$202=, $310, $202
+	call    	__ashlti3@FUNCTION, $202, $5, $6, $46
+	i32.const	$203=, 48
+	i32.add 	$203=, $310, $203
+	call    	__lshrti3@FUNCTION, $203, $3, $4, $44
+	i32.const	$204=, 176
+	i32.add 	$204=, $310, $204
+	call    	__lshrti3@FUNCTION, $204, $7, $8, $43
+	i32.const	$205=, 288
+	i32.add 	$205=, $310, $205
+	call    	__lshrti3@FUNCTION, $205, $1, $2, $34
+	i32.const	$206=, 272
+	i32.add 	$206=, $310, $206
+	call    	__ashlti3@FUNCTION, $206, $3, $4, $35
+	i32.const	$207=, 304
+	i32.add 	$207=, $310, $207
+	call    	__ashlti3@FUNCTION, $207, $1, $2, $36
+	i32.const	$208=, 128
+	i32.add 	$208=, $310, $208
+	call    	__lshrti3@FUNCTION, $208, $5, $6, $43
+	i64.const	$push480=, 384
+	i64.sub 	$push17=, $pop480, $44
+	i32.const	$209=, 144
+	i32.add 	$209=, $310, $209
+	call    	__ashlti3@FUNCTION, $209, $7, $8, $pop17
+	i32.const	$210=, 160
+	i32.add 	$210=, $310, $210
+	call    	__lshrti3@FUNCTION, $210, $7, $8, $41
+	i32.const	$211=, 0
+	i32.add 	$211=, $310, $211
+	call    	__lshrti3@FUNCTION, $211, $1, $2, $44
+	i32.const	$212=, 16
+	i32.add 	$212=, $310, $212
+	call    	__ashlti3@FUNCTION, $212, $3, $4, $46
+	i32.const	$213=, 32
+	i32.add 	$213=, $310, $213
+	call    	__lshrti3@FUNCTION, $213, $3, $4, $38
+	i32.const	$214=, 64
+	i32.add 	$214=, $310, $214
+	call    	__ashlti3@FUNCTION, $214, $5, $6, $45
+	i32.const	$215=, 896
+	i32.add 	$215=, $310, $215
+	call    	__ashlti3@FUNCTION, $215, $9, $10, $17
+	i32.const	$216=, 256
+	i32.add 	$216=, $310, $216
+	call    	__ashlti3@FUNCTION, $216, $1, $2, $35
+	i32.const	$217=, 192
+	i32.add 	$217=, $310, $217
+	call    	__lshrti3@FUNCTION, $217, $5, $6, $44
+	i32.const	$218=, 208
+	i32.add 	$218=, $310, $218
+	call    	__ashlti3@FUNCTION, $218, $7, $8, $46
+	i32.const	$219=, 224
+	i32.add 	$219=, $310, $219
+	call    	__lshrti3@FUNCTION, $219, $7, $8, $38
+	i32.const	$220=, 768
+	i32.add 	$220=, $310, $220
+	call    	__lshrti3@FUNCTION, $220, $9, $10, $43
+	i64.const	$push479=, 128
+	i64.sub 	$46=, $pop479, $43
+	i32.const	$221=, 784
+	i32.add 	$221=, $310, $221
+	call    	__ashlti3@FUNCTION, $221, $11, $12, $46
+	i32.const	$222=, 800
+	i32.add 	$222=, $310, $222
+	call    	__lshrti3@FUNCTION, $222, $11, $12, $41
+	i32.const	$223=, 992
+	i32.add 	$223=, $310, $223
+	call    	__ashlti3@FUNCTION, $223, $13, $14, $17
+	i32.const	$224=, 832
+	i32.add 	$224=, $310, $224
+	call    	__ashlti3@FUNCTION, $224, $9, $10, $39
+	i32.const	$225=, 384
+	i32.add 	$225=, $310, $225
+	call    	__lshrti3@FUNCTION, $225, $1, $2, $37
+	i64.const	$push478=, 128
+	i64.sub 	$push18=, $pop478, $37
+	i32.const	$226=, 400
+	i32.add 	$226=, $310, $226
+	call    	__ashlti3@FUNCTION, $226, $3, $4, $pop18
+	i32.const	$227=, 416
+	i32.add 	$227=, $310, $227
+	call    	__lshrti3@FUNCTION, $227, $3, $4, $34
+	i32.const	$228=, 320
+	i32.add 	$228=, $310, $228
+	call    	__ashlti3@FUNCTION, $228, $5, $6, $35
+	i32.const	$229=, 448
+	i32.add 	$229=, $310, $229
+	call    	__ashlti3@FUNCTION, $229, $1, $2, $33
+	i32.const	$230=, 736
+	i32.add 	$230=, $310, $230
+	call    	__lshrti3@FUNCTION, $230, $1, $2, $38
+	i32.const	$231=, 720
+	i32.add 	$231=, $310, $231
+	call    	__ashlti3@FUNCTION, $231, $3, $4, $39
+	i32.const	$232=, 752
+	i32.add 	$232=, $310, $232
+	call    	__ashlti3@FUNCTION, $232, $1, $2, $40
+	i32.const	$233=, 592
+	i32.add 	$233=, $310, $233
+	call    	__ashlti3@FUNCTION, $233, $7, $8, $17
+	i32.const	$234=, 608
+	i32.add 	$234=, $310, $234
+	call    	__lshrti3@FUNCTION, $234, $5, $6, $41
+	i32.const	$235=, 624
+	i32.add 	$235=, $310, $235
+	call    	__ashlti3@FUNCTION, $235, $5, $6, $42
+	i32.const	$236=, 688
+	i32.add 	$236=, $310, $236
+	call    	__lshrti3@FUNCTION, $236, $3, $4, $43
+	i32.const	$237=, 640
+	i32.add 	$237=, $310, $237
+	call    	__lshrti3@FUNCTION, $237, $1, $2, $43
+	i32.const	$238=, 656
+	i32.add 	$238=, $310, $238
+	call    	__ashlti3@FUNCTION, $238, $3, $4, $46
+	i32.const	$239=, 672
+	i32.add 	$239=, $310, $239
+	call    	__lshrti3@FUNCTION, $239, $3, $4, $41
+	i32.const	$240=, 576
+	i32.add 	$240=, $310, $240
+	call    	__ashlti3@FUNCTION, $240, $5, $6, $17
+	i32.const	$241=, 704
+	i32.add 	$241=, $310, $241
+	call    	__ashlti3@FUNCTION, $241, $1, $2, $39
+	i32.const	$242=, 528
+	i32.add 	$242=, $310, $242
+	call    	__ashlti3@FUNCTION, $242, $3, $4, $17
+	i32.const	$243=, 544
+	i32.add 	$243=, $310, $243
+	call    	__lshrti3@FUNCTION, $243, $1, $2, $41
+	i32.const	$244=, 560
+	i32.add 	$244=, $310, $244
+	call    	__ashlti3@FUNCTION, $244, $1, $2, $42
+	i32.const	$245=, 512
+	i32.add 	$245=, $310, $245
+	call    	__ashlti3@FUNCTION, $245, $1, $2, $17
+	i32.const	$push45=, 8
+	i32.const	$246=, 480
+	i32.add 	$246=, $310, $246
+	i32.add 	$push46=, $246, $pop45
+	i64.load	$41=, 0($pop46)
+	i32.const	$push477=, 8
+	i32.const	$247=, 464
+	i32.add 	$247=, $310, $247
+	i32.add 	$push47=, $247, $pop477
+	i64.load	$38=, 0($pop47)
+	i64.load	$42=, 480($310)
+	i64.load	$34=, 464($310)
+	i64.load	$46=, 496($310)
+	i32.const	$push476=, 8
+	i32.const	$248=, 496
+	i32.add 	$248=, $310, $248
+	i32.add 	$push49=, $248, $pop476
+	i64.load	$36=, 0($pop49)
+	i64.const	$push475=, 128
+	i64.lt_u	$47=, $33, $pop475
+	i64.const	$push21=, 0
+	i64.eq  	$48=, $33, $pop21
+	i32.const	$push474=, 8
+	i32.const	$249=, 352
+	i32.add 	$249=, $310, $249
+	i32.add 	$push52=, $249, $pop474
+	i64.load	$33=, 0($pop52)
+	i32.const	$push473=, 8
+	i32.const	$250=, 336
+	i32.add 	$250=, $310, $250
+	i32.add 	$push53=, $250, $pop473
+	i64.load	$40=, 0($pop53)
+	i32.const	$push472=, 8
+	i32.const	$251=, 368
+	i32.add 	$251=, $310, $251
+	i32.add 	$push55=, $251, $pop472
+	i64.load	$74=, 0($pop55)
+	i32.const	$push471=, 8
+	i32.const	$252=, 432
+	i32.add 	$252=, $310, $252
+	i32.add 	$push58=, $252, $pop471
+	i64.load	$75=, 0($pop58)
+	i32.const	$push470=, 8
+	i32.const	$253=, 864
+	i32.add 	$253=, $310, $253
+	i32.add 	$push63=, $253, $pop470
+	i64.load	$76=, 0($pop63)
+	i32.const	$push469=, 8
+	i32.const	$254=, 848
+	i32.add 	$254=, $310, $254
+	i32.add 	$push64=, $254, $pop469
+	i64.load	$77=, 0($pop64)
+	i32.const	$push468=, 8
+	i32.const	$255=, 880
+	i32.add 	$255=, $310, $255
+	i32.add 	$push66=, $255, $pop468
+	i64.load	$78=, 0($pop66)
+	i32.const	$push467=, 8
+	i32.const	$256=, 1008
+	i32.add 	$256=, $310, $256
+	i32.add 	$push69=, $256, $pop467
+	i64.load	$79=, 0($pop69)
+	i32.const	$push466=, 8
+	i32.const	$257=, 960
+	i32.add 	$257=, $310, $257
+	i32.add 	$push70=, $257, $pop466
+	i64.load	$80=, 0($pop70)
+	i32.const	$push465=, 8
+	i32.const	$258=, 976
+	i32.add 	$258=, $310, $258
+	i32.add 	$push72=, $258, $pop465
+	i64.load	$81=, 0($pop72)
+	i32.const	$push464=, 8
+	i32.const	$259=, 816
+	i32.add 	$259=, $310, $259
+	i32.add 	$push75=, $259, $pop464
+	i64.load	$82=, 0($pop75)
+	i32.const	$push463=, 8
+	i32.const	$260=, 240
+	i32.add 	$260=, $310, $260
+	i32.add 	$push80=, $260, $pop463
+	i64.load	$83=, 0($pop80)
+	i32.const	$push462=, 8
+	i32.const	$261=, 912
+	i32.add 	$261=, $310, $261
+	i32.add 	$push105=, $261, $pop462
+	i64.load	$97=, 0($pop105)
+	i32.const	$push461=, 8
+	i32.const	$262=, 928
+	i32.add 	$262=, $310, $262
+	i32.add 	$push106=, $262, $pop461
+	i64.load	$98=, 0($pop106)
+	i32.const	$push460=, 8
+	i32.const	$263=, 944
+	i32.add 	$263=, $310, $263
+	i32.add 	$push108=, $263, $pop460
+	i64.load	$99=, 0($pop108)
+	i32.const	$push459=, 8
+	i32.const	$264=, 80
+	i32.add 	$264=, $310, $264
+	i32.add 	$push112=, $264, $pop459
+	i64.load	$100=, 0($pop112)
+	i32.const	$push458=, 8
+	i32.const	$265=, 96
+	i32.add 	$265=, $310, $265
+	i32.add 	$push113=, $265, $pop458
+	i64.load	$101=, 0($pop113)
+	i64.load	$87=, 80($310)
+	i64.load	$88=, 96($310)
+	i64.load	$89=, 112($310)
+	i32.const	$push457=, 8
+	i32.const	$266=, 112
+	i32.add 	$266=, $310, $266
+	i32.add 	$push115=, $266, $pop457
+	i64.load	$102=, 0($pop115)
+	i64.const	$push456=, 128
+	i64.lt_u	$86=, $45, $pop456
+	i64.const	$push455=, 0
+	i64.eq  	$90=, $45, $pop455
+	i32.const	$push454=, 8
+	i32.const	$267=, 48
+	i32.add 	$267=, $310, $267
+	i32.add 	$push118=, $267, $pop454
+	i64.load	$45=, 0($pop118)
+	i32.const	$push453=, 8
+	i32.const	$268=, 176
+	i32.add 	$268=, $310, $268
+	i32.add 	$push121=, $268, $pop453
+	i64.load	$103=, 0($pop121)
+	i32.const	$push452=, 8
+	i32.const	$269=, 288
+	i32.add 	$269=, $310, $269
+	i32.add 	$push126=, $269, $pop452
+	i64.load	$104=, 0($pop126)
+	i32.const	$push451=, 8
+	i32.const	$270=, 272
+	i32.add 	$270=, $310, $270
+	i32.add 	$push127=, $270, $pop451
+	i64.load	$105=, 0($pop127)
+	i32.const	$push450=, 8
+	i32.const	$271=, 304
+	i32.add 	$271=, $310, $271
+	i32.add 	$push129=, $271, $pop450
+	i64.load	$106=, 0($pop129)
+	i32.const	$push449=, 8
+	i32.const	$272=, 128
+	i32.add 	$272=, $310, $272
+	i32.add 	$push152=, $272, $pop449
+	i64.load	$117=, 0($pop152)
+	i32.const	$push448=, 8
+	i32.const	$273=, 144
+	i32.add 	$273=, $310, $273
+	i32.add 	$push153=, $273, $pop448
+	i64.load	$118=, 0($pop153)
+	i32.const	$push447=, 8
+	i32.const	$274=, 160
+	i32.add 	$274=, $310, $274
+	i32.add 	$push155=, $274, $pop447
+	i64.load	$119=, 0($pop155)
+	i32.const	$push446=, 8
+	i32.const	$275=, 0
+	i32.add 	$275=, $310, $275
+	i32.add 	$push158=, $275, $pop446
+	i64.load	$120=, 0($pop158)
+	i32.const	$push445=, 8
+	i32.const	$276=, 16
+	i32.add 	$276=, $310, $276
+	i32.add 	$push159=, $276, $pop445
+	i64.load	$121=, 0($pop159)
+	i64.load	$114=, 64($310)
+	i32.const	$push444=, 8
+	i32.const	$277=, 64
+	i32.add 	$277=, $310, $277
+	i32.add 	$push164=, $277, $pop444
+	i64.load	$123=, 0($pop164)
+	i32.const	$push443=, 8
+	i32.const	$278=, 32
+	i32.add 	$278=, $310, $278
+	i32.add 	$push161=, $278, $pop443
+	i64.load	$122=, 0($pop161)
+	i32.const	$push442=, 8
+	i32.const	$279=, 896
+	i32.add 	$279=, $310, $279
+	i32.add 	$push169=, $279, $pop442
+	i64.load	$124=, 0($pop169)
+	i32.const	$push441=, 8
+	i32.const	$280=, 256
+	i32.add 	$280=, $310, $280
+	i32.add 	$push173=, $280, $pop441
+	i64.load	$125=, 0($pop173)
+	i32.const	$push440=, 8
+	i32.const	$281=, 192
+	i32.add 	$281=, $310, $281
+	i32.add 	$push201=, $281, $pop440
+	i64.load	$140=, 0($pop201)
+	i32.const	$push439=, 8
+	i32.const	$282=, 208
+	i32.add 	$282=, $310, $282
+	i32.add 	$push202=, $282, $pop439
+	i64.load	$141=, 0($pop202)
+	i64.load	$111=, 0($310)
+	i64.load	$112=, 16($310)
+	i64.load	$113=, 32($310)
+	i32.const	$push438=, 8
+	i32.const	$283=, 224
+	i32.add 	$283=, $310, $283
+	i32.add 	$push204=, $283, $pop438
+	i64.load	$142=, 0($pop204)
+	i64.const	$push437=, 128
+	i64.lt_u	$70=, $44, $pop437
+	i64.load	$126=, 192($310)
+	i64.load	$127=, 208($310)
+	i64.load	$128=, 224($310)
+	i64.load	$71=, 240($310)
+	i64.load	$91=, 48($310)
+	i64.load	$92=, 176($310)
+	i64.const	$push436=, 128
+	i64.lt_u	$67=, $43, $pop436
+	i64.const	$push435=, 256
+	i64.lt_u	$72=, $44, $pop435
+	i64.load	$107=, 128($310)
+	i64.load	$108=, 144($310)
+	i64.load	$109=, 160($310)
+	i64.const	$push434=, 0
+	i64.eq  	$93=, $44, $pop434
+	i64.const	$push433=, 0
+	i64.eq  	$110=, $43, $pop433
+	i32.const	$push432=, 8
+	i32.const	$284=, 768
+	i32.add 	$284=, $310, $284
+	i32.add 	$push208=, $284, $pop432
+	i64.load	$44=, 0($pop208)
+	i32.const	$push431=, 8
+	i32.const	$285=, 784
+	i32.add 	$285=, $310, $285
+	i32.add 	$push209=, $285, $pop431
+	i64.load	$43=, 0($pop209)
+	i32.const	$push430=, 8
+	i32.const	$286=, 800
+	i32.add 	$286=, $310, $286
+	i32.add 	$push211=, $286, $pop430
+	i64.load	$143=, 0($pop211)
+	i32.const	$push429=, 8
+	i32.const	$287=, 992
+	i32.add 	$287=, $310, $287
+	i32.add 	$push214=, $287, $pop429
+	i64.load	$144=, 0($pop214)
+	i32.const	$push428=, 8
+	i32.const	$288=, 832
+	i32.add 	$288=, $310, $288
+	i32.add 	$push217=, $288, $pop428
+	i64.load	$145=, 0($pop217)
+	i32.const	$push427=, 8
+	i32.const	$289=, 384
+	i32.add 	$289=, $310, $289
+	i32.add 	$push222=, $289, $pop427
+	i64.load	$146=, 0($pop222)
+	i32.const	$push426=, 8
+	i32.const	$290=, 400
+	i32.add 	$290=, $310, $290
+	i32.add 	$push223=, $290, $pop426
+	i64.load	$147=, 0($pop223)
+	i64.load	$134=, 384($310)
+	i64.load	$135=, 400($310)
+	i32.const	$push425=, 8
+	i32.const	$291=, 416
+	i32.add 	$291=, $310, $291
+	i32.add 	$push225=, $291, $pop425
+	i64.load	$148=, 0($pop225)
+	i64.load	$136=, 416($310)
+	i64.load	$55=, 432($310)
+	i64.const	$push424=, 128
+	i64.lt_u	$54=, $37, $pop424
+	i64.const	$push423=, 0
+	i64.eq  	$137=, $37, $pop423
+	i32.const	$push422=, 8
+	i32.const	$292=, 320
+	i32.add 	$292=, $310, $292
+	i32.add 	$push228=, $292, $pop422
+	i64.load	$37=, 0($pop228)
+	i64.load	$50=, 352($310)
+	i64.load	$51=, 336($310)
+	i64.load	$52=, 368($310)
+	i32.const	$push421=, 8
+	i32.const	$293=, 448
+	i32.add 	$293=, $310, $293
+	i32.add 	$push231=, $293, $pop421
+	i64.load	$149=, 0($pop231)
+	i64.const	$push420=, 128
+	i64.lt_u	$49=, $35, $pop420
+	i64.load	$94=, 288($310)
+	i64.load	$95=, 272($310)
+	i64.load	$96=, 304($310)
+	i64.load	$116=, 256($310)
+	i64.load	$138=, 320($310)
+	i64.load	$139=, 448($310)
+	i64.const	$push419=, 0
+	i64.eq  	$53=, $35, $pop419
+	i64.const	$push418=, 256
+	i64.lt_u	$56=, $35, $pop418
+	i32.const	$push417=, 8
+	i32.const	$294=, 736
+	i32.add 	$294=, $310, $294
+	i32.add 	$push248=, $294, $pop417
+	i64.load	$35=, 0($pop248)
+	i32.const	$push416=, 8
+	i32.const	$295=, 720
+	i32.add 	$295=, $310, $295
+	i32.add 	$push249=, $295, $pop416
+	i64.load	$157=, 0($pop249)
+	i64.load	$58=, 864($310)
+	i64.load	$59=, 848($310)
+	i64.load	$60=, 880($310)
+	i32.const	$push415=, 8
+	i32.const	$296=, 752
+	i32.add 	$296=, $310, $296
+	i32.add 	$push251=, $296, $pop415
+	i64.load	$158=, 0($pop251)
+	i64.const	$push414=, 128
+	i64.lt_u	$57=, $39, $pop414
+	i64.load	$150=, 736($310)
+	i64.load	$151=, 720($310)
+	i64.load	$152=, 752($310)
+	i64.const	$push413=, 0
+	i64.eq  	$61=, $39, $pop413
+	i32.const	$push412=, 8
+	i32.const	$297=, 592
+	i32.add 	$297=, $310, $297
+	i32.add 	$push254=, $297, $pop412
+	i64.load	$39=, 0($pop254)
+	i32.const	$push411=, 8
+	i32.const	$298=, 608
+	i32.add 	$298=, $310, $298
+	i32.add 	$push255=, $298, $pop411
+	i64.load	$159=, 0($pop255)
+	i32.const	$push410=, 8
+	i32.const	$299=, 624
+	i32.add 	$299=, $310, $299
+	i32.add 	$push257=, $299, $pop410
+	i64.load	$160=, 0($pop257)
+	i32.const	$push409=, 8
+	i32.const	$300=, 688
+	i32.add 	$300=, $310, $300
+	i32.add 	$push260=, $300, $pop409
+	i64.load	$161=, 0($pop260)
+	i32.const	$push408=, 8
+	i32.const	$301=, 640
+	i32.add 	$301=, $310, $301
+	i32.add 	$push275=, $301, $pop408
+	i64.load	$167=, 0($pop275)
+	i32.const	$push407=, 8
+	i32.const	$302=, 656
+	i32.add 	$302=, $310, $302
+	i32.add 	$push276=, $302, $pop407
+	i64.load	$168=, 0($pop276)
+	i64.load	$129=, 768($310)
+	i64.load	$130=, 784($310)
+	i64.load	$131=, 800($310)
+	i32.const	$push406=, 8
+	i32.const	$303=, 672
+	i32.add 	$303=, $310, $303
+	i32.add 	$push278=, $303, $pop406
+	i64.load	$169=, 0($pop278)
+	i64.load	$162=, 640($310)
+	i64.load	$163=, 656($310)
+	i64.load	$164=, 672($310)
+	i64.load	$68=, 816($310)
+	i64.load	$156=, 688($310)
+	i32.const	$push405=, 8
+	i32.const	$304=, 576
+	i32.add 	$304=, $310, $304
+	i32.add 	$push281=, $304, $pop405
+	i64.load	$170=, 0($pop281)
+	i64.load	$133=, 832($310)
+	i64.load	$166=, 704($310)
+	i32.const	$push404=, 8
+	i32.const	$305=, 704
+	i32.add 	$305=, $310, $305
+	i32.add 	$push284=, $305, $pop404
+	i64.load	$171=, 0($pop284)
+	i32.const	$push403=, 8
+	i32.const	$306=, 528
+	i32.add 	$306=, $310, $306
+	i32.add 	$push294=, $306, $pop403
+	i64.load	$175=, 0($pop294)
+	i32.const	$push402=, 8
+	i32.const	$307=, 544
+	i32.add 	$307=, $310, $307
+	i32.add 	$push295=, $307, $pop402
+	i64.load	$176=, 0($pop295)
+	i32.const	$push401=, 8
+	i32.const	$308=, 560
+	i32.add 	$308=, $310, $308
+	i32.add 	$push297=, $308, $pop401
+	i64.load	$177=, 0($pop297)
+	i64.load	$63=, 1008($310)
+	i64.load	$64=, 960($310)
+	i64.load	$65=, 976($310)
+	i64.const	$push400=, 128
+	i64.lt_u	$62=, $17, $pop400
+	i64.const	$push399=, 0
+	i64.eq  	$66=, $17, $pop399
+	i64.const	$push398=, 256
+	i64.lt_u	$69=, $17, $pop398
+	i64.const	$push397=, 512
+	i64.lt_u	$73=, $17, $pop397
+	i64.load	$17=, 912($310)
+	i64.load	$84=, 928($310)
+	i64.load	$85=, 944($310)
+	i64.load	$115=, 896($310)
+	i64.load	$132=, 992($310)
+	i64.load	$153=, 592($310)
+	i64.load	$154=, 608($310)
+	i64.load	$155=, 624($310)
+	i64.load	$165=, 576($310)
+	i64.load	$172=, 528($310)
+	i64.load	$173=, 544($310)
+	i64.load	$174=, 560($310)
+	i64.load	$178=, 512($310)
+	i32.const	$push396=, 8
+	i32.add 	$push310=, $0, $pop396
+	i32.const	$push395=, 8
+	i32.const	$309=, 512
+	i32.add 	$309=, $310, $309
+	i32.add 	$push305=, $309, $pop395
+	i64.load	$push306=, 0($pop305)
+	i64.const	$push394=, 0
+	i64.select	$push307=, $62, $pop306, $pop394
+	i64.const	$push393=, 0
+	i64.select	$push308=, $69, $pop307, $pop393
+	i64.const	$push392=, 0
+	i64.select	$push309=, $73, $pop308, $pop392
+	i64.store	$discard=, 0($pop310), $pop309
+	i64.const	$push391=, 0
+	i64.select	$push302=, $62, $178, $pop391
+	i64.const	$push390=, 0
+	i64.select	$push303=, $69, $pop302, $pop390
+	i64.const	$push389=, 0
+	i64.select	$push304=, $73, $pop303, $pop389
+	i64.store	$discard=, 0($0), $pop304
+	i32.const	$push311=, 24
 	i32.add 	$push312=, $0, $pop311
-	i64.or  	$push234=, $155, $156
-	i64.select	$push235=, $66, $pop234, $157
-	i64.select	$push236=, $70, $7, $pop235
-	i64.select	$push237=, $71, $158, $51
-	i64.or  	$push238=, $pop236, $pop237
-	i64.or  	$push231=, $153, $152
-	i64.select	$push232=, $61, $pop231, $154
-	i64.select	$push233=, $65, $3, $pop232
-	i64.select	$push239=, $73, $pop238, $pop233
-	i64.select	$push240=, $70, $7, $pop239
-	i64.select	$push241=, $77, $pop240, $51
-	i64.store	$discard=, 0($pop312), $pop241
-	i32.const	$push313=, 40
+	i64.or  	$push296=, $175, $176
+	i64.select	$push298=, $62, $pop296, $177
+	i64.select	$push299=, $66, $4, $pop298
+	i64.const	$push388=, 0
+	i64.select	$push300=, $69, $pop299, $pop388
+	i64.const	$push387=, 0
+	i64.select	$push301=, $73, $pop300, $pop387
+	i64.store	$discard=, 0($pop312), $pop301
+	i32.const	$push313=, 16
 	i32.add 	$push314=, $0, $pop313
-	i64.select	$push276=, $66, $172, $51
-	i64.or  	$push271=, $169, $170
-	i64.select	$push273=, $71, $pop271, $171
-	i64.select	$push274=, $113, $2, $pop273
-	i64.or  	$push277=, $pop276, $pop274
-	i64.select	$push279=, $61, $173, $51
-	i64.select	$push280=, $73, $pop277, $pop279
-	i64.select	$push281=, $70, $6, $pop280
-	i64.select	$push282=, $77, $pop281, $51
-	i64.store	$discard=, 0($pop314), $pop282
-	i32.const	$push315=, 32
+	i64.or  	$push289=, $172, $173
+	i64.select	$push290=, $62, $pop289, $174
+	i64.select	$push291=, $66, $3, $pop290
+	i64.const	$push386=, 0
+	i64.select	$push292=, $69, $pop291, $pop386
+	i64.const	$push385=, 0
+	i64.select	$push293=, $73, $pop292, $pop385
+	i64.store	$discard=, 0($pop314), $pop293
+	i32.const	$push315=, 56
 	i32.add 	$push316=, $0, $pop315
-	i64.select	$push263=, $66, $167, $51
-	i64.or  	$push260=, $164, $165
-	i64.select	$push261=, $71, $pop260, $166
-	i64.select	$push262=, $113, $1, $pop261
-	i64.or  	$push264=, $pop263, $pop262
-	i64.select	$push265=, $61, $168, $51
-	i64.select	$push266=, $73, $pop264, $pop265
-	i64.select	$push267=, $70, $5, $pop266
-	i64.select	$push268=, $77, $pop267, $51
-	i64.store	$discard=, 0($pop316), $pop268
-	i32.const	$push317=, 120
+	i64.or  	$push256=, $39, $159
+	i64.select	$push258=, $62, $pop256, $160
+	i64.select	$push259=, $66, $8, $pop258
+	i64.const	$push384=, 0
+	i64.select	$push261=, $67, $161, $pop384
+	i64.or  	$push262=, $pop259, $pop261
+	i64.or  	$push250=, $157, $35
+	i64.select	$push252=, $57, $pop250, $158
+	i64.select	$push253=, $61, $4, $pop252
+	i64.select	$push263=, $69, $pop262, $pop253
+	i64.select	$push264=, $66, $8, $pop263
+	i64.const	$push383=, 0
+	i64.select	$push265=, $73, $pop264, $pop383
+	i64.store	$discard=, 0($pop316), $pop265
+	i32.const	$push317=, 48
 	i32.add 	$push318=, $0, $pop317
-	i64.or  	$push65=, $84, $85
-	i64.select	$push67=, $66, $pop65, $86
-	i64.select	$push68=, $70, $16, $pop67
-	i64.select	$push70=, $71, $87, $51
-	i64.or  	$push71=, $pop68, $pop70
-	i64.or  	$push59=, $82, $81
-	i64.select	$push61=, $61, $pop59, $83
-	i64.select	$push62=, $65, $12, $pop61
-	i64.select	$push72=, $73, $pop71, $pop62
-	i64.select	$push73=, $70, $16, $pop72
-	i64.select	$push75=, $74, $88, $51
-	i64.select	$push76=, $76, $pop75, $51
-	i64.or  	$push77=, $pop73, $pop76
-	i64.or  	$push48=, $41, $33
-	i64.select	$push50=, $53, $pop48, $79
-	i64.select	$push51=, $57, $8, $pop50
-	i64.select	$push53=, $58, $80, $51
-	i64.or  	$push54=, $pop51, $pop53
-	i64.or  	$push42=, $43, $39
-	i64.select	$push44=, $50, $pop42, $38
-	i64.select	$push45=, $52, $4, $pop44
-	i64.select	$push55=, $60, $pop54, $pop45
-	i64.select	$push56=, $57, $8, $pop55
-	i64.select	$push78=, $77, $pop77, $pop56
-	i64.select	$push79=, $70, $16, $pop78
-	i64.store	$discard=, 0($pop318), $pop79
-	i32.const	$push319=, 112
+	i64.or  	$push240=, $153, $154
+	i64.select	$push241=, $62, $pop240, $155
+	i64.select	$push242=, $66, $7, $pop241
+	i64.const	$push382=, 0
+	i64.select	$push243=, $67, $156, $pop382
+	i64.or  	$push244=, $pop242, $pop243
+	i64.or  	$push237=, $151, $150
+	i64.select	$push238=, $57, $pop237, $152
+	i64.select	$push239=, $61, $3, $pop238
+	i64.select	$push245=, $69, $pop244, $pop239
+	i64.select	$push246=, $66, $7, $pop245
+	i64.const	$push381=, 0
+	i64.select	$push247=, $73, $pop246, $pop381
+	i64.store	$discard=, 0($pop318), $pop247
+	i32.const	$push319=, 40
 	i32.add 	$push320=, $0, $pop319
-	i64.or  	$push28=, $67, $68
-	i64.select	$push29=, $66, $pop28, $69
-	i64.select	$push30=, $70, $15, $pop29
-	i64.select	$push31=, $71, $72, $51
-	i64.or  	$push32=, $pop30, $pop31
-	i64.or  	$push25=, $63, $62
-	i64.select	$push26=, $61, $pop25, $64
-	i64.select	$push27=, $65, $11, $pop26
-	i64.select	$push33=, $73, $pop32, $pop27
-	i64.select	$push34=, $70, $15, $pop33
-	i64.select	$push35=, $74, $75, $51
-	i64.select	$push36=, $76, $pop35, $51
-	i64.or  	$push37=, $pop34, $pop36
-	i64.or  	$push18=, $55, $54
-	i64.select	$push19=, $53, $pop18, $56
-	i64.select	$push20=, $57, $7, $pop19
-	i64.select	$push21=, $58, $59, $51
-	i64.or  	$push22=, $pop20, $pop21
-	i64.or  	$push15=, $49, $34
-	i64.select	$push16=, $50, $pop15, $36
-	i64.select	$push17=, $52, $3, $pop16
-	i64.select	$push23=, $60, $pop22, $pop17
-	i64.select	$push24=, $57, $7, $pop23
-	i64.select	$push38=, $77, $pop37, $pop24
-	i64.select	$push39=, $70, $15, $pop38
-	i64.store	$discard=, 0($pop320), $pop39
-	i32.const	$push321=, 104
+	i64.const	$push380=, 0
+	i64.select	$push282=, $62, $170, $pop380
+	i64.or  	$push277=, $167, $168
+	i64.select	$push279=, $67, $pop277, $169
+	i64.select	$push280=, $110, $2, $pop279
+	i64.or  	$push283=, $pop282, $pop280
+	i64.const	$push379=, 0
+	i64.select	$push285=, $57, $171, $pop379
+	i64.select	$push286=, $69, $pop283, $pop285
+	i64.select	$push287=, $66, $6, $pop286
+	i64.const	$push378=, 0
+	i64.select	$push288=, $73, $pop287, $pop378
+	i64.store	$discard=, 0($pop320), $pop288
+	i32.const	$push321=, 32
 	i32.add 	$push322=, $0, $pop321
-	i64.select	$push209=, $66, $146, $51
-	i64.or  	$push204=, $47, $45
-	i64.select	$push206=, $71, $pop204, $145
-	i64.select	$push207=, $113, $10, $pop206
-	i64.or  	$push210=, $pop209, $pop207
-	i64.select	$push212=, $61, $147, $51
-	i64.select	$push213=, $73, $pop210, $pop212
-	i64.select	$push214=, $70, $14, $pop213
-	i64.or  	$push197=, $142, $143
-	i64.select	$push199=, $74, $pop197, $144
-	i64.select	$push200=, $96, $6, $pop199
-	i64.select	$push201=, $76, $pop200, $51
-	i64.or  	$push215=, $pop214, $pop201
-	i64.select	$push223=, $53, $37, $51
-	i64.or  	$push218=, $148, $149
-	i64.select	$push220=, $58, $pop218, $150
-	i64.select	$push221=, $139, $2, $pop220
-	i64.or  	$push224=, $pop223, $pop221
-	i64.select	$push226=, $50, $151, $51
-	i64.select	$push227=, $60, $pop224, $pop226
-	i64.select	$push228=, $57, $6, $pop227
-	i64.select	$push229=, $77, $pop215, $pop228
-	i64.select	$push230=, $70, $14, $pop229
-	i64.store	$discard=, 0($pop322), $pop230
-	i32.const	$push323=, 96
+	i64.const	$push377=, 0
+	i64.select	$push269=, $62, $165, $pop377
+	i64.or  	$push266=, $162, $163
+	i64.select	$push267=, $67, $pop266, $164
+	i64.select	$push268=, $110, $1, $pop267
+	i64.or  	$push270=, $pop269, $pop268
+	i64.const	$push376=, 0
+	i64.select	$push271=, $57, $166, $pop376
+	i64.select	$push272=, $69, $pop270, $pop271
+	i64.select	$push273=, $66, $5, $pop272
+	i64.const	$push375=, 0
+	i64.select	$push274=, $73, $pop273, $pop375
+	i64.store	$discard=, 0($pop322), $pop274
+	i32.const	$push323=, 120
 	i32.add 	$push324=, $0, $pop323
-	i64.select	$push179=, $66, $134, $51
-	i64.or  	$push176=, $131, $132
-	i64.select	$push177=, $71, $pop176, $133
-	i64.select	$push178=, $113, $9, $pop177
-	i64.or  	$push180=, $pop179, $pop178
-	i64.select	$push181=, $61, $135, $51
-	i64.select	$push182=, $73, $pop180, $pop181
-	i64.select	$push183=, $70, $13, $pop182
-	i64.or  	$push172=, $128, $129
-	i64.select	$push173=, $74, $pop172, $130
-	i64.select	$push174=, $96, $5, $pop173
-	i64.select	$push175=, $76, $pop174, $51
-	i64.or  	$push184=, $pop183, $pop175
-	i64.select	$push188=, $53, $140, $51
-	i64.or  	$push185=, $136, $137
-	i64.select	$push186=, $58, $pop185, $138
-	i64.select	$push187=, $139, $1, $pop186
-	i64.or  	$push189=, $pop188, $pop187
-	i64.select	$push190=, $50, $141, $51
-	i64.select	$push191=, $60, $pop189, $pop190
-	i64.select	$push192=, $57, $5, $pop191
-	i64.select	$push193=, $77, $pop184, $pop192
-	i64.select	$push194=, $70, $13, $pop193
-	i64.store	$discard=, 0($pop324), $pop194
-	i32.const	$push325=, 72
+	i64.or  	$push71=, $79, $80
+	i64.select	$push73=, $62, $pop71, $81
+	i64.select	$push74=, $66, $16, $pop73
+	i64.const	$push374=, 0
+	i64.select	$push76=, $67, $82, $pop374
+	i64.or  	$push77=, $pop74, $pop76
+	i64.or  	$push65=, $77, $76
+	i64.select	$push67=, $57, $pop65, $78
+	i64.select	$push68=, $61, $12, $pop67
+	i64.select	$push78=, $69, $pop77, $pop68
+	i64.select	$push79=, $66, $16, $pop78
+	i64.const	$push373=, 0
+	i64.select	$push81=, $70, $83, $pop373
+	i64.const	$push372=, 0
+	i64.select	$push82=, $72, $pop81, $pop372
+	i64.or  	$push83=, $pop79, $pop82
+	i64.or  	$push54=, $40, $33
+	i64.select	$push56=, $49, $pop54, $74
+	i64.select	$push57=, $53, $8, $pop56
+	i64.const	$push371=, 0
+	i64.select	$push59=, $54, $75, $pop371
+	i64.or  	$push60=, $pop57, $pop59
+	i64.or  	$push48=, $38, $41
+	i64.select	$push50=, $47, $pop48, $36
+	i64.select	$push51=, $48, $4, $pop50
+	i64.select	$push61=, $56, $pop60, $pop51
+	i64.select	$push62=, $53, $8, $pop61
+	i64.select	$push84=, $73, $pop83, $pop62
+	i64.select	$push85=, $66, $16, $pop84
+	i64.store	$discard=, 0($pop324), $pop85
+	i32.const	$push325=, 112
 	i32.add 	$push326=, $0, $pop325
-	i64.select	$push164=, $66, $126, $51
-	i64.select	$push165=, $73, $pop164, $51
-	i64.or  	$push154=, $122, $123
-	i64.select	$push156=, $74, $pop154, $124
-	i64.select	$push157=, $96, $2, $pop156
-	i64.select	$push159=, $89, $125, $51
-	i64.or  	$push160=, $pop157, $pop159
-	i64.or  	$push148=, $119, $120
-	i64.select	$push150=, $71, $pop148, $121
-	i64.select	$push151=, $113, $6, $pop150
-	i64.select	$push161=, $76, $pop160, $pop151
-	i64.select	$push162=, $96, $2, $pop161
-	i64.or  	$push166=, $pop165, $pop162
-	i64.select	$push168=, $53, $127, $51
-	i64.select	$push169=, $60, $pop168, $51
-	i64.select	$push170=, $77, $pop166, $pop169
-	i64.select	$push171=, $70, $10, $pop170
-	i64.store	$discard=, 0($pop326), $pop171
-	i32.const	$push327=, 64
+	i64.or  	$push33=, $63, $64
+	i64.select	$push34=, $62, $pop33, $65
+	i64.select	$push35=, $66, $15, $pop34
+	i64.const	$push370=, 0
+	i64.select	$push36=, $67, $68, $pop370
+	i64.or  	$push37=, $pop35, $pop36
+	i64.or  	$push30=, $59, $58
+	i64.select	$push31=, $57, $pop30, $60
+	i64.select	$push32=, $61, $11, $pop31
+	i64.select	$push38=, $69, $pop37, $pop32
+	i64.select	$push39=, $66, $15, $pop38
+	i64.const	$push369=, 0
+	i64.select	$push40=, $70, $71, $pop369
+	i64.const	$push368=, 0
+	i64.select	$push41=, $72, $pop40, $pop368
+	i64.or  	$push42=, $pop39, $pop41
+	i64.or  	$push23=, $51, $50
+	i64.select	$push24=, $49, $pop23, $52
+	i64.select	$push25=, $53, $7, $pop24
+	i64.const	$push367=, 0
+	i64.select	$push26=, $54, $55, $pop367
+	i64.or  	$push27=, $pop25, $pop26
+	i64.or  	$push19=, $34, $42
+	i64.select	$push20=, $47, $pop19, $46
+	i64.select	$push22=, $48, $3, $pop20
+	i64.select	$push28=, $56, $pop27, $pop22
+	i64.select	$push29=, $53, $7, $pop28
+	i64.select	$push43=, $73, $pop42, $pop29
+	i64.select	$push44=, $66, $15, $pop43
+	i64.store	$discard=, 0($pop326), $pop44
+	i32.const	$push327=, 104
 	i32.add 	$push328=, $0, $pop327
-	i64.select	$push139=, $66, $46, $51
-	i64.select	$push140=, $73, $pop139, $51
-	i64.or  	$push132=, $114, $115
-	i64.select	$push133=, $74, $pop132, $116
-	i64.select	$push134=, $96, $1, $pop133
-	i64.select	$push135=, $89, $117, $51
-	i64.or  	$push136=, $pop134, $pop135
-	i64.or  	$push129=, $110, $111
-	i64.select	$push130=, $71, $pop129, $112
-	i64.select	$push131=, $113, $5, $pop130
-	i64.select	$push137=, $76, $pop136, $pop131
-	i64.select	$push138=, $96, $1, $pop137
-	i64.or  	$push141=, $pop140, $pop138
-	i64.select	$push142=, $53, $118, $51
-	i64.select	$push143=, $60, $pop142, $51
-	i64.select	$push144=, $77, $pop141, $pop143
-	i64.select	$push145=, $70, $9, $pop144
-	i64.store	$discard=, 0($pop328), $pop145
-	i32.const	$push329=, 88
+	i64.const	$push366=, 0
+	i64.select	$push215=, $62, $144, $pop366
+	i64.or  	$push210=, $44, $43
+	i64.select	$push212=, $67, $pop210, $143
+	i64.select	$push213=, $110, $10, $pop212
+	i64.or  	$push216=, $pop215, $pop213
+	i64.const	$push365=, 0
+	i64.select	$push218=, $57, $145, $pop365
+	i64.select	$push219=, $69, $pop216, $pop218
+	i64.select	$push220=, $66, $14, $pop219
+	i64.or  	$push203=, $140, $141
+	i64.select	$push205=, $70, $pop203, $142
+	i64.select	$push206=, $93, $6, $pop205
+	i64.const	$push364=, 0
+	i64.select	$push207=, $72, $pop206, $pop364
+	i64.or  	$push221=, $pop220, $pop207
+	i64.const	$push363=, 0
+	i64.select	$push229=, $49, $37, $pop363
+	i64.or  	$push224=, $146, $147
+	i64.select	$push226=, $54, $pop224, $148
+	i64.select	$push227=, $137, $2, $pop226
+	i64.or  	$push230=, $pop229, $pop227
+	i64.const	$push362=, 0
+	i64.select	$push232=, $47, $149, $pop362
+	i64.select	$push233=, $56, $pop230, $pop232
+	i64.select	$push234=, $53, $6, $pop233
+	i64.select	$push235=, $73, $pop221, $pop234
+	i64.select	$push236=, $66, $14, $pop235
+	i64.store	$discard=, 0($pop328), $pop236
+	i32.const	$push329=, 96
 	i32.add 	$push330=, $0, $pop329
-	i64.or  	$push101=, $100, $101
-	i64.select	$push103=, $66, $pop101, $102
-	i64.select	$push104=, $70, $12, $pop103
-	i64.select	$push105=, $73, $pop104, $51
-	i64.select	$push113=, $74, $48, $51
-	i64.or  	$push108=, $103, $104
-	i64.select	$push110=, $89, $pop108, $105
-	i64.select	$push111=, $93, $8, $pop110
-	i64.or  	$push114=, $pop113, $pop111
-	i64.select	$push116=, $71, $106, $51
-	i64.select	$push117=, $76, $pop114, $pop116
-	i64.select	$push118=, $96, $4, $pop117
-	i64.or  	$push119=, $pop105, $pop118
-	i64.or  	$push122=, $108, $107
-	i64.select	$push124=, $53, $pop122, $109
-	i64.select	$push125=, $57, $4, $pop124
-	i64.select	$push126=, $60, $pop125, $51
-	i64.select	$push127=, $77, $pop119, $pop126
-	i64.select	$push128=, $70, $12, $pop127
-	i64.store	$discard=, 0($pop330), $pop128
-	i32.const	$push331=, 80
+	i64.const	$push361=, 0
+	i64.select	$push185=, $62, $132, $pop361
+	i64.or  	$push182=, $129, $130
+	i64.select	$push183=, $67, $pop182, $131
+	i64.select	$push184=, $110, $9, $pop183
+	i64.or  	$push186=, $pop185, $pop184
+	i64.const	$push360=, 0
+	i64.select	$push187=, $57, $133, $pop360
+	i64.select	$push188=, $69, $pop186, $pop187
+	i64.select	$push189=, $66, $13, $pop188
+	i64.or  	$push178=, $126, $127
+	i64.select	$push179=, $70, $pop178, $128
+	i64.select	$push180=, $93, $5, $pop179
+	i64.const	$push359=, 0
+	i64.select	$push181=, $72, $pop180, $pop359
+	i64.or  	$push190=, $pop189, $pop181
+	i64.const	$push358=, 0
+	i64.select	$push194=, $49, $138, $pop358
+	i64.or  	$push191=, $134, $135
+	i64.select	$push192=, $54, $pop191, $136
+	i64.select	$push193=, $137, $1, $pop192
+	i64.or  	$push195=, $pop194, $pop193
+	i64.const	$push357=, 0
+	i64.select	$push196=, $47, $139, $pop357
+	i64.select	$push197=, $56, $pop195, $pop196
+	i64.select	$push198=, $53, $5, $pop197
+	i64.select	$push199=, $73, $pop190, $pop198
+	i64.select	$push200=, $66, $13, $pop199
+	i64.store	$discard=, 0($pop330), $pop200
+	i32.const	$push331=, 72
 	i32.add 	$push332=, $0, $pop331
-	i64.or  	$push80=, $17, $42
-	i64.select	$push81=, $66, $pop80, $44
-	i64.select	$push82=, $70, $11, $pop81
-	i64.select	$push83=, $73, $pop82, $51
-	i64.select	$push87=, $74, $94, $51
-	i64.or  	$push84=, $90, $91
-	i64.select	$push85=, $89, $pop84, $92
-	i64.select	$push86=, $93, $7, $pop85
-	i64.or  	$push88=, $pop87, $pop86
-	i64.select	$push89=, $71, $95, $51
-	i64.select	$push90=, $76, $pop88, $pop89
-	i64.select	$push91=, $96, $3, $pop90
-	i64.or  	$push92=, $pop83, $pop91
-	i64.or  	$push93=, $98, $97
-	i64.select	$push94=, $53, $pop93, $99
-	i64.select	$push95=, $57, $3, $pop94
-	i64.select	$push96=, $60, $pop95, $51
-	i64.select	$push97=, $77, $pop92, $pop96
-	i64.select	$push98=, $70, $11, $pop97
-	i64.store	$discard=, 0($pop332), $pop98
-	i32.const	$183=, 1024
-	i32.add 	$279=, $279, $183
-	i32.const	$183=, __stack_pointer
-	i32.store	$279=, 0($183), $279
+	i64.const	$push356=, 0
+	i64.select	$push170=, $62, $124, $pop356
+	i64.const	$push355=, 0
+	i64.select	$push171=, $69, $pop170, $pop355
+	i64.or  	$push160=, $120, $121
+	i64.select	$push162=, $70, $pop160, $122
+	i64.select	$push163=, $93, $2, $pop162
+	i64.const	$push354=, 0
+	i64.select	$push165=, $86, $123, $pop354
+	i64.or  	$push166=, $pop163, $pop165
+	i64.or  	$push154=, $117, $118
+	i64.select	$push156=, $67, $pop154, $119
+	i64.select	$push157=, $110, $6, $pop156
+	i64.select	$push167=, $72, $pop166, $pop157
+	i64.select	$push168=, $93, $2, $pop167
+	i64.or  	$push172=, $pop171, $pop168
+	i64.const	$push353=, 0
+	i64.select	$push174=, $49, $125, $pop353
+	i64.const	$push352=, 0
+	i64.select	$push175=, $56, $pop174, $pop352
+	i64.select	$push176=, $73, $pop172, $pop175
+	i64.select	$push177=, $66, $10, $pop176
+	i64.store	$discard=, 0($pop332), $pop177
+	i32.const	$push333=, 64
+	i32.add 	$push334=, $0, $pop333
+	i64.const	$push351=, 0
+	i64.select	$push145=, $62, $115, $pop351
+	i64.const	$push350=, 0
+	i64.select	$push146=, $69, $pop145, $pop350
+	i64.or  	$push138=, $111, $112
+	i64.select	$push139=, $70, $pop138, $113
+	i64.select	$push140=, $93, $1, $pop139
+	i64.const	$push349=, 0
+	i64.select	$push141=, $86, $114, $pop349
+	i64.or  	$push142=, $pop140, $pop141
+	i64.or  	$push135=, $107, $108
+	i64.select	$push136=, $67, $pop135, $109
+	i64.select	$push137=, $110, $5, $pop136
+	i64.select	$push143=, $72, $pop142, $pop137
+	i64.select	$push144=, $93, $1, $pop143
+	i64.or  	$push147=, $pop146, $pop144
+	i64.const	$push348=, 0
+	i64.select	$push148=, $49, $116, $pop348
+	i64.const	$push347=, 0
+	i64.select	$push149=, $56, $pop148, $pop347
+	i64.select	$push150=, $73, $pop147, $pop149
+	i64.select	$push151=, $66, $9, $pop150
+	i64.store	$discard=, 0($pop334), $pop151
+	i32.const	$push335=, 88
+	i32.add 	$push336=, $0, $pop335
+	i64.or  	$push107=, $97, $98
+	i64.select	$push109=, $62, $pop107, $99
+	i64.select	$push110=, $66, $12, $pop109
+	i64.const	$push346=, 0
+	i64.select	$push111=, $69, $pop110, $pop346
+	i64.const	$push345=, 0
+	i64.select	$push119=, $70, $45, $pop345
+	i64.or  	$push114=, $100, $101
+	i64.select	$push116=, $86, $pop114, $102
+	i64.select	$push117=, $90, $8, $pop116
+	i64.or  	$push120=, $pop119, $pop117
+	i64.const	$push344=, 0
+	i64.select	$push122=, $67, $103, $pop344
+	i64.select	$push123=, $72, $pop120, $pop122
+	i64.select	$push124=, $93, $4, $pop123
+	i64.or  	$push125=, $pop111, $pop124
+	i64.or  	$push128=, $105, $104
+	i64.select	$push130=, $49, $pop128, $106
+	i64.select	$push131=, $53, $4, $pop130
+	i64.const	$push343=, 0
+	i64.select	$push132=, $56, $pop131, $pop343
+	i64.select	$push133=, $73, $pop125, $pop132
+	i64.select	$push134=, $66, $12, $pop133
+	i64.store	$discard=, 0($pop336), $pop134
+	i32.const	$push337=, 80
+	i32.add 	$push338=, $0, $pop337
+	i64.or  	$push86=, $17, $84
+	i64.select	$push87=, $62, $pop86, $85
+	i64.select	$push88=, $66, $11, $pop87
+	i64.const	$push342=, 0
+	i64.select	$push89=, $69, $pop88, $pop342
+	i64.const	$push341=, 0
+	i64.select	$push93=, $70, $91, $pop341
+	i64.or  	$push90=, $87, $88
+	i64.select	$push91=, $86, $pop90, $89
+	i64.select	$push92=, $90, $7, $pop91
+	i64.or  	$push94=, $pop93, $pop92
+	i64.const	$push340=, 0
+	i64.select	$push95=, $67, $92, $pop340
+	i64.select	$push96=, $72, $pop94, $pop95
+	i64.select	$push97=, $93, $3, $pop96
+	i64.or  	$push98=, $pop89, $pop97
+	i64.or  	$push99=, $95, $94
+	i64.select	$push100=, $49, $pop99, $96
+	i64.select	$push101=, $53, $3, $pop100
+	i64.const	$push339=, 0
+	i64.select	$push102=, $56, $pop101, $pop339
+	i64.select	$push103=, $73, $pop98, $pop102
+	i64.select	$push104=, $66, $11, $pop103
+	i64.store	$discard=, 0($pop338), $pop104
+	i32.const	$181=, 1024
+	i32.add 	$310=, $310, $181
+	i32.const	$181=, __stack_pointer
+	i32.store	$310=, 0($181), $310
 	return
+	.endfunc
 .Lfunc_end5:
 	.size	bigshift, .Lfunc_end5-bigshift
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/legalize.wast b/test/llvm_autogenerated/legalize.wast
index 6a6c87461..6feec29d4 100644
--- a/test/llvm_autogenerated/legalize.wast
+++ b/test/llvm_autogenerated/legalize.wast
@@ -40,19 +40,15 @@
     )
   )
   (func $sext_in_reg_i32_i64 (param $$0 i64) (result i64)
-    (local $$1 i64)
     (block $fake_return_waka123
       (block
-        (set_local $$1
-          (i64.const 32)
-        )
         (br $fake_return_waka123
           (i64.shr_s
             (i64.shl
               (get_local $$0)
-              (get_local $$1)
+              (i64.const 32)
             )
-            (get_local $$1)
+            (i64.const 32)
           )
         )
       )
@@ -99,38 +95,38 @@
     (local $$44 i64)
     (local $$45 i64)
     (local $$46 i64)
-    (local $$47 i64)
-    (local $$48 i64)
-    (local $$49 i64)
-    (local $$50 i32)
+    (local $$47 i32)
+    (local $$48 i32)
+    (local $$49 i32)
+    (local $$50 i64)
     (local $$51 i64)
-    (local $$52 i32)
+    (local $$52 i64)
     (local $$53 i32)
-    (local $$54 i64)
+    (local $$54 i32)
     (local $$55 i64)
-    (local $$56 i64)
+    (local $$56 i32)
     (local $$57 i32)
-    (local $$58 i32)
+    (local $$58 i64)
     (local $$59 i64)
-    (local $$60 i32)
+    (local $$60 i64)
     (local $$61 i32)
-    (local $$62 i64)
+    (local $$62 i32)
     (local $$63 i64)
     (local $$64 i64)
-    (local $$65 i32)
+    (local $$65 i64)
     (local $$66 i32)
-    (local $$67 i64)
+    (local $$67 i32)
     (local $$68 i64)
-    (local $$69 i64)
+    (local $$69 i32)
     (local $$70 i32)
-    (local $$71 i32)
-    (local $$72 i64)
+    (local $$71 i64)
+    (local $$72 i32)
     (local $$73 i32)
-    (local $$74 i32)
+    (local $$74 i64)
     (local $$75 i64)
-    (local $$76 i32)
-    (local $$77 i32)
-    (local $$78 i32)
+    (local $$76 i64)
+    (local $$77 i64)
+    (local $$78 i64)
     (local $$79 i64)
     (local $$80 i64)
     (local $$81 i64)
@@ -138,17 +134,17 @@
     (local $$83 i64)
     (local $$84 i64)
     (local $$85 i64)
-    (local $$86 i64)
+    (local $$86 i32)
     (local $$87 i64)
     (local $$88 i64)
-    (local $$89 i32)
-    (local $$90 i64)
+    (local $$89 i64)
+    (local $$90 i32)
     (local $$91 i64)
     (local $$92 i64)
     (local $$93 i32)
     (local $$94 i64)
     (local $$95 i64)
-    (local $$96 i32)
+    (local $$96 i64)
     (local $$97 i64)
     (local $$98 i64)
     (local $$99 i64)
@@ -162,10 +158,10 @@
     (local $$107 i64)
     (local $$108 i64)
     (local $$109 i64)
-    (local $$110 i64)
+    (local $$110 i32)
     (local $$111 i64)
     (local $$112 i64)
-    (local $$113 i32)
+    (local $$113 i64)
     (local $$114 i64)
     (local $$115 i64)
     (local $$116 i64)
@@ -189,9 +185,9 @@
     (local $$134 i64)
     (local $$135 i64)
     (local $$136 i64)
-    (local $$137 i64)
+    (local $$137 i32)
     (local $$138 i64)
-    (local $$139 i32)
+    (local $$139 i64)
     (local $$140 i64)
     (local $$141 i64)
     (local $$142 i64)
@@ -231,8 +227,8 @@
     (local $$176 i64)
     (local $$177 i64)
     (local $$178 i64)
-    (local $$179 i64)
-    (local $$180 i64)
+    (local $$179 i32)
+    (local $$180 i32)
     (local $$181 i32)
     (local $$182 i32)
     (local $$183 i32)
@@ -363,47 +359,45 @@
     (local $$308 i32)
     (local $$309 i32)
     (local $$310 i32)
-    (local $$311 i32)
-    (local $$312 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$181
+        (set_local $$179
           (i32.const 1)
         )
-        (set_local $$181
+        (set_local $$179
           (i32.load align=4
-            (get_local $$181)
+            (get_local $$179)
           )
         )
-        (set_local $$182
+        (set_local $$180
           (i32.const 1024)
         )
-        (set_local $$279
+        (set_local $$310
           (i32.sub
-            (get_local $$181)
-            (get_local $$182)
+            (get_local $$179)
+            (get_local $$180)
           )
         )
-        (set_local $$182
+        (set_local $$180
           (i32.const 1)
         )
-        (set_local $$279
+        (set_local $$310
           (i32.store align=4
-            (get_local $$182)
-            (get_local $$279)
+            (get_local $$180)
+            (get_local $$310)
           )
         )
-        (set_local $$184
+        (set_local $$182
           (i32.const 480)
         )
-        (set_local $$184
+        (set_local $$182
           (i32.add
-            (get_local $$279)
-            (get_local $$184)
+            (get_local $$310)
+            (get_local $$182)
           )
         )
         (call_import $__lshrti3
-          (get_local $$184)
+          (get_local $$182)
           (get_local $$1)
           (get_local $$2)
           (i64.sub
@@ -417,32 +411,32 @@
             (i64.const -768)
           )
         )
-        (set_local $$185
+        (set_local $$183
           (i32.const 464)
         )
-        (set_local $$185
+        (set_local $$183
           (i32.add
-            (get_local $$279)
-            (get_local $$185)
+            (get_local $$310)
+            (get_local $$183)
           )
         )
         (call_import $__ashlti3
-          (get_local $$185)
+          (get_local $$183)
           (get_local $$3)
           (get_local $$4)
           (get_local $$33)
         )
-        (set_local $$186
+        (set_local $$184
           (i32.const 496)
         )
-        (set_local $$186
+        (set_local $$184
           (i32.add
-            (get_local $$279)
-            (get_local $$186)
+            (get_local $$310)
+            (get_local $$184)
           )
         )
         (call_import $__ashlti3
-          (get_local $$186)
+          (get_local $$184)
           (get_local $$1)
           (get_local $$2)
           (i64.add
@@ -456,17 +450,17 @@
             (get_local $$17)
           )
         )
-        (set_local $$187
+        (set_local $$185
           (i32.const 352)
         )
-        (set_local $$187
+        (set_local $$185
           (i32.add
-            (get_local $$279)
-            (get_local $$187)
+            (get_local $$310)
+            (get_local $$185)
           )
         )
         (call_import $__lshrti3
-          (get_local $$187)
+          (get_local $$185)
           (get_local $$5)
           (get_local $$6)
           (get_local $$34)
@@ -477,17 +471,17 @@
             (i64.const -512)
           )
         )
-        (set_local $$188
+        (set_local $$186
           (i32.const 336)
         )
-        (set_local $$188
+        (set_local $$186
           (i32.add
-            (get_local $$279)
-            (get_local $$188)
+            (get_local $$310)
+            (get_local $$186)
           )
         )
         (call_import $__ashlti3
-          (get_local $$188)
+          (get_local $$186)
           (get_local $$7)
           (get_local $$8)
           (get_local $$35)
@@ -498,17 +492,17 @@
             (i64.const -640)
           )
         )
-        (set_local $$189
+        (set_local $$187
           (i32.const 368)
         )
-        (set_local $$189
+        (set_local $$187
           (i32.add
-            (get_local $$279)
-            (get_local $$189)
+            (get_local $$310)
+            (get_local $$187)
           )
         )
         (call_import $__ashlti3
-          (get_local $$189)
+          (get_local $$187)
           (get_local $$5)
           (get_local $$6)
           (get_local $$36)
@@ -519,1465 +513,1481 @@
             (get_local $$17)
           )
         )
-        (set_local $$190
+        (set_local $$188
           (i32.const 432)
         )
-        (set_local $$190
+        (set_local $$188
           (i32.add
-            (get_local $$279)
-            (get_local $$190)
+            (get_local $$310)
+            (get_local $$188)
           )
         )
         (call_import $__lshrti3
-          (get_local $$190)
+          (get_local $$188)
           (get_local $$3)
           (get_local $$4)
           (get_local $$37)
         )
         (set_local $$38
-          (i64.const 384)
-        )
-        (set_local $$39
           (i64.sub
-            (get_local $$38)
+            (i64.const 384)
             (get_local $$17)
           )
         )
-        (set_local $$191
+        (set_local $$189
           (i32.const 864)
         )
-        (set_local $$191
+        (set_local $$189
           (i32.add
-            (get_local $$279)
-            (get_local $$191)
+            (get_local $$310)
+            (get_local $$189)
           )
         )
         (call_import $__lshrti3
-          (get_local $$191)
+          (get_local $$189)
           (get_local $$9)
           (get_local $$10)
-          (get_local $$39)
+          (get_local $$38)
         )
-        (set_local $$40
+        (set_local $$39
           (i64.add
             (get_local $$17)
             (i64.const -256)
           )
         )
-        (set_local $$192
+        (set_local $$190
           (i32.const 848)
         )
-        (set_local $$192
+        (set_local $$190
           (i32.add
-            (get_local $$279)
-            (get_local $$192)
+            (get_local $$310)
+            (get_local $$190)
           )
         )
         (call_import $__ashlti3
-          (get_local $$192)
+          (get_local $$190)
           (get_local $$11)
           (get_local $$12)
-          (get_local $$40)
+          (get_local $$39)
         )
-        (set_local $$41
+        (set_local $$40
           (i64.add
             (get_local $$17)
             (i64.const -384)
           )
         )
-        (set_local $$193
+        (set_local $$191
           (i32.const 880)
         )
-        (set_local $$193
+        (set_local $$191
           (i32.add
-            (get_local $$279)
-            (get_local $$193)
+            (get_local $$310)
+            (get_local $$191)
           )
         )
         (call_import $__ashlti3
-          (get_local $$193)
+          (get_local $$191)
           (get_local $$9)
           (get_local $$10)
-          (get_local $$41)
+          (get_local $$40)
         )
-        (set_local $$194
+        (set_local $$192
           (i32.const 1008)
         )
-        (set_local $$194
+        (set_local $$192
           (i32.add
-            (get_local $$279)
-            (get_local $$194)
+            (get_local $$310)
+            (get_local $$192)
           )
         )
         (call_import $__ashlti3
-          (get_local $$194)
+          (get_local $$192)
           (get_local $$15)
           (get_local $$16)
           (get_local $$17)
         )
-        (set_local $$42
-          (i64.const 128)
-        )
-        (set_local $$51
+        (set_local $$41
           (i64.sub
-            (get_local $$42)
+            (i64.const 128)
             (get_local $$17)
           )
         )
-        (set_local $$195
+        (set_local $$193
           (i32.const 960)
         )
-        (set_local $$195
+        (set_local $$193
           (i32.add
-            (get_local $$279)
-            (get_local $$195)
+            (get_local $$310)
+            (get_local $$193)
           )
         )
         (call_import $__lshrti3
-          (get_local $$195)
+          (get_local $$193)
           (get_local $$13)
           (get_local $$14)
-          (get_local $$51)
+          (get_local $$41)
         )
-        (set_local $$43
+        (set_local $$42
           (i64.add
             (get_local $$17)
             (i64.const -128)
           )
         )
-        (set_local $$196
+        (set_local $$194
           (i32.const 976)
         )
-        (set_local $$196
+        (set_local $$194
           (i32.add
-            (get_local $$279)
-            (get_local $$196)
+            (get_local $$310)
+            (get_local $$194)
           )
         )
         (call_import $__ashlti3
-          (get_local $$196)
+          (get_local $$194)
           (get_local $$13)
           (get_local $$14)
-          (get_local $$43)
+          (get_local $$42)
         )
-        (set_local $$44
-          (i64.const 256)
-        )
-        (set_local $$45
+        (set_local $$43
           (i64.sub
-            (get_local $$44)
+            (i64.const 256)
             (get_local $$17)
           )
         )
-        (set_local $$197
+        (set_local $$195
           (i32.const 816)
         )
-        (set_local $$197
+        (set_local $$195
           (i32.add
-            (get_local $$279)
-            (get_local $$197)
+            (get_local $$310)
+            (get_local $$195)
           )
         )
         (call_import $__lshrti3
-          (get_local $$197)
+          (get_local $$195)
           (get_local $$11)
           (get_local $$12)
-          (get_local $$45)
-        )
-        (set_local $$46
-          (i64.const 512)
+          (get_local $$43)
         )
-        (set_local $$47
+        (set_local $$44
           (i64.sub
-            (get_local $$46)
+            (i64.const 512)
             (get_local $$17)
           )
         )
-        (set_local $$198
+        (set_local $$196
           (i32.const 240)
         )
-        (set_local $$198
+        (set_local $$196
           (i32.add
-            (get_local $$279)
-            (get_local $$198)
+            (get_local $$310)
+            (get_local $$196)
           )
         )
         (call_import $__lshrti3
-          (get_local $$198)
+          (get_local $$196)
           (get_local $$7)
           (get_local $$8)
-          (get_local $$47)
+          (get_local $$44)
         )
-        (set_local $$199
+        (set_local $$197
           (i32.const 912)
         )
-        (set_local $$199
+        (set_local $$197
           (i32.add
-            (get_local $$279)
-            (get_local $$199)
+            (get_local $$310)
+            (get_local $$197)
           )
         )
         (call_import $__ashlti3
-          (get_local $$199)
+          (get_local $$197)
           (get_local $$11)
           (get_local $$12)
           (get_local $$17)
         )
-        (set_local $$200
+        (set_local $$198
           (i32.const 928)
         )
-        (set_local $$200
+        (set_local $$198
           (i32.add
-            (get_local $$279)
-            (get_local $$200)
+            (get_local $$310)
+            (get_local $$198)
           )
         )
         (call_import $__lshrti3
-          (get_local $$200)
+          (get_local $$198)
           (get_local $$9)
           (get_local $$10)
-          (get_local $$51)
+          (get_local $$41)
         )
-        (set_local $$201
+        (set_local $$199
           (i32.const 944)
         )
-        (set_local $$201
+        (set_local $$199
           (i32.add
-            (get_local $$279)
-            (get_local $$201)
+            (get_local $$310)
+            (get_local $$199)
           )
         )
         (call_import $__ashlti3
-          (get_local $$201)
+          (get_local $$199)
           (get_local $$9)
           (get_local $$10)
-          (get_local $$43)
+          (get_local $$42)
         )
-        (set_local $$48
+        (set_local $$45
           (i64.sub
+            (i64.const 256)
             (get_local $$44)
-            (get_local $$47)
           )
         )
-        (set_local $$202
+        (set_local $$200
           (i32.const 80)
         )
-        (set_local $$202
+        (set_local $$200
           (i32.add
-            (get_local $$279)
-            (get_local $$202)
+            (get_local $$310)
+            (get_local $$200)
           )
         )
         (call_import $__ashlti3
-          (get_local $$202)
+          (get_local $$200)
           (get_local $$7)
           (get_local $$8)
-          (get_local $$48)
+          (get_local $$45)
         )
-        (set_local $$203
+        (set_local $$201
           (i32.const 96)
         )
-        (set_local $$203
+        (set_local $$201
           (i32.add
-            (get_local $$279)
-            (get_local $$203)
+            (get_local $$310)
+            (get_local $$201)
           )
         )
         (call_import $__lshrti3
-          (get_local $$203)
+          (get_local $$201)
           (get_local $$5)
           (get_local $$6)
           (i64.sub
-            (get_local $$42)
-            (get_local $$48)
+            (i64.const 128)
+            (get_local $$45)
           )
         )
-        (set_local $$49
+        (set_local $$46
           (i64.sub
-            (get_local $$42)
-            (get_local $$47)
+            (i64.const 128)
+            (get_local $$44)
           )
         )
-        (set_local $$204
+        (set_local $$202
           (i32.const 112)
         )
-        (set_local $$204
+        (set_local $$202
           (i32.add
-            (get_local $$279)
-            (get_local $$204)
+            (get_local $$310)
+            (get_local $$202)
           )
         )
         (call_import $__ashlti3
-          (get_local $$204)
+          (get_local $$202)
           (get_local $$5)
           (get_local $$6)
-          (get_local $$49)
+          (get_local $$46)
         )
-        (set_local $$205
+        (set_local $$203
           (i32.const 48)
         )
-        (set_local $$205
+        (set_local $$203
           (i32.add
-            (get_local $$279)
-            (get_local $$205)
+            (get_local $$310)
+            (get_local $$203)
           )
         )
         (call_import $__lshrti3
-          (get_local $$205)
+          (get_local $$203)
           (get_local $$3)
           (get_local $$4)
-          (get_local $$47)
+          (get_local $$44)
         )
-        (set_local $$206
+        (set_local $$204
           (i32.const 176)
         )
-        (set_local $$206
+        (set_local $$204
           (i32.add
-            (get_local $$279)
-            (get_local $$206)
+            (get_local $$310)
+            (get_local $$204)
           )
         )
         (call_import $__lshrti3
-          (get_local $$206)
+          (get_local $$204)
           (get_local $$7)
           (get_local $$8)
-          (get_local $$45)
+          (get_local $$43)
         )
-        (set_local $$207
+        (set_local $$205
           (i32.const 288)
         )
-        (set_local $$207
+        (set_local $$205
           (i32.add
-            (get_local $$279)
-            (get_local $$207)
+            (get_local $$310)
+            (get_local $$205)
           )
         )
         (call_import $__lshrti3
-          (get_local $$207)
+          (get_local $$205)
           (get_local $$1)
           (get_local $$2)
           (get_local $$34)
         )
-        (set_local $$208
+        (set_local $$206
           (i32.const 272)
         )
-        (set_local $$208
+        (set_local $$206
           (i32.add
-            (get_local $$279)
-            (get_local $$208)
+            (get_local $$310)
+            (get_local $$206)
           )
         )
         (call_import $__ashlti3
-          (get_local $$208)
+          (get_local $$206)
           (get_local $$3)
           (get_local $$4)
           (get_local $$35)
         )
-        (set_local $$209
+        (set_local $$207
           (i32.const 304)
         )
-        (set_local $$209
+        (set_local $$207
           (i32.add
-            (get_local $$279)
-            (get_local $$209)
+            (get_local $$310)
+            (get_local $$207)
           )
         )
         (call_import $__ashlti3
-          (get_local $$209)
+          (get_local $$207)
           (get_local $$1)
           (get_local $$2)
           (get_local $$36)
         )
-        (set_local $$210
+        (set_local $$208
           (i32.const 128)
         )
-        (set_local $$210
+        (set_local $$208
           (i32.add
-            (get_local $$279)
-            (get_local $$210)
+            (get_local $$310)
+            (get_local $$208)
           )
         )
         (call_import $__lshrti3
-          (get_local $$210)
+          (get_local $$208)
           (get_local $$5)
           (get_local $$6)
-          (get_local $$45)
+          (get_local $$43)
         )
-        (set_local $$211
+        (set_local $$209
           (i32.const 144)
         )
-        (set_local $$211
+        (set_local $$209
           (i32.add
-            (get_local $$279)
-            (get_local $$211)
+            (get_local $$310)
+            (get_local $$209)
           )
         )
         (call_import $__ashlti3
-          (get_local $$211)
+          (get_local $$209)
           (get_local $$7)
           (get_local $$8)
           (i64.sub
-            (get_local $$38)
-            (get_local $$47)
+            (i64.const 384)
+            (get_local $$44)
           )
         )
-        (set_local $$212
+        (set_local $$210
           (i32.const 160)
         )
-        (set_local $$212
+        (set_local $$210
           (i32.add
-            (get_local $$279)
-            (get_local $$212)
+            (get_local $$310)
+            (get_local $$210)
           )
         )
         (call_import $__lshrti3
-          (get_local $$212)
+          (get_local $$210)
           (get_local $$7)
           (get_local $$8)
-          (get_local $$51)
+          (get_local $$41)
         )
-        (set_local $$213
+        (set_local $$211
           (i32.const 0)
         )
-        (set_local $$213
+        (set_local $$211
           (i32.add
-            (get_local $$279)
-            (get_local $$213)
+            (get_local $$310)
+            (get_local $$211)
           )
         )
         (call_import $__lshrti3
-          (get_local $$213)
+          (get_local $$211)
           (get_local $$1)
           (get_local $$2)
-          (get_local $$47)
+          (get_local $$44)
         )
-        (set_local $$214
+        (set_local $$212
           (i32.const 16)
         )
-        (set_local $$214
+        (set_local $$212
           (i32.add
-            (get_local $$279)
-            (get_local $$214)
+            (get_local $$310)
+            (get_local $$212)
           )
         )
         (call_import $__ashlti3
-          (get_local $$214)
+          (get_local $$212)
           (get_local $$3)
           (get_local $$4)
-          (get_local $$49)
+          (get_local $$46)
         )
-        (set_local $$215
+        (set_local $$213
           (i32.const 32)
         )
-        (set_local $$215
+        (set_local $$213
           (i32.add
-            (get_local $$279)
-            (get_local $$215)
+            (get_local $$310)
+            (get_local $$213)
           )
         )
         (call_import $__lshrti3
-          (get_local $$215)
+          (get_local $$213)
           (get_local $$3)
           (get_local $$4)
-          (get_local $$39)
+          (get_local $$38)
         )
-        (set_local $$216
+        (set_local $$214
           (i32.const 64)
         )
-        (set_local $$216
+        (set_local $$214
           (i32.add
-            (get_local $$279)
-            (get_local $$216)
+            (get_local $$310)
+            (get_local $$214)
           )
         )
         (call_import $__ashlti3
-          (get_local $$216)
+          (get_local $$214)
           (get_local $$5)
           (get_local $$6)
-          (get_local $$48)
+          (get_local $$45)
         )
-        (set_local $$217
+        (set_local $$215
           (i32.const 896)
         )
-        (set_local $$217
+        (set_local $$215
           (i32.add
-            (get_local $$279)
-            (get_local $$217)
+            (get_local $$310)
+            (get_local $$215)
           )
         )
         (call_import $__ashlti3
-          (get_local $$217)
+          (get_local $$215)
           (get_local $$9)
           (get_local $$10)
           (get_local $$17)
         )
-        (set_local $$218
+        (set_local $$216
           (i32.const 256)
         )
-        (set_local $$218
+        (set_local $$216
           (i32.add
-            (get_local $$279)
-            (get_local $$218)
+            (get_local $$310)
+            (get_local $$216)
           )
         )
         (call_import $__ashlti3
-          (get_local $$218)
+          (get_local $$216)
           (get_local $$1)
           (get_local $$2)
           (get_local $$35)
         )
-        (set_local $$219
+        (set_local $$217
           (i32.const 192)
         )
-        (set_local $$219
+        (set_local $$217
           (i32.add
-            (get_local $$279)
-            (get_local $$219)
+            (get_local $$310)
+            (get_local $$217)
           )
         )
         (call_import $__lshrti3
-          (get_local $$219)
+          (get_local $$217)
           (get_local $$5)
           (get_local $$6)
-          (get_local $$47)
+          (get_local $$44)
         )
-        (set_local $$220
+        (set_local $$218
           (i32.const 208)
         )
-        (set_local $$220
+        (set_local $$218
           (i32.add
-            (get_local $$279)
-            (get_local $$220)
+            (get_local $$310)
+            (get_local $$218)
           )
         )
         (call_import $__ashlti3
-          (get_local $$220)
+          (get_local $$218)
           (get_local $$7)
           (get_local $$8)
-          (get_local $$49)
+          (get_local $$46)
         )
-        (set_local $$221
+        (set_local $$219
           (i32.const 224)
         )
-        (set_local $$221
+        (set_local $$219
           (i32.add
-            (get_local $$279)
-            (get_local $$221)
+            (get_local $$310)
+            (get_local $$219)
           )
         )
         (call_import $__lshrti3
-          (get_local $$221)
+          (get_local $$219)
           (get_local $$7)
           (get_local $$8)
-          (get_local $$39)
+          (get_local $$38)
         )
-        (set_local $$222
+        (set_local $$220
           (i32.const 768)
         )
-        (set_local $$222
+        (set_local $$220
           (i32.add
-            (get_local $$279)
-            (get_local $$222)
+            (get_local $$310)
+            (get_local $$220)
           )
         )
         (call_import $__lshrti3
-          (get_local $$222)
+          (get_local $$220)
           (get_local $$9)
           (get_local $$10)
-          (get_local $$45)
+          (get_local $$43)
         )
-        (set_local $$49
+        (set_local $$46
           (i64.sub
-            (get_local $$42)
-            (get_local $$45)
+            (i64.const 128)
+            (get_local $$43)
           )
         )
-        (set_local $$223
+        (set_local $$221
           (i32.const 784)
         )
-        (set_local $$223
+        (set_local $$221
           (i32.add
-            (get_local $$279)
-            (get_local $$223)
+            (get_local $$310)
+            (get_local $$221)
           )
         )
         (call_import $__ashlti3
-          (get_local $$223)
+          (get_local $$221)
           (get_local $$11)
           (get_local $$12)
-          (get_local $$49)
+          (get_local $$46)
         )
-        (set_local $$224
+        (set_local $$222
           (i32.const 800)
         )
-        (set_local $$224
+        (set_local $$222
           (i32.add
-            (get_local $$279)
-            (get_local $$224)
+            (get_local $$310)
+            (get_local $$222)
           )
         )
         (call_import $__lshrti3
-          (get_local $$224)
+          (get_local $$222)
           (get_local $$11)
           (get_local $$12)
-          (get_local $$51)
+          (get_local $$41)
         )
-        (set_local $$225
+        (set_local $$223
           (i32.const 992)
         )
-        (set_local $$225
+        (set_local $$223
           (i32.add
-            (get_local $$279)
-            (get_local $$225)
+            (get_local $$310)
+            (get_local $$223)
           )
         )
         (call_import $__ashlti3
-          (get_local $$225)
+          (get_local $$223)
           (get_local $$13)
           (get_local $$14)
           (get_local $$17)
         )
-        (set_local $$226
+        (set_local $$224
           (i32.const 832)
         )
-        (set_local $$226
+        (set_local $$224
           (i32.add
-            (get_local $$279)
-            (get_local $$226)
+            (get_local $$310)
+            (get_local $$224)
           )
         )
         (call_import $__ashlti3
-          (get_local $$226)
+          (get_local $$224)
           (get_local $$9)
           (get_local $$10)
-          (get_local $$40)
+          (get_local $$39)
         )
-        (set_local $$227
+        (set_local $$225
           (i32.const 384)
         )
-        (set_local $$227
+        (set_local $$225
           (i32.add
-            (get_local $$279)
-            (get_local $$227)
+            (get_local $$310)
+            (get_local $$225)
           )
         )
         (call_import $__lshrti3
-          (get_local $$227)
+          (get_local $$225)
           (get_local $$1)
           (get_local $$2)
           (get_local $$37)
         )
-        (set_local $$228
+        (set_local $$226
           (i32.const 400)
         )
-        (set_local $$228
+        (set_local $$226
           (i32.add
-            (get_local $$279)
-            (get_local $$228)
+            (get_local $$310)
+            (get_local $$226)
           )
         )
         (call_import $__ashlti3
-          (get_local $$228)
+          (get_local $$226)
           (get_local $$3)
           (get_local $$4)
           (i64.sub
-            (get_local $$42)
+            (i64.const 128)
             (get_local $$37)
           )
         )
-        (set_local $$229
+        (set_local $$227
           (i32.const 416)
         )
-        (set_local $$229
+        (set_local $$227
           (i32.add
-            (get_local $$279)
-            (get_local $$229)
+            (get_local $$310)
+            (get_local $$227)
           )
         )
         (call_import $__lshrti3
-          (get_local $$229)
+          (get_local $$227)
           (get_local $$3)
           (get_local $$4)
           (get_local $$34)
         )
-        (set_local $$230
+        (set_local $$228
           (i32.const 320)
         )
-        (set_local $$230
+        (set_local $$228
           (i32.add
-            (get_local $$279)
-            (get_local $$230)
+            (get_local $$310)
+            (get_local $$228)
           )
         )
         (call_import $__ashlti3
-          (get_local $$230)
+          (get_local $$228)
           (get_local $$5)
           (get_local $$6)
           (get_local $$35)
         )
-        (set_local $$231
+        (set_local $$229
           (i32.const 448)
         )
-        (set_local $$231
+        (set_local $$229
           (i32.add
-            (get_local $$279)
-            (get_local $$231)
+            (get_local $$310)
+            (get_local $$229)
           )
         )
         (call_import $__ashlti3
-          (get_local $$231)
+          (get_local $$229)
           (get_local $$1)
           (get_local $$2)
           (get_local $$33)
         )
-        (set_local $$232
+        (set_local $$230
           (i32.const 736)
         )
+        (set_local $$230
+          (i32.add
+            (get_local $$310)
+            (get_local $$230)
+          )
+        )
+        (call_import $__lshrti3
+          (get_local $$230)
+          (get_local $$1)
+          (get_local $$2)
+          (get_local $$38)
+        )
+        (set_local $$231
+          (i32.const 720)
+        )
+        (set_local $$231
+          (i32.add
+            (get_local $$310)
+            (get_local $$231)
+          )
+        )
+        (call_import $__ashlti3
+          (get_local $$231)
+          (get_local $$3)
+          (get_local $$4)
+          (get_local $$39)
+        )
+        (set_local $$232
+          (i32.const 752)
+        )
         (set_local $$232
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$232)
           )
         )
-        (call_import $__lshrti3
+        (call_import $__ashlti3
           (get_local $$232)
           (get_local $$1)
           (get_local $$2)
-          (get_local $$39)
+          (get_local $$40)
         )
         (set_local $$233
-          (i32.const 720)
+          (i32.const 592)
         )
         (set_local $$233
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$233)
           )
         )
         (call_import $__ashlti3
           (get_local $$233)
-          (get_local $$3)
-          (get_local $$4)
-          (get_local $$40)
+          (get_local $$7)
+          (get_local $$8)
+          (get_local $$17)
         )
         (set_local $$234
-          (i32.const 752)
+          (i32.const 608)
         )
         (set_local $$234
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$234)
           )
         )
-        (call_import $__ashlti3
+        (call_import $__lshrti3
           (get_local $$234)
-          (get_local $$1)
-          (get_local $$2)
+          (get_local $$5)
+          (get_local $$6)
           (get_local $$41)
         )
         (set_local $$235
-          (i32.const 592)
+          (i32.const 624)
         )
         (set_local $$235
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$235)
           )
         )
         (call_import $__ashlti3
           (get_local $$235)
-          (get_local $$7)
-          (get_local $$8)
-          (get_local $$17)
+          (get_local $$5)
+          (get_local $$6)
+          (get_local $$42)
         )
         (set_local $$236
-          (i32.const 608)
+          (i32.const 688)
         )
         (set_local $$236
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$236)
           )
         )
         (call_import $__lshrti3
           (get_local $$236)
-          (get_local $$5)
-          (get_local $$6)
-          (get_local $$51)
+          (get_local $$3)
+          (get_local $$4)
+          (get_local $$43)
         )
         (set_local $$237
-          (i32.const 624)
+          (i32.const 640)
         )
         (set_local $$237
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$237)
           )
         )
-        (call_import $__ashlti3
+        (call_import $__lshrti3
           (get_local $$237)
-          (get_local $$5)
-          (get_local $$6)
+          (get_local $$1)
+          (get_local $$2)
           (get_local $$43)
         )
         (set_local $$238
-          (i32.const 688)
+          (i32.const 656)
         )
         (set_local $$238
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$238)
           )
         )
-        (call_import $__lshrti3
+        (call_import $__ashlti3
           (get_local $$238)
           (get_local $$3)
           (get_local $$4)
-          (get_local $$45)
+          (get_local $$46)
         )
         (set_local $$239
-          (i32.const 640)
+          (i32.const 672)
         )
         (set_local $$239
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$239)
           )
         )
         (call_import $__lshrti3
           (get_local $$239)
-          (get_local $$1)
-          (get_local $$2)
-          (get_local $$45)
+          (get_local $$3)
+          (get_local $$4)
+          (get_local $$41)
         )
         (set_local $$240
-          (i32.const 656)
+          (i32.const 576)
         )
         (set_local $$240
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$240)
           )
         )
         (call_import $__ashlti3
           (get_local $$240)
-          (get_local $$3)
-          (get_local $$4)
-          (get_local $$49)
+          (get_local $$5)
+          (get_local $$6)
+          (get_local $$17)
         )
         (set_local $$241
-          (i32.const 672)
+          (i32.const 704)
         )
         (set_local $$241
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$241)
           )
         )
-        (call_import $__lshrti3
+        (call_import $__ashlti3
           (get_local $$241)
-          (get_local $$3)
-          (get_local $$4)
-          (get_local $$51)
+          (get_local $$1)
+          (get_local $$2)
+          (get_local $$39)
         )
         (set_local $$242
-          (i32.const 576)
+          (i32.const 528)
         )
         (set_local $$242
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$242)
           )
         )
         (call_import $__ashlti3
           (get_local $$242)
-          (get_local $$5)
-          (get_local $$6)
+          (get_local $$3)
+          (get_local $$4)
           (get_local $$17)
         )
         (set_local $$243
-          (i32.const 704)
+          (i32.const 544)
         )
         (set_local $$243
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$243)
           )
         )
-        (call_import $__ashlti3
+        (call_import $__lshrti3
           (get_local $$243)
           (get_local $$1)
           (get_local $$2)
-          (get_local $$40)
+          (get_local $$41)
         )
         (set_local $$244
-          (i32.const 528)
+          (i32.const 560)
         )
         (set_local $$244
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$244)
           )
         )
         (call_import $__ashlti3
           (get_local $$244)
-          (get_local $$3)
-          (get_local $$4)
-          (get_local $$17)
+          (get_local $$1)
+          (get_local $$2)
+          (get_local $$42)
         )
         (set_local $$245
-          (i32.const 544)
+          (i32.const 512)
         )
         (set_local $$245
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$245)
           )
         )
-        (call_import $__lshrti3
+        (call_import $__ashlti3
           (get_local $$245)
           (get_local $$1)
           (get_local $$2)
-          (get_local $$51)
+          (get_local $$17)
         )
         (set_local $$246
-          (i32.const 560)
+          (i32.const 480)
         )
         (set_local $$246
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$246)
           )
         )
-        (call_import $__ashlti3
-          (get_local $$246)
-          (get_local $$1)
-          (get_local $$2)
-          (get_local $$43)
+        (set_local $$41
+          (i64.load align=8
+            (i32.add
+              (get_local $$246)
+              (i32.const 8)
+            )
+          )
         )
         (set_local $$247
-          (i32.const 512)
+          (i32.const 464)
         )
         (set_local $$247
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$247)
           )
         )
-        (call_import $__ashlti3
-          (get_local $$247)
-          (get_local $$1)
-          (get_local $$2)
-          (get_local $$17)
+        (set_local $$38
+          (i64.load align=8
+            (i32.add
+              (get_local $$247)
+              (i32.const 8)
+            )
+          )
         )
-        (set_local $$78
-          (i32.const 8)
+        (set_local $$42
+          (i64.load offset=480 align=8
+            (get_local $$310)
+          )
+        )
+        (set_local $$34
+          (i64.load offset=464 align=8
+            (get_local $$310)
+          )
+        )
+        (set_local $$46
+          (i64.load offset=496 align=8
+            (get_local $$310)
+          )
         )
         (set_local $$248
-          (i32.const 480)
+          (i32.const 496)
         )
         (set_local $$248
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$248)
           )
         )
-        (set_local $$39
+        (set_local $$36
           (i64.load align=8
             (i32.add
               (get_local $$248)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
+        (set_local $$47
+          (i64.lt_u
+            (get_local $$33)
+            (i64.const 128)
+          )
+        )
+        (set_local $$48
+          (i64.eq
+            (get_local $$33)
+            (i64.const 0)
+          )
+        )
         (set_local $$249
-          (i32.const 464)
+          (i32.const 352)
         )
         (set_local $$249
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$249)
           )
         )
-        (set_local $$43
+        (set_local $$33
           (i64.load align=8
             (i32.add
               (get_local $$249)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$34
-          (i64.load offset=480 align=8
-            (get_local $$279)
-          )
-        )
-        (set_local $$49
-          (i64.load offset=464 align=8
-            (get_local $$279)
-          )
-        )
-        (set_local $$36
-          (i64.load offset=496 align=8
-            (get_local $$279)
-          )
-        )
         (set_local $$250
-          (i32.const 496)
+          (i32.const 336)
         )
         (set_local $$250
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$250)
           )
         )
-        (set_local $$38
+        (set_local $$40
           (i64.load align=8
             (i32.add
               (get_local $$250)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$50
-          (i64.lt_u
-            (get_local $$33)
-            (get_local $$42)
-          )
-        )
-        (set_local $$51
-          (i64.const 0)
-        )
-        (set_local $$52
-          (i64.eq
-            (get_local $$33)
-            (get_local $$51)
-          )
-        )
         (set_local $$251
-          (i32.const 352)
+          (i32.const 368)
         )
         (set_local $$251
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$251)
           )
         )
-        (set_local $$33
+        (set_local $$74
           (i64.load align=8
             (i32.add
               (get_local $$251)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$252
-          (i32.const 336)
+          (i32.const 432)
         )
         (set_local $$252
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$252)
           )
         )
-        (set_local $$41
+        (set_local $$75
           (i64.load align=8
             (i32.add
               (get_local $$252)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$253
-          (i32.const 368)
+          (i32.const 864)
         )
         (set_local $$253
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$253)
           )
         )
-        (set_local $$79
+        (set_local $$76
           (i64.load align=8
             (i32.add
               (get_local $$253)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$254
-          (i32.const 432)
+          (i32.const 848)
         )
         (set_local $$254
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$254)
           )
         )
-        (set_local $$80
+        (set_local $$77
           (i64.load align=8
             (i32.add
               (get_local $$254)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$255
-          (i32.const 864)
+          (i32.const 880)
         )
         (set_local $$255
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$255)
           )
         )
-        (set_local $$81
+        (set_local $$78
           (i64.load align=8
             (i32.add
               (get_local $$255)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$256
-          (i32.const 848)
+          (i32.const 1008)
         )
         (set_local $$256
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$256)
           )
         )
-        (set_local $$82
+        (set_local $$79
           (i64.load align=8
             (i32.add
               (get_local $$256)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$257
-          (i32.const 880)
+          (i32.const 960)
         )
         (set_local $$257
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$257)
           )
         )
-        (set_local $$83
+        (set_local $$80
           (i64.load align=8
             (i32.add
               (get_local $$257)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$258
-          (i32.const 1008)
+          (i32.const 976)
         )
         (set_local $$258
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$258)
           )
         )
-        (set_local $$84
+        (set_local $$81
           (i64.load align=8
             (i32.add
               (get_local $$258)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$259
-          (i32.const 960)
+          (i32.const 816)
         )
         (set_local $$259
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$259)
           )
         )
-        (set_local $$85
+        (set_local $$82
           (i64.load align=8
             (i32.add
               (get_local $$259)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$260
-          (i32.const 976)
+          (i32.const 240)
         )
         (set_local $$260
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$260)
           )
         )
-        (set_local $$86
+        (set_local $$83
           (i64.load align=8
             (i32.add
               (get_local $$260)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$261
-          (i32.const 816)
+          (i32.const 912)
         )
         (set_local $$261
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$261)
           )
         )
-        (set_local $$87
+        (set_local $$97
           (i64.load align=8
             (i32.add
               (get_local $$261)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$262
-          (i32.const 240)
+          (i32.const 928)
         )
         (set_local $$262
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$262)
           )
         )
-        (set_local $$88
+        (set_local $$98
           (i64.load align=8
             (i32.add
               (get_local $$262)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$263
-          (i32.const 912)
+          (i32.const 944)
         )
         (set_local $$263
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$263)
           )
         )
-        (set_local $$100
+        (set_local $$99
           (i64.load align=8
             (i32.add
               (get_local $$263)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$264
-          (i32.const 928)
+          (i32.const 80)
         )
         (set_local $$264
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$264)
           )
         )
-        (set_local $$101
+        (set_local $$100
           (i64.load align=8
             (i32.add
               (get_local $$264)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$265
-          (i32.const 944)
+          (i32.const 96)
         )
         (set_local $$265
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$265)
           )
         )
-        (set_local $$102
+        (set_local $$101
           (i64.load align=8
             (i32.add
               (get_local $$265)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
+        (set_local $$87
+          (i64.load offset=80 align=8
+            (get_local $$310)
+          )
+        )
+        (set_local $$88
+          (i64.load offset=96 align=8
+            (get_local $$310)
+          )
+        )
+        (set_local $$89
+          (i64.load offset=112 align=8
+            (get_local $$310)
+          )
+        )
         (set_local $$266
-          (i32.const 80)
+          (i32.const 112)
         )
         (set_local $$266
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$266)
           )
         )
-        (set_local $$103
+        (set_local $$102
           (i64.load align=8
             (i32.add
               (get_local $$266)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
+        (set_local $$86
+          (i64.lt_u
+            (get_local $$45)
+            (i64.const 128)
+          )
+        )
+        (set_local $$90
+          (i64.eq
+            (get_local $$45)
+            (i64.const 0)
+          )
+        )
         (set_local $$267
-          (i32.const 96)
+          (i32.const 48)
         )
         (set_local $$267
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$267)
           )
         )
-        (set_local $$104
+        (set_local $$45
           (i64.load align=8
             (i32.add
               (get_local $$267)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$90
-          (i64.load offset=80 align=8
-            (get_local $$279)
-          )
-        )
-        (set_local $$91
-          (i64.load offset=96 align=8
-            (get_local $$279)
-          )
-        )
-        (set_local $$92
-          (i64.load offset=112 align=8
-            (get_local $$279)
-          )
-        )
         (set_local $$268
-          (i32.const 112)
+          (i32.const 176)
         )
         (set_local $$268
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$268)
           )
         )
-        (set_local $$105
+        (set_local $$103
           (i64.load align=8
             (i32.add
               (get_local $$268)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$89
-          (i64.lt_u
-            (get_local $$48)
-            (get_local $$42)
-          )
-        )
-        (set_local $$93
-          (i64.eq
-            (get_local $$48)
-            (get_local $$51)
-          )
-        )
         (set_local $$269
-          (i32.const 48)
+          (i32.const 288)
         )
         (set_local $$269
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$269)
           )
         )
-        (set_local $$48
+        (set_local $$104
           (i64.load align=8
             (i32.add
               (get_local $$269)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$270
-          (i32.const 176)
+          (i32.const 272)
         )
         (set_local $$270
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$270)
           )
         )
-        (set_local $$106
+        (set_local $$105
           (i64.load align=8
             (i32.add
               (get_local $$270)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$271
-          (i32.const 288)
+          (i32.const 304)
         )
         (set_local $$271
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$271)
           )
         )
-        (set_local $$107
+        (set_local $$106
           (i64.load align=8
             (i32.add
               (get_local $$271)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$272
-          (i32.const 272)
+          (i32.const 128)
         )
         (set_local $$272
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$272)
           )
         )
-        (set_local $$108
+        (set_local $$117
           (i64.load align=8
             (i32.add
               (get_local $$272)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$273
-          (i32.const 304)
+          (i32.const 144)
         )
         (set_local $$273
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$273)
           )
         )
-        (set_local $$109
+        (set_local $$118
           (i64.load align=8
             (i32.add
               (get_local $$273)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$274
-          (i32.const 128)
+          (i32.const 160)
         )
         (set_local $$274
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$274)
           )
         )
@@ -1985,16 +1995,16 @@
           (i64.load align=8
             (i32.add
               (get_local $$274)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$275
-          (i32.const 144)
+          (i32.const 0)
         )
         (set_local $$275
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$275)
           )
         )
@@ -2002,16 +2012,16 @@
           (i64.load align=8
             (i32.add
               (get_local $$275)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$276
-          (i32.const 160)
+          (i32.const 16)
         )
         (set_local $$276
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$276)
           )
         )
@@ -2019,1021 +2029,987 @@
           (i64.load align=8
             (i32.add
               (get_local $$276)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
+        (set_local $$114
+          (i64.load offset=64 align=8
+            (get_local $$310)
+          )
+        )
         (set_local $$277
-          (i32.const 0)
+          (i32.const 64)
         )
         (set_local $$277
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$277)
           )
         )
-        (set_local $$122
+        (set_local $$123
           (i64.load align=8
             (i32.add
               (get_local $$277)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$278
-          (i32.const 16)
+          (i32.const 32)
         )
         (set_local $$278
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$278)
           )
         )
-        (set_local $$123
+        (set_local $$122
           (i64.load align=8
             (i32.add
               (get_local $$278)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$117
-          (i64.load offset=64 align=8
-            (get_local $$279)
-          )
-        )
         (set_local $$279
-          (i32.const 64)
+          (i32.const 896)
         )
         (set_local $$279
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$279)
           )
         )
-        (set_local $$125
+        (set_local $$124
           (i64.load align=8
             (i32.add
               (get_local $$279)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$280
-          (i32.const 32)
+          (i32.const 256)
         )
         (set_local $$280
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$280)
           )
         )
-        (set_local $$124
+        (set_local $$125
           (i64.load align=8
             (i32.add
               (get_local $$280)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$281
-          (i32.const 896)
+          (i32.const 192)
         )
         (set_local $$281
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$281)
           )
         )
-        (set_local $$126
+        (set_local $$140
           (i64.load align=8
             (i32.add
               (get_local $$281)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
         (set_local $$282
-          (i32.const 256)
+          (i32.const 208)
         )
         (set_local $$282
           (i32.add
-            (get_local $$279)
+            (get_local $$310)
             (get_local $$282)
           )
         )
-        (set_local $$127
+        (set_local $$141
           (i64.load align=8
             (i32.add
               (get_local $$282)
-              (get_local $$78)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$283
-          (i32.const 192)
-        )
-        (set_local $$283
-          (i32.add
-            (get_local $$279)
-            (get_local $$283)
-          )
-        )
-        (set_local $$142
-          (i64.load align=8
-            (i32.add
-              (get_local $$283)
-              (get_local $$78)
-            )
-          )
-        )
-        (set_local $$284
-          (i32.const 208)
-        )
-        (set_local $$284
-          (i32.add
-            (get_local $$279)
-            (get_local $$284)
-          )
-        )
-        (set_local $$143
-          (i64.load align=8
-            (i32.add
-              (get_local $$284)
-              (get_local $$78)
-            )
-          )
-        )
-        (set_local $$114
+        (set_local $$111
           (i64.load align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$115
+        (set_local $$112
           (i64.load offset=16 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$116
+        (set_local $$113
           (i64.load offset=32 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$285
+        (set_local $$283
           (i32.const 224)
         )
-        (set_local $$285
+        (set_local $$283
           (i32.add
-            (get_local $$279)
-            (get_local $$285)
+            (get_local $$310)
+            (get_local $$283)
           )
         )
-        (set_local $$144
+        (set_local $$142
           (i64.load align=8
             (i32.add
-              (get_local $$285)
-              (get_local $$78)
+              (get_local $$283)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$74
+        (set_local $$70
           (i64.lt_u
-            (get_local $$47)
-            (get_local $$42)
+            (get_local $$44)
+            (i64.const 128)
           )
         )
-        (set_local $$128
+        (set_local $$126
           (i64.load offset=192 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$129
+        (set_local $$127
           (i64.load offset=208 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$130
+        (set_local $$128
           (i64.load offset=224 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$75
+        (set_local $$71
           (i64.load offset=240 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$94
+        (set_local $$91
           (i64.load offset=48 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$95
+        (set_local $$92
           (i64.load offset=176 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$71
+        (set_local $$67
           (i64.lt_u
-            (get_local $$45)
-            (get_local $$42)
+            (get_local $$43)
+            (i64.const 128)
           )
         )
-        (set_local $$76
+        (set_local $$72
           (i64.lt_u
-            (get_local $$47)
             (get_local $$44)
+            (i64.const 256)
           )
         )
-        (set_local $$110
+        (set_local $$107
           (i64.load offset=128 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$111
+        (set_local $$108
           (i64.load offset=144 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$112
+        (set_local $$109
           (i64.load offset=160 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$96
+        (set_local $$93
           (i64.eq
-            (get_local $$47)
-            (get_local $$51)
+            (get_local $$44)
+            (i64.const 0)
           )
         )
-        (set_local $$113
+        (set_local $$110
           (i64.eq
-            (get_local $$45)
-            (get_local $$51)
+            (get_local $$43)
+            (i64.const 0)
           )
         )
-        (set_local $$286
+        (set_local $$284
           (i32.const 768)
         )
-        (set_local $$286
+        (set_local $$284
           (i32.add
-            (get_local $$279)
-            (get_local $$286)
+            (get_local $$310)
+            (get_local $$284)
           )
         )
-        (set_local $$47
+        (set_local $$44
           (i64.load align=8
             (i32.add
-              (get_local $$286)
-              (get_local $$78)
+              (get_local $$284)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$287
+        (set_local $$285
           (i32.const 784)
         )
-        (set_local $$287
+        (set_local $$285
           (i32.add
-            (get_local $$279)
-            (get_local $$287)
+            (get_local $$310)
+            (get_local $$285)
           )
         )
-        (set_local $$45
+        (set_local $$43
           (i64.load align=8
             (i32.add
-              (get_local $$287)
-              (get_local $$78)
+              (get_local $$285)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$288
+        (set_local $$286
           (i32.const 800)
         )
-        (set_local $$288
+        (set_local $$286
           (i32.add
-            (get_local $$279)
-            (get_local $$288)
+            (get_local $$310)
+            (get_local $$286)
           )
         )
-        (set_local $$145
+        (set_local $$143
           (i64.load align=8
             (i32.add
-              (get_local $$288)
-              (get_local $$78)
+              (get_local $$286)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$289
+        (set_local $$287
           (i32.const 992)
         )
-        (set_local $$289
+        (set_local $$287
           (i32.add
-            (get_local $$279)
-            (get_local $$289)
+            (get_local $$310)
+            (get_local $$287)
           )
         )
-        (set_local $$146
+        (set_local $$144
           (i64.load align=8
             (i32.add
-              (get_local $$289)
-              (get_local $$78)
+              (get_local $$287)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$290
+        (set_local $$288
           (i32.const 832)
         )
-        (set_local $$290
+        (set_local $$288
           (i32.add
-            (get_local $$279)
-            (get_local $$290)
+            (get_local $$310)
+            (get_local $$288)
           )
         )
-        (set_local $$147
+        (set_local $$145
           (i64.load align=8
             (i32.add
-              (get_local $$290)
-              (get_local $$78)
+              (get_local $$288)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$291
+        (set_local $$289
           (i32.const 384)
         )
-        (set_local $$291
+        (set_local $$289
           (i32.add
-            (get_local $$279)
-            (get_local $$291)
+            (get_local $$310)
+            (get_local $$289)
           )
         )
-        (set_local $$148
+        (set_local $$146
           (i64.load align=8
             (i32.add
-              (get_local $$291)
-              (get_local $$78)
+              (get_local $$289)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$292
+        (set_local $$290
           (i32.const 400)
         )
-        (set_local $$292
+        (set_local $$290
           (i32.add
-            (get_local $$279)
-            (get_local $$292)
+            (get_local $$310)
+            (get_local $$290)
           )
         )
-        (set_local $$149
+        (set_local $$147
           (i64.load align=8
             (i32.add
-              (get_local $$292)
-              (get_local $$78)
+              (get_local $$290)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$136
+        (set_local $$134
           (i64.load offset=384 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$137
+        (set_local $$135
           (i64.load offset=400 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$293
+        (set_local $$291
           (i32.const 416)
         )
-        (set_local $$293
+        (set_local $$291
           (i32.add
-            (get_local $$279)
-            (get_local $$293)
+            (get_local $$310)
+            (get_local $$291)
           )
         )
-        (set_local $$150
+        (set_local $$148
           (i64.load align=8
             (i32.add
-              (get_local $$293)
-              (get_local $$78)
+              (get_local $$291)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$138
+        (set_local $$136
           (i64.load offset=416 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$59
+        (set_local $$55
           (i64.load offset=432 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$58
+        (set_local $$54
           (i64.lt_u
             (get_local $$37)
-            (get_local $$42)
+            (i64.const 128)
           )
         )
-        (set_local $$139
+        (set_local $$137
           (i64.eq
             (get_local $$37)
-            (get_local $$51)
+            (i64.const 0)
           )
         )
-        (set_local $$294
+        (set_local $$292
           (i32.const 320)
         )
-        (set_local $$294
+        (set_local $$292
           (i32.add
-            (get_local $$279)
-            (get_local $$294)
+            (get_local $$310)
+            (get_local $$292)
           )
         )
         (set_local $$37
           (i64.load align=8
             (i32.add
-              (get_local $$294)
-              (get_local $$78)
+              (get_local $$292)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$54
+        (set_local $$50
           (i64.load offset=352 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$55
+        (set_local $$51
           (i64.load offset=336 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$56
+        (set_local $$52
           (i64.load offset=368 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$295
+        (set_local $$293
           (i32.const 448)
         )
-        (set_local $$295
+        (set_local $$293
           (i32.add
-            (get_local $$279)
-            (get_local $$295)
+            (get_local $$310)
+            (get_local $$293)
           )
         )
-        (set_local $$151
+        (set_local $$149
           (i64.load align=8
             (i32.add
-              (get_local $$295)
-              (get_local $$78)
+              (get_local $$293)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$53
+        (set_local $$49
           (i64.lt_u
             (get_local $$35)
-            (get_local $$42)
+            (i64.const 128)
           )
         )
-        (set_local $$97
+        (set_local $$94
           (i64.load offset=288 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$98
+        (set_local $$95
           (i64.load offset=272 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$99
+        (set_local $$96
           (i64.load offset=304 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$118
+        (set_local $$116
           (i64.load offset=256 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$140
+        (set_local $$138
           (i64.load offset=320 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$141
+        (set_local $$139
           (i64.load offset=448 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$57
+        (set_local $$53
           (i64.eq
             (get_local $$35)
-            (get_local $$51)
+            (i64.const 0)
           )
         )
-        (set_local $$60
+        (set_local $$56
           (i64.lt_u
             (get_local $$35)
-            (get_local $$44)
+            (i64.const 256)
           )
         )
-        (set_local $$296
+        (set_local $$294
           (i32.const 736)
         )
-        (set_local $$296
+        (set_local $$294
           (i32.add
-            (get_local $$279)
-            (get_local $$296)
+            (get_local $$310)
+            (get_local $$294)
           )
         )
         (set_local $$35
           (i64.load align=8
             (i32.add
-              (get_local $$296)
-              (get_local $$78)
+              (get_local $$294)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$297
+        (set_local $$295
           (i32.const 720)
         )
-        (set_local $$297
+        (set_local $$295
           (i32.add
-            (get_local $$279)
-            (get_local $$297)
+            (get_local $$310)
+            (get_local $$295)
           )
         )
-        (set_local $$159
+        (set_local $$157
           (i64.load align=8
             (i32.add
-              (get_local $$297)
-              (get_local $$78)
+              (get_local $$295)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$62
+        (set_local $$58
           (i64.load offset=864 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$63
+        (set_local $$59
           (i64.load offset=848 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$64
+        (set_local $$60
           (i64.load offset=880 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$298
+        (set_local $$296
           (i32.const 752)
         )
-        (set_local $$298
+        (set_local $$296
           (i32.add
-            (get_local $$279)
-            (get_local $$298)
+            (get_local $$310)
+            (get_local $$296)
           )
         )
-        (set_local $$160
+        (set_local $$158
           (i64.load align=8
             (i32.add
-              (get_local $$298)
-              (get_local $$78)
+              (get_local $$296)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$61
+        (set_local $$57
           (i64.lt_u
-            (get_local $$40)
-            (get_local $$42)
+            (get_local $$39)
+            (i64.const 128)
           )
         )
-        (set_local $$152
+        (set_local $$150
           (i64.load offset=736 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$153
+        (set_local $$151
           (i64.load offset=720 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$154
+        (set_local $$152
           (i64.load offset=752 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$65
+        (set_local $$61
           (i64.eq
-            (get_local $$40)
-            (get_local $$51)
+            (get_local $$39)
+            (i64.const 0)
           )
         )
-        (set_local $$299
+        (set_local $$297
           (i32.const 592)
         )
-        (set_local $$299
+        (set_local $$297
           (i32.add
-            (get_local $$279)
-            (get_local $$299)
+            (get_local $$310)
+            (get_local $$297)
           )
         )
-        (set_local $$40
+        (set_local $$39
           (i64.load align=8
             (i32.add
-              (get_local $$299)
-              (get_local $$78)
+              (get_local $$297)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$300
+        (set_local $$298
           (i32.const 608)
         )
-        (set_local $$300
+        (set_local $$298
           (i32.add
-            (get_local $$279)
-            (get_local $$300)
+            (get_local $$310)
+            (get_local $$298)
           )
         )
-        (set_local $$161
+        (set_local $$159
           (i64.load align=8
             (i32.add
-              (get_local $$300)
-              (get_local $$78)
+              (get_local $$298)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$301
+        (set_local $$299
           (i32.const 624)
         )
-        (set_local $$301
+        (set_local $$299
           (i32.add
-            (get_local $$279)
-            (get_local $$301)
+            (get_local $$310)
+            (get_local $$299)
           )
         )
-        (set_local $$162
+        (set_local $$160
           (i64.load align=8
             (i32.add
-              (get_local $$301)
-              (get_local $$78)
+              (get_local $$299)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$302
+        (set_local $$300
           (i32.const 688)
         )
-        (set_local $$302
+        (set_local $$300
           (i32.add
-            (get_local $$279)
-            (get_local $$302)
+            (get_local $$310)
+            (get_local $$300)
           )
         )
-        (set_local $$163
+        (set_local $$161
           (i64.load align=8
             (i32.add
-              (get_local $$302)
-              (get_local $$78)
+              (get_local $$300)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$303
+        (set_local $$301
           (i32.const 640)
         )
-        (set_local $$303
+        (set_local $$301
           (i32.add
-            (get_local $$279)
-            (get_local $$303)
+            (get_local $$310)
+            (get_local $$301)
           )
         )
-        (set_local $$169
+        (set_local $$167
           (i64.load align=8
             (i32.add
-              (get_local $$303)
-              (get_local $$78)
+              (get_local $$301)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$304
+        (set_local $$302
           (i32.const 656)
         )
-        (set_local $$304
+        (set_local $$302
           (i32.add
-            (get_local $$279)
-            (get_local $$304)
+            (get_local $$310)
+            (get_local $$302)
           )
         )
-        (set_local $$170
+        (set_local $$168
           (i64.load align=8
             (i32.add
-              (get_local $$304)
-              (get_local $$78)
+              (get_local $$302)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$131
+        (set_local $$129
           (i64.load offset=768 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$132
+        (set_local $$130
           (i64.load offset=784 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$133
+        (set_local $$131
           (i64.load offset=800 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$305
+        (set_local $$303
           (i32.const 672)
         )
-        (set_local $$305
+        (set_local $$303
           (i32.add
-            (get_local $$279)
-            (get_local $$305)
+            (get_local $$310)
+            (get_local $$303)
           )
         )
-        (set_local $$171
+        (set_local $$169
           (i64.load align=8
             (i32.add
-              (get_local $$305)
-              (get_local $$78)
+              (get_local $$303)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$164
+        (set_local $$162
           (i64.load offset=640 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$165
+        (set_local $$163
           (i64.load offset=656 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$166
+        (set_local $$164
           (i64.load offset=672 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$72
+        (set_local $$68
           (i64.load offset=816 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$158
+        (set_local $$156
           (i64.load offset=688 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$306
+        (set_local $$304
           (i32.const 576)
         )
-        (set_local $$306
+        (set_local $$304
           (i32.add
-            (get_local $$279)
-            (get_local $$306)
+            (get_local $$310)
+            (get_local $$304)
           )
         )
-        (set_local $$172
+        (set_local $$170
           (i64.load align=8
             (i32.add
-              (get_local $$306)
-              (get_local $$78)
+              (get_local $$304)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$135
+        (set_local $$133
           (i64.load offset=832 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$168
+        (set_local $$166
           (i64.load offset=704 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$307
+        (set_local $$305
           (i32.const 704)
         )
-        (set_local $$307
+        (set_local $$305
           (i32.add
-            (get_local $$279)
-            (get_local $$307)
+            (get_local $$310)
+            (get_local $$305)
           )
         )
-        (set_local $$173
+        (set_local $$171
           (i64.load align=8
             (i32.add
-              (get_local $$307)
-              (get_local $$78)
+              (get_local $$305)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$308
+        (set_local $$306
           (i32.const 528)
         )
-        (set_local $$308
+        (set_local $$306
           (i32.add
-            (get_local $$279)
-            (get_local $$308)
+            (get_local $$310)
+            (get_local $$306)
           )
         )
-        (set_local $$177
+        (set_local $$175
           (i64.load align=8
             (i32.add
-              (get_local $$308)
-              (get_local $$78)
+              (get_local $$306)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$309
+        (set_local $$307
           (i32.const 544)
         )
-        (set_local $$309
+        (set_local $$307
           (i32.add
-            (get_local $$279)
-            (get_local $$309)
+            (get_local $$310)
+            (get_local $$307)
           )
         )
-        (set_local $$178
+        (set_local $$176
           (i64.load align=8
             (i32.add
-              (get_local $$309)
-              (get_local $$78)
+              (get_local $$307)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$310
+        (set_local $$308
           (i32.const 560)
         )
-        (set_local $$310
+        (set_local $$308
           (i32.add
-            (get_local $$279)
             (get_local $$310)
+            (get_local $$308)
           )
         )
-        (set_local $$179
+        (set_local $$177
           (i64.load align=8
             (i32.add
-              (get_local $$310)
-              (get_local $$78)
+              (get_local $$308)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$67
+        (set_local $$63
           (i64.load offset=1008 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$68
+        (set_local $$64
           (i64.load offset=960 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$69
+        (set_local $$65
           (i64.load offset=976 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$66
+        (set_local $$62
           (i64.lt_u
             (get_local $$17)
-            (get_local $$42)
+            (i64.const 128)
           )
         )
-        (set_local $$70
+        (set_local $$66
           (i64.eq
             (get_local $$17)
-            (get_local $$51)
+            (i64.const 0)
           )
         )
-        (set_local $$73
+        (set_local $$69
           (i64.lt_u
             (get_local $$17)
-            (get_local $$44)
+            (i64.const 256)
           )
         )
-        (set_local $$77
+        (set_local $$73
           (i64.lt_u
             (get_local $$17)
-            (get_local $$46)
+            (i64.const 512)
           )
         )
         (set_local $$17
           (i64.load offset=912 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$42
+        (set_local $$84
           (i64.load offset=928 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$44
+        (set_local $$85
           (i64.load offset=944 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$46
+        (set_local $$115
           (i64.load offset=896 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$134
+        (set_local $$132
           (i64.load offset=992 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$155
+        (set_local $$153
           (i64.load offset=592 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$156
+        (set_local $$154
           (i64.load offset=608 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$157
+        (set_local $$155
           (i64.load offset=624 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$167
+        (set_local $$165
           (i64.load offset=576 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$174
+        (set_local $$172
           (i64.load offset=528 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$175
+        (set_local $$173
           (i64.load offset=544 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$176
+        (set_local $$174
           (i64.load offset=560 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$180
+        (set_local $$178
           (i64.load offset=512 align=8
-            (get_local $$279)
+            (get_local $$310)
           )
         )
-        (set_local $$311
+        (set_local $$309
           (i32.const 512)
         )
-        (set_local $$311
+        (set_local $$309
           (i32.add
-            (get_local $$279)
-            (get_local $$311)
+            (get_local $$310)
+            (get_local $$309)
           )
         )
         (i64.store align=8
           (i32.add
             (get_local $$0)
-            (get_local $$78)
+            (i32.const 8)
           )
           (i64.select
-            (get_local $$77)
+            (get_local $$73)
             (i64.select
-              (get_local $$73)
+              (get_local $$69)
               (i64.select
-                (get_local $$66)
+                (get_local $$62)
                 (i64.load align=8
                   (i32.add
-                    (get_local $$311)
-                    (get_local $$78)
+                    (get_local $$309)
+                    (i32.const 8)
                   )
                 )
-                (get_local $$51)
+                (i64.const 0)
               )
-              (get_local $$51)
+              (i64.const 0)
             )
-            (get_local $$51)
+            (i64.const 0)
           )
         )
         (i64.store align=8
           (get_local $$0)
           (i64.select
-            (get_local $$77)
+            (get_local $$73)
             (i64.select
-              (get_local $$73)
+              (get_local $$69)
               (i64.select
-                (get_local $$66)
-                (get_local $$180)
-                (get_local $$51)
+                (get_local $$62)
+                (get_local $$178)
+                (i64.const 0)
               )
-              (get_local $$51)
+              (i64.const 0)
             )
-            (get_local $$51)
+            (i64.const 0)
           )
         )
         (i64.store align=8
@@ -3042,24 +3018,24 @@
             (i32.const 24)
           )
           (i64.select
-            (get_local $$77)
+            (get_local $$73)
             (i64.select
-              (get_local $$73)
+              (get_local $$69)
               (i64.select
-                (get_local $$70)
+                (get_local $$66)
                 (get_local $$4)
                 (i64.select
-                  (get_local $$66)
+                  (get_local $$62)
                   (i64.or
-                    (get_local $$177)
-                    (get_local $$178)
+                    (get_local $$175)
+                    (get_local $$176)
                   )
-                  (get_local $$179)
+                  (get_local $$177)
                 )
               )
-              (get_local $$51)
+              (i64.const 0)
             )
-            (get_local $$51)
+            (i64.const 0)
           )
         )
         (i64.store align=8
@@ -3068,24 +3044,24 @@
             (i32.const 16)
           )
           (i64.select
-            (get_local $$77)
+            (get_local $$73)
             (i64.select
-              (get_local $$73)
+              (get_local $$69)
               (i64.select
-                (get_local $$70)
+                (get_local $$66)
                 (get_local $$3)
                 (i64.select
-                  (get_local $$66)
+                  (get_local $$62)
                   (i64.or
-                    (get_local $$174)
-                    (get_local $$175)
+                    (get_local $$172)
+                    (get_local $$173)
                   )
-                  (get_local $$176)
+                  (get_local $$174)
                 )
               )
-              (get_local $$51)
+              (i64.const 0)
             )
-            (get_local $$51)
+            (i64.const 0)
           )
         )
         (i64.store align=8
@@ -3094,46 +3070,46 @@
             (i32.const 56)
           )
           (i64.select
-            (get_local $$77)
+            (get_local $$73)
             (i64.select
-              (get_local $$70)
+              (get_local $$66)
               (get_local $$8)
               (i64.select
-                (get_local $$73)
+                (get_local $$69)
                 (i64.or
                   (i64.select
-                    (get_local $$70)
+                    (get_local $$66)
                     (get_local $$8)
                     (i64.select
-                      (get_local $$66)
+                      (get_local $$62)
                       (i64.or
-                        (get_local $$40)
-                        (get_local $$161)
+                        (get_local $$39)
+                        (get_local $$159)
                       )
-                      (get_local $$162)
+                      (get_local $$160)
                     )
                   )
                   (i64.select
-                    (get_local $$71)
-                    (get_local $$163)
-                    (get_local $$51)
+                    (get_local $$67)
+                    (get_local $$161)
+                    (i64.const 0)
                   )
                 )
                 (i64.select
-                  (get_local $$65)
+                  (get_local $$61)
                   (get_local $$4)
                   (i64.select
-                    (get_local $$61)
+                    (get_local $$57)
                     (i64.or
-                      (get_local $$159)
+                      (get_local $$157)
                       (get_local $$35)
                     )
-                    (get_local $$160)
+                    (get_local $$158)
                   )
                 )
               )
             )
-            (get_local $$51)
+            (i64.const 0)
           )
         )
         (i64.store align=8
@@ -3142,46 +3118,46 @@
             (i32.const 48)
           )
           (i64.select
-            (get_local $$77)
+            (get_local $$73)
             (i64.select
-              (get_local $$70)
+              (get_local $$66)
               (get_local $$7)
               (i64.select
-                (get_local $$73)
+                (get_local $$69)
                 (i64.or
                   (i64.select
-                    (get_local $$70)
+                    (get_local $$66)
                     (get_local $$7)
                     (i64.select
-                      (get_local $$66)
+                      (get_local $$62)
                       (i64.or
-                        (get_local $$155)
-                        (get_local $$156)
+                        (get_local $$153)
+                        (get_local $$154)
                       )
-                      (get_local $$157)
+                      (get_local $$155)
                     )
                   )
                   (i64.select
-                    (get_local $$71)
-                    (get_local $$158)
-                    (get_local $$51)
+                    (get_local $$67)
+                    (get_local $$156)
+                    (i64.const 0)
                   )
                 )
                 (i64.select
-                  (get_local $$65)
+                  (get_local $$61)
                   (get_local $$3)
                   (i64.select
-                    (get_local $$61)
+                    (get_local $$57)
                     (i64.or
-                      (get_local $$153)
-                      (get_local $$152)
+                      (get_local $$151)
+                      (get_local $$150)
                     )
-                    (get_local $$154)
+                    (get_local $$152)
                   )
                 )
               )
             )
-            (get_local $$51)
+            (i64.const 0)
           )
         )
         (i64.store align=8
@@ -3190,39 +3166,39 @@
             (i32.const 40)
           )
           (i64.select
-            (get_local $$77)
+            (get_local $$73)
             (i64.select
-              (get_local $$70)
+              (get_local $$66)
               (get_local $$6)
               (i64.select
-                (get_local $$73)
+                (get_local $$69)
                 (i64.or
                   (i64.select
-                    (get_local $$66)
-                    (get_local $$172)
-                    (get_local $$51)
+                    (get_local $$62)
+                    (get_local $$170)
+                    (i64.const 0)
                   )
                   (i64.select
-                    (get_local $$113)
+                    (get_local $$110)
                     (get_local $$2)
                     (i64.select
-                      (get_local $$71)
+                      (get_local $$67)
                       (i64.or
-                        (get_local $$169)
-                        (get_local $$170)
+                        (get_local $$167)
+                        (get_local $$168)
                       )
-                      (get_local $$171)
+                      (get_local $$169)
                     )
                   )
                 )
                 (i64.select
-                  (get_local $$61)
-                  (get_local $$173)
-                  (get_local $$51)
+                  (get_local $$57)
+                  (get_local $$171)
+                  (i64.const 0)
                 )
               )
             )
-            (get_local $$51)
+            (i64.const 0)
           )
         )
         (i64.store align=8
@@ -3231,39 +3207,39 @@
             (i32.const 32)
           )
           (i64.select
-            (get_local $$77)
+            (get_local $$73)
             (i64.select
-              (get_local $$70)
+              (get_local $$66)
               (get_local $$5)
               (i64.select
-                (get_local $$73)
+                (get_local $$69)
                 (i64.or
                   (i64.select
-                    (get_local $$66)
-                    (get_local $$167)
-                    (get_local $$51)
+                    (get_local $$62)
+                    (get_local $$165)
+                    (i64.const 0)
                   )
                   (i64.select
-                    (get_local $$113)
+                    (get_local $$110)
                     (get_local $$1)
                     (i64.select
-                      (get_local $$71)
+                      (get_local $$67)
                       (i64.or
-                        (get_local $$164)
-                        (get_local $$165)
+                        (get_local $$162)
+                        (get_local $$163)
                       )
-                      (get_local $$166)
+                      (get_local $$164)
                     )
                   )
                 )
                 (i64.select
-                  (get_local $$61)
-                  (get_local $$168)
-                  (get_local $$51)
+                  (get_local $$57)
+                  (get_local $$166)
+                  (i64.const 0)
                 )
               )
             )
-            (get_local $$51)
+            (i64.const 0)
           )
         )
         (i64.store align=8
@@ -3272,93 +3248,93 @@
             (i32.const 120)
           )
           (i64.select
-            (get_local $$70)
+            (get_local $$66)
             (get_local $$16)
             (i64.select
-              (get_local $$77)
+              (get_local $$73)
               (i64.or
                 (i64.select
-                  (get_local $$70)
+                  (get_local $$66)
                   (get_local $$16)
                   (i64.select
-                    (get_local $$73)
+                    (get_local $$69)
                     (i64.or
                       (i64.select
-                        (get_local $$70)
+                        (get_local $$66)
                         (get_local $$16)
                         (i64.select
-                          (get_local $$66)
+                          (get_local $$62)
                           (i64.or
-                            (get_local $$84)
-                            (get_local $$85)
+                            (get_local $$79)
+                            (get_local $$80)
                           )
-                          (get_local $$86)
+                          (get_local $$81)
                         )
                       )
                       (i64.select
-                        (get_local $$71)
-                        (get_local $$87)
-                        (get_local $$51)
+                        (get_local $$67)
+                        (get_local $$82)
+                        (i64.const 0)
                       )
                     )
                     (i64.select
-                      (get_local $$65)
+                      (get_local $$61)
                       (get_local $$12)
                       (i64.select
-                        (get_local $$61)
+                        (get_local $$57)
                         (i64.or
-                          (get_local $$82)
-                          (get_local $$81)
+                          (get_local $$77)
+                          (get_local $$76)
                         )
-                        (get_local $$83)
+                        (get_local $$78)
                       )
                     )
                   )
                 )
                 (i64.select
-                  (get_local $$76)
+                  (get_local $$72)
                   (i64.select
-                    (get_local $$74)
-                    (get_local $$88)
-                    (get_local $$51)
+                    (get_local $$70)
+                    (get_local $$83)
+                    (i64.const 0)
                   )
-                  (get_local $$51)
+                  (i64.const 0)
                 )
               )
               (i64.select
-                (get_local $$57)
+                (get_local $$53)
                 (get_local $$8)
                 (i64.select
-                  (get_local $$60)
+                  (get_local $$56)
                   (i64.or
                     (i64.select
-                      (get_local $$57)
+                      (get_local $$53)
                       (get_local $$8)
                       (i64.select
-                        (get_local $$53)
+                        (get_local $$49)
                         (i64.or
-                          (get_local $$41)
+                          (get_local $$40)
                           (get_local $$33)
                         )
-                        (get_local $$79)
+                        (get_local $$74)
                       )
                     )
                     (i64.select
-                      (get_local $$58)
-                      (get_local $$80)
-                      (get_local $$51)
+                      (get_local $$54)
+                      (get_local $$75)
+                      (i64.const 0)
                     )
                   )
                   (i64.select
-                    (get_local $$52)
+                    (get_local $$48)
                     (get_local $$4)
                     (i64.select
-                      (get_local $$50)
+                      (get_local $$47)
                       (i64.or
-                        (get_local $$43)
-                        (get_local $$39)
+                        (get_local $$38)
+                        (get_local $$41)
                       )
-                      (get_local $$38)
+                      (get_local $$36)
                     )
                   )
                 )
@@ -3372,93 +3348,93 @@
             (i32.const 112)
           )
           (i64.select
-            (get_local $$70)
+            (get_local $$66)
             (get_local $$15)
             (i64.select
-              (get_local $$77)
+              (get_local $$73)
               (i64.or
                 (i64.select
-                  (get_local $$70)
+                  (get_local $$66)
                   (get_local $$15)
                   (i64.select
-                    (get_local $$73)
+                    (get_local $$69)
                     (i64.or
                       (i64.select
-                        (get_local $$70)
+                        (get_local $$66)
                         (get_local $$15)
                         (i64.select
-                          (get_local $$66)
+                          (get_local $$62)
                           (i64.or
-                            (get_local $$67)
-                            (get_local $$68)
+                            (get_local $$63)
+                            (get_local $$64)
                           )
-                          (get_local $$69)
+                          (get_local $$65)
                         )
                       )
                       (i64.select
-                        (get_local $$71)
-                        (get_local $$72)
-                        (get_local $$51)
+                        (get_local $$67)
+                        (get_local $$68)
+                        (i64.const 0)
                       )
                     )
                     (i64.select
-                      (get_local $$65)
+                      (get_local $$61)
                       (get_local $$11)
                       (i64.select
-                        (get_local $$61)
+                        (get_local $$57)
                         (i64.or
-                          (get_local $$63)
-                          (get_local $$62)
+                          (get_local $$59)
+                          (get_local $$58)
                         )
-                        (get_local $$64)
+                        (get_local $$60)
                       )
                     )
                   )
                 )
                 (i64.select
-                  (get_local $$76)
+                  (get_local $$72)
                   (i64.select
-                    (get_local $$74)
-                    (get_local $$75)
-                    (get_local $$51)
+                    (get_local $$70)
+                    (get_local $$71)
+                    (i64.const 0)
                   )
-                  (get_local $$51)
+                  (i64.const 0)
                 )
               )
               (i64.select
-                (get_local $$57)
+                (get_local $$53)
                 (get_local $$7)
                 (i64.select
-                  (get_local $$60)
+                  (get_local $$56)
                   (i64.or
                     (i64.select
-                      (get_local $$57)
+                      (get_local $$53)
                       (get_local $$7)
                       (i64.select
-                        (get_local $$53)
+                        (get_local $$49)
                         (i64.or
-                          (get_local $$55)
-                          (get_local $$54)
+                          (get_local $$51)
+                          (get_local $$50)
                         )
-                        (get_local $$56)
+                        (get_local $$52)
                       )
                     )
                     (i64.select
-                      (get_local $$58)
-                      (get_local $$59)
-                      (get_local $$51)
+                      (get_local $$54)
+                      (get_local $$55)
+                      (i64.const 0)
                     )
                   )
                   (i64.select
-                    (get_local $$52)
+                    (get_local $$48)
                     (get_local $$3)
                     (i64.select
-                      (get_local $$50)
+                      (get_local $$47)
                       (i64.or
-                        (get_local $$49)
                         (get_local $$34)
+                        (get_local $$42)
                       )
-                      (get_local $$36)
+                      (get_local $$46)
                     )
                   )
                 )
@@ -3472,87 +3448,87 @@
             (i32.const 104)
           )
           (i64.select
-            (get_local $$70)
+            (get_local $$66)
             (get_local $$14)
             (i64.select
-              (get_local $$77)
+              (get_local $$73)
               (i64.or
                 (i64.select
-                  (get_local $$70)
+                  (get_local $$66)
                   (get_local $$14)
                   (i64.select
-                    (get_local $$73)
+                    (get_local $$69)
                     (i64.or
                       (i64.select
-                        (get_local $$66)
-                        (get_local $$146)
-                        (get_local $$51)
+                        (get_local $$62)
+                        (get_local $$144)
+                        (i64.const 0)
                       )
                       (i64.select
-                        (get_local $$113)
+                        (get_local $$110)
                         (get_local $$10)
                         (i64.select
-                          (get_local $$71)
+                          (get_local $$67)
                           (i64.or
-                            (get_local $$47)
-                            (get_local $$45)
+                            (get_local $$44)
+                            (get_local $$43)
                           )
-                          (get_local $$145)
+                          (get_local $$143)
                         )
                       )
                     )
                     (i64.select
-                      (get_local $$61)
-                      (get_local $$147)
-                      (get_local $$51)
+                      (get_local $$57)
+                      (get_local $$145)
+                      (i64.const 0)
                     )
                   )
                 )
                 (i64.select
-                  (get_local $$76)
+                  (get_local $$72)
                   (i64.select
-                    (get_local $$96)
+                    (get_local $$93)
                     (get_local $$6)
                     (i64.select
-                      (get_local $$74)
+                      (get_local $$70)
                       (i64.or
-                        (get_local $$142)
-                        (get_local $$143)
+                        (get_local $$140)
+                        (get_local $$141)
                       )
-                      (get_local $$144)
+                      (get_local $$142)
                     )
                   )
-                  (get_local $$51)
+                  (i64.const 0)
                 )
               )
               (i64.select
-                (get_local $$57)
+                (get_local $$53)
                 (get_local $$6)
                 (i64.select
-                  (get_local $$60)
+                  (get_local $$56)
                   (i64.or
                     (i64.select
-                      (get_local $$53)
+                      (get_local $$49)
                       (get_local $$37)
-                      (get_local $$51)
+                      (i64.const 0)
                     )
                     (i64.select
-                      (get_local $$139)
+                      (get_local $$137)
                       (get_local $$2)
                       (i64.select
-                        (get_local $$58)
+                        (get_local $$54)
                         (i64.or
-                          (get_local $$148)
-                          (get_local $$149)
+                          (get_local $$146)
+                          (get_local $$147)
                         )
-                        (get_local $$150)
+                        (get_local $$148)
                       )
                     )
                   )
                   (i64.select
-                    (get_local $$50)
-                    (get_local $$151)
-                    (get_local $$51)
+                    (get_local $$47)
+                    (get_local $$149)
+                    (i64.const 0)
                   )
                 )
               )
@@ -3565,87 +3541,87 @@
             (i32.const 96)
           )
           (i64.select
-            (get_local $$70)
+            (get_local $$66)
             (get_local $$13)
             (i64.select
-              (get_local $$77)
+              (get_local $$73)
               (i64.or
                 (i64.select
-                  (get_local $$70)
+                  (get_local $$66)
                   (get_local $$13)
                   (i64.select
-                    (get_local $$73)
+                    (get_local $$69)
                     (i64.or
                       (i64.select
-                        (get_local $$66)
-                        (get_local $$134)
-                        (get_local $$51)
+                        (get_local $$62)
+                        (get_local $$132)
+                        (i64.const 0)
                       )
                       (i64.select
-                        (get_local $$113)
+                        (get_local $$110)
                         (get_local $$9)
                         (i64.select
-                          (get_local $$71)
+                          (get_local $$67)
                           (i64.or
-                            (get_local $$131)
-                            (get_local $$132)
+                            (get_local $$129)
+                            (get_local $$130)
                           )
-                          (get_local $$133)
+                          (get_local $$131)
                         )
                       )
                     )
                     (i64.select
-                      (get_local $$61)
-                      (get_local $$135)
-                      (get_local $$51)
+                      (get_local $$57)
+                      (get_local $$133)
+                      (i64.const 0)
                     )
                   )
                 )
                 (i64.select
-                  (get_local $$76)
+                  (get_local $$72)
                   (i64.select
-                    (get_local $$96)
+                    (get_local $$93)
                     (get_local $$5)
                     (i64.select
-                      (get_local $$74)
+                      (get_local $$70)
                       (i64.or
-                        (get_local $$128)
-                        (get_local $$129)
+                        (get_local $$126)
+                        (get_local $$127)
                       )
-                      (get_local $$130)
+                      (get_local $$128)
                     )
                   )
-                  (get_local $$51)
+                  (i64.const 0)
                 )
               )
               (i64.select
-                (get_local $$57)
+                (get_local $$53)
                 (get_local $$5)
                 (i64.select
-                  (get_local $$60)
+                  (get_local $$56)
                   (i64.or
                     (i64.select
-                      (get_local $$53)
-                      (get_local $$140)
-                      (get_local $$51)
+                      (get_local $$49)
+                      (get_local $$138)
+                      (i64.const 0)
                     )
                     (i64.select
-                      (get_local $$139)
+                      (get_local $$137)
                       (get_local $$1)
                       (i64.select
-                        (get_local $$58)
+                        (get_local $$54)
                         (i64.or
-                          (get_local $$136)
-                          (get_local $$137)
+                          (get_local $$134)
+                          (get_local $$135)
                         )
-                        (get_local $$138)
+                        (get_local $$136)
                       )
                     )
                   )
                   (i64.select
-                    (get_local $$50)
-                    (get_local $$141)
-                    (get_local $$51)
+                    (get_local $$47)
+                    (get_local $$139)
+                    (i64.const 0)
                   )
                 )
               )
@@ -3658,67 +3634,67 @@
             (i32.const 72)
           )
           (i64.select
-            (get_local $$70)
+            (get_local $$66)
             (get_local $$10)
             (i64.select
-              (get_local $$77)
+              (get_local $$73)
               (i64.or
                 (i64.select
-                  (get_local $$73)
+                  (get_local $$69)
                   (i64.select
-                    (get_local $$66)
-                    (get_local $$126)
-                    (get_local $$51)
+                    (get_local $$62)
+                    (get_local $$124)
+                    (i64.const 0)
                   )
-                  (get_local $$51)
+                  (i64.const 0)
                 )
                 (i64.select
-                  (get_local $$96)
+                  (get_local $$93)
                   (get_local $$2)
                   (i64.select
-                    (get_local $$76)
+                    (get_local $$72)
                     (i64.or
                       (i64.select
-                        (get_local $$96)
+                        (get_local $$93)
                         (get_local $$2)
                         (i64.select
-                          (get_local $$74)
+                          (get_local $$70)
                           (i64.or
-                            (get_local $$122)
-                            (get_local $$123)
+                            (get_local $$120)
+                            (get_local $$121)
                           )
-                          (get_local $$124)
+                          (get_local $$122)
                         )
                       )
                       (i64.select
-                        (get_local $$89)
-                        (get_local $$125)
-                        (get_local $$51)
+                        (get_local $$86)
+                        (get_local $$123)
+                        (i64.const 0)
                       )
                     )
                     (i64.select
-                      (get_local $$113)
+                      (get_local $$110)
                       (get_local $$6)
                       (i64.select
-                        (get_local $$71)
+                        (get_local $$67)
                         (i64.or
-                          (get_local $$119)
-                          (get_local $$120)
+                          (get_local $$117)
+                          (get_local $$118)
                         )
-                        (get_local $$121)
+                        (get_local $$119)
                       )
                     )
                   )
                 )
               )
               (i64.select
-                (get_local $$60)
+                (get_local $$56)
                 (i64.select
-                  (get_local $$53)
-                  (get_local $$127)
-                  (get_local $$51)
+                  (get_local $$49)
+                  (get_local $$125)
+                  (i64.const 0)
                 )
-                (get_local $$51)
+                (i64.const 0)
               )
             )
           )
@@ -3729,67 +3705,67 @@
             (i32.const 64)
           )
           (i64.select
-            (get_local $$70)
+            (get_local $$66)
             (get_local $$9)
             (i64.select
-              (get_local $$77)
+              (get_local $$73)
               (i64.or
                 (i64.select
-                  (get_local $$73)
+                  (get_local $$69)
                   (i64.select
-                    (get_local $$66)
-                    (get_local $$46)
-                    (get_local $$51)
+                    (get_local $$62)
+                    (get_local $$115)
+                    (i64.const 0)
                   )
-                  (get_local $$51)
+                  (i64.const 0)
                 )
                 (i64.select
-                  (get_local $$96)
+                  (get_local $$93)
                   (get_local $$1)
                   (i64.select
-                    (get_local $$76)
+                    (get_local $$72)
                     (i64.or
                       (i64.select
-                        (get_local $$96)
+                        (get_local $$93)
                         (get_local $$1)
                         (i64.select
-                          (get_local $$74)
+                          (get_local $$70)
                           (i64.or
-                            (get_local $$114)
-                            (get_local $$115)
+                            (get_local $$111)
+                            (get_local $$112)
                           )
-                          (get_local $$116)
+                          (get_local $$113)
                         )
                       )
                       (i64.select
-                        (get_local $$89)
-                        (get_local $$117)
-                        (get_local $$51)
+                        (get_local $$86)
+                        (get_local $$114)
+                        (i64.const 0)
                       )
                     )
                     (i64.select
-                      (get_local $$113)
+                      (get_local $$110)
                       (get_local $$5)
                       (i64.select
-                        (get_local $$71)
+                        (get_local $$67)
                         (i64.or
-                          (get_local $$110)
-                          (get_local $$111)
+                          (get_local $$107)
+                          (get_local $$108)
                         )
-                        (get_local $$112)
+                        (get_local $$109)
                       )
                     )
                   )
                 )
               )
               (i64.select
-                (get_local $$60)
+                (get_local $$56)
                 (i64.select
-                  (get_local $$53)
-                  (get_local $$118)
-                  (get_local $$51)
+                  (get_local $$49)
+                  (get_local $$116)
+                  (i64.const 0)
                 )
-                (get_local $$51)
+                (i64.const 0)
               )
             )
           )
@@ -3800,74 +3776,74 @@
             (i32.const 88)
           )
           (i64.select
-            (get_local $$70)
+            (get_local $$66)
             (get_local $$12)
             (i64.select
-              (get_local $$77)
+              (get_local $$73)
               (i64.or
                 (i64.select
-                  (get_local $$73)
+                  (get_local $$69)
                   (i64.select
-                    (get_local $$70)
+                    (get_local $$66)
                     (get_local $$12)
                     (i64.select
-                      (get_local $$66)
+                      (get_local $$62)
                       (i64.or
-                        (get_local $$100)
-                        (get_local $$101)
+                        (get_local $$97)
+                        (get_local $$98)
                       )
-                      (get_local $$102)
+                      (get_local $$99)
                     )
                   )
-                  (get_local $$51)
+                  (i64.const 0)
                 )
                 (i64.select
-                  (get_local $$96)
+                  (get_local $$93)
                   (get_local $$4)
                   (i64.select
-                    (get_local $$76)
+                    (get_local $$72)
                     (i64.or
                       (i64.select
-                        (get_local $$74)
-                        (get_local $$48)
-                        (get_local $$51)
+                        (get_local $$70)
+                        (get_local $$45)
+                        (i64.const 0)
                       )
                       (i64.select
-                        (get_local $$93)
+                        (get_local $$90)
                         (get_local $$8)
                         (i64.select
-                          (get_local $$89)
+                          (get_local $$86)
                           (i64.or
-                            (get_local $$103)
-                            (get_local $$104)
+                            (get_local $$100)
+                            (get_local $$101)
                           )
-                          (get_local $$105)
+                          (get_local $$102)
                         )
                       )
                     )
                     (i64.select
-                      (get_local $$71)
-                      (get_local $$106)
-                      (get_local $$51)
+                      (get_local $$67)
+                      (get_local $$103)
+                      (i64.const 0)
                     )
                   )
                 )
               )
               (i64.select
-                (get_local $$60)
+                (get_local $$56)
                 (i64.select
-                  (get_local $$57)
+                  (get_local $$53)
                   (get_local $$4)
                   (i64.select
-                    (get_local $$53)
+                    (get_local $$49)
                     (i64.or
-                      (get_local $$108)
-                      (get_local $$107)
+                      (get_local $$105)
+                      (get_local $$104)
                     )
-                    (get_local $$109)
+                    (get_local $$106)
                   )
                 )
-                (get_local $$51)
+                (i64.const 0)
               )
             )
           )
@@ -3878,94 +3854,94 @@
             (i32.const 80)
           )
           (i64.select
-            (get_local $$70)
+            (get_local $$66)
             (get_local $$11)
             (i64.select
-              (get_local $$77)
+              (get_local $$73)
               (i64.or
                 (i64.select
-                  (get_local $$73)
+                  (get_local $$69)
                   (i64.select
-                    (get_local $$70)
+                    (get_local $$66)
                     (get_local $$11)
                     (i64.select
-                      (get_local $$66)
+                      (get_local $$62)
                       (i64.or
                         (get_local $$17)
-                        (get_local $$42)
+                        (get_local $$84)
                       )
-                      (get_local $$44)
+                      (get_local $$85)
                     )
                   )
-                  (get_local $$51)
+                  (i64.const 0)
                 )
                 (i64.select
-                  (get_local $$96)
+                  (get_local $$93)
                   (get_local $$3)
                   (i64.select
-                    (get_local $$76)
+                    (get_local $$72)
                     (i64.or
                       (i64.select
-                        (get_local $$74)
-                        (get_local $$94)
-                        (get_local $$51)
+                        (get_local $$70)
+                        (get_local $$91)
+                        (i64.const 0)
                       )
                       (i64.select
-                        (get_local $$93)
+                        (get_local $$90)
                         (get_local $$7)
                         (i64.select
-                          (get_local $$89)
+                          (get_local $$86)
                           (i64.or
-                            (get_local $$90)
-                            (get_local $$91)
+                            (get_local $$87)
+                            (get_local $$88)
                           )
-                          (get_local $$92)
+                          (get_local $$89)
                         )
                       )
                     )
                     (i64.select
-                      (get_local $$71)
-                      (get_local $$95)
-                      (get_local $$51)
+                      (get_local $$67)
+                      (get_local $$92)
+                      (i64.const 0)
                     )
                   )
                 )
               )
               (i64.select
-                (get_local $$60)
+                (get_local $$56)
                 (i64.select
-                  (get_local $$57)
+                  (get_local $$53)
                   (get_local $$3)
                   (i64.select
-                    (get_local $$53)
+                    (get_local $$49)
                     (i64.or
-                      (get_local $$98)
-                      (get_local $$97)
+                      (get_local $$95)
+                      (get_local $$94)
                     )
-                    (get_local $$99)
+                    (get_local $$96)
                   )
                 )
-                (get_local $$51)
+                (i64.const 0)
               )
             )
           )
         )
-        (set_local $$183
+        (set_local $$181
           (i32.const 1024)
         )
-        (set_local $$279
+        (set_local $$310
           (i32.add
-            (get_local $$279)
-            (get_local $$183)
+            (get_local $$310)
+            (get_local $$181)
           )
         )
-        (set_local $$183
+        (set_local $$181
           (i32.const 1)
         )
-        (set_local $$279
+        (set_local $$310
           (i32.store align=4
-            (get_local $$183)
-            (get_local $$279)
+            (get_local $$181)
+            (get_local $$310)
           )
         )
         (br $fake_return_waka123)
diff --git a/test/llvm_autogenerated/llvm-to-s.py b/test/llvm_autogenerated/llvm-to-s.py
index aea7d722c..05a03da2b 100755
--- a/test/llvm_autogenerated/llvm-to-s.py
+++ b/test/llvm_autogenerated/llvm-to-s.py
@@ -57,6 +57,7 @@ def main(args):
                      'returned',   # external global symbol
                      'vtable',     # external global symbol
                      'offset-folding', # external global symbol
+                     'address-offsets', # external global symbol
                      ]
         if name_noext in BLACKLIST:
           continue
diff --git a/test/llvm_autogenerated/load-ext.s b/test/llvm_autogenerated/load-ext.s
index a7c4e6810..1bda3b4f7 100644
--- a/test/llvm_autogenerated/load-ext.s
+++ b/test/llvm_autogenerated/load-ext.s
@@ -7,6 +7,7 @@ sext_i8_i32:
 	.result 	i32
 	i32.load8_s	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	sext_i8_i32, .Lfunc_end0-sext_i8_i32
 
@@ -17,6 +18,7 @@ zext_i8_i32:
 	.result 	i32
 	i32.load8_u	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	zext_i8_i32, .Lfunc_end1-zext_i8_i32
 
@@ -27,6 +29,7 @@ sext_i16_i32:
 	.result 	i32
 	i32.load16_s	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	sext_i16_i32, .Lfunc_end2-sext_i16_i32
 
@@ -37,6 +40,7 @@ zext_i16_i32:
 	.result 	i32
 	i32.load16_u	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	zext_i16_i32, .Lfunc_end3-zext_i16_i32
 
@@ -47,6 +51,7 @@ sext_i8_i64:
 	.result 	i64
 	i64.load8_s	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	sext_i8_i64, .Lfunc_end4-sext_i8_i64
 
@@ -57,6 +62,7 @@ zext_i8_i64:
 	.result 	i64
 	i64.load8_u	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	zext_i8_i64, .Lfunc_end5-zext_i8_i64
 
@@ -67,6 +73,7 @@ sext_i16_i64:
 	.result 	i64
 	i64.load16_s	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	sext_i16_i64, .Lfunc_end6-sext_i16_i64
 
@@ -77,6 +84,7 @@ zext_i16_i64:
 	.result 	i64
 	i64.load16_u	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	zext_i16_i64, .Lfunc_end7-zext_i16_i64
 
@@ -87,6 +95,7 @@ sext_i32_i64:
 	.result 	i64
 	i64.load32_s	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	sext_i32_i64, .Lfunc_end8-sext_i32_i64
 
@@ -97,8 +106,8 @@ zext_i32_i64:
 	.result 	i64
 	i64.load32_u	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end9:
 	.size	zext_i32_i64, .Lfunc_end9-zext_i32_i64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/load-store-i1.s b/test/llvm_autogenerated/load-store-i1.s
index 0deb7b440..f5557b6b9 100644
--- a/test/llvm_autogenerated/load-store-i1.s
+++ b/test/llvm_autogenerated/load-store-i1.s
@@ -7,6 +7,7 @@ load_u_i1_i32:
 	.result 	i32
 	i32.load8_u	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	load_u_i1_i32, .Lfunc_end0-load_u_i1_i32
 
@@ -15,12 +16,13 @@ load_u_i1_i32:
 load_s_i1_i32:
 	.param  	i32
 	.result 	i32
-	.local  	i32
-	i32.const	$1=, 31
 	i32.load8_u	$push0=, 0($0)
-	i32.shl 	$push1=, $pop0, $1
-	i32.shr_s	$push2=, $pop1, $1
-	return  	$pop2
+	i32.const	$push1=, 31
+	i32.shl 	$push2=, $pop0, $pop1
+	i32.const	$push4=, 31
+	i32.shr_s	$push3=, $pop2, $pop4
+	return  	$pop3
+	.endfunc
 .Lfunc_end1:
 	.size	load_s_i1_i32, .Lfunc_end1-load_s_i1_i32
 
@@ -31,6 +33,7 @@ load_u_i1_i64:
 	.result 	i64
 	i64.load8_u	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	load_u_i1_i64, .Lfunc_end2-load_u_i1_i64
 
@@ -39,12 +42,13 @@ load_u_i1_i64:
 load_s_i1_i64:
 	.param  	i32
 	.result 	i64
-	.local  	i64
-	i64.const	$1=, 63
 	i64.load8_u	$push0=, 0($0)
-	i64.shl 	$push1=, $pop0, $1
-	i64.shr_s	$push2=, $pop1, $1
-	return  	$pop2
+	i64.const	$push1=, 63
+	i64.shl 	$push2=, $pop0, $pop1
+	i64.const	$push4=, 63
+	i64.shr_s	$push3=, $pop2, $pop4
+	return  	$pop3
+	.endfunc
 .Lfunc_end3:
 	.size	load_s_i1_i64, .Lfunc_end3-load_s_i1_i64
 
@@ -56,6 +60,7 @@ store_i32_i1:
 	i32.and 	$push1=, $1, $pop0
 	i32.store8	$discard=, 0($0), $pop1
 	return
+	.endfunc
 .Lfunc_end4:
 	.size	store_i32_i1, .Lfunc_end4-store_i32_i1
 
@@ -67,8 +72,8 @@ store_i64_i1:
 	i64.and 	$push1=, $1, $pop0
 	i64.store8	$discard=, 0($0), $pop1
 	return
+	.endfunc
 .Lfunc_end5:
 	.size	store_i64_i1, .Lfunc_end5-store_i64_i1
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/load-store-i1.wast b/test/llvm_autogenerated/load-store-i1.wast
index 863384950..63156af87 100644
--- a/test/llvm_autogenerated/load-store-i1.wast
+++ b/test/llvm_autogenerated/load-store-i1.wast
@@ -18,21 +18,17 @@
     )
   )
   (func $load_s_i1_i32 (param $$0 i32) (result i32)
-    (local $$1 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$1
-          (i32.const 31)
-        )
         (br $fake_return_waka123
           (i32.shr_s
             (i32.shl
               (i32.load8_u align=1
                 (get_local $$0)
               )
-              (get_local $$1)
+              (i32.const 31)
             )
-            (get_local $$1)
+            (i32.const 31)
           )
         )
       )
@@ -50,21 +46,17 @@
     )
   )
   (func $load_s_i1_i64 (param $$0 i32) (result i64)
-    (local $$1 i64)
     (block $fake_return_waka123
       (block
-        (set_local $$1
-          (i64.const 63)
-        )
         (br $fake_return_waka123
           (i64.shr_s
             (i64.shl
               (i64.load8_u align=1
                 (get_local $$0)
               )
-              (get_local $$1)
+              (i64.const 63)
             )
-            (get_local $$1)
+            (i64.const 63)
           )
         )
       )
diff --git a/test/llvm_autogenerated/load.s b/test/llvm_autogenerated/load.s
index 98a078c05..f00f83985 100644
--- a/test/llvm_autogenerated/load.s
+++ b/test/llvm_autogenerated/load.s
@@ -7,6 +7,7 @@ ldi32:
 	.result 	i32
 	i32.load	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	ldi32, .Lfunc_end0-ldi32
 
@@ -17,6 +18,7 @@ ldi64:
 	.result 	i64
 	i64.load	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	ldi64, .Lfunc_end1-ldi64
 
@@ -27,6 +29,7 @@ ldf32:
 	.result 	f32
 	f32.load	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	ldf32, .Lfunc_end2-ldf32
 
@@ -37,8 +40,8 @@ ldf64:
 	.result 	f64
 	f64.load	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	ldf64, .Lfunc_end3-ldf64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/memory-addr32.s b/test/llvm_autogenerated/memory-addr32.s
index 7cc7655bc..f693e4769 100644
--- a/test/llvm_autogenerated/memory-addr32.s
+++ b/test/llvm_autogenerated/memory-addr32.s
@@ -6,6 +6,7 @@ memory_size:
 	.result 	i32
 	memory_size	$push0=
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	memory_size, .Lfunc_end0-memory_size
 
@@ -15,8 +16,8 @@ grow_memory:
 	.param  	i32
 	grow_memory	$0
 	return
+	.endfunc
 .Lfunc_end1:
 	.size	grow_memory, .Lfunc_end1-grow_memory
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/memory-addr64.s b/test/llvm_autogenerated/memory-addr64.s
index 8f42acc5e..7a8ca3e8e 100644
--- a/test/llvm_autogenerated/memory-addr64.s
+++ b/test/llvm_autogenerated/memory-addr64.s
@@ -6,6 +6,7 @@ memory_size:
 	.result 	i64
 	memory_size	$push0=
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	memory_size, .Lfunc_end0-memory_size
 
@@ -15,8 +16,8 @@ grow_memory:
 	.param  	i64
 	grow_memory	$0
 	return
+	.endfunc
 .Lfunc_end1:
 	.size	grow_memory, .Lfunc_end1-grow_memory
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/non-executable-stack.s b/test/llvm_autogenerated/non-executable-stack.s
new file mode 100644
index 000000000..a331dc616
--- /dev/null
+++ b/test/llvm_autogenerated/non-executable-stack.s
@@ -0,0 +1,3 @@
+	.text
+	.file	"/s/llvm/llvm/test/CodeGen/WebAssembly/non-executable-stack.ll"
+
diff --git a/test/llvm_autogenerated/non-executable-stack.wast b/test/llvm_autogenerated/non-executable-stack.wast
new file mode 100644
index 000000000..62e304e84
--- /dev/null
+++ b/test/llvm_autogenerated/non-executable-stack.wast
@@ -0,0 +1,4 @@
+(module
+  (memory 0 4294967295)
+)
+;; METADATA: { "asmConsts": {},"staticBump": 4 }
diff --git a/test/llvm_autogenerated/offset.s b/test/llvm_autogenerated/offset.s
index e1252c121..46f3173de 100644
--- a/test/llvm_autogenerated/offset.s
+++ b/test/llvm_autogenerated/offset.s
@@ -7,6 +7,7 @@ load_i32_with_folded_offset:
 	.result 	i32
 	i32.load	$push0=, 24($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	load_i32_with_folded_offset, .Lfunc_end0-load_i32_with_folded_offset
 
@@ -17,6 +18,7 @@ load_i32_with_folded_gep_offset:
 	.result 	i32
 	i32.load	$push0=, 24($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	load_i32_with_folded_gep_offset, .Lfunc_end1-load_i32_with_folded_gep_offset
 
@@ -29,6 +31,7 @@ load_i32_with_unfolded_gep_negative_offset:
 	i32.add 	$push1=, $0, $pop0
 	i32.load	$push2=, 0($pop1)
 	return  	$pop2
+	.endfunc
 .Lfunc_end2:
 	.size	load_i32_with_unfolded_gep_negative_offset, .Lfunc_end2-load_i32_with_unfolded_gep_negative_offset
 
@@ -41,6 +44,7 @@ load_i32_with_unfolded_offset:
 	i32.add 	$push1=, $0, $pop0
 	i32.load	$push2=, 0($pop1)
 	return  	$pop2
+	.endfunc
 .Lfunc_end3:
 	.size	load_i32_with_unfolded_offset, .Lfunc_end3-load_i32_with_unfolded_offset
 
@@ -53,6 +57,7 @@ load_i32_with_unfolded_gep_offset:
 	i32.add 	$push1=, $0, $pop0
 	i32.load	$push2=, 0($pop1)
 	return  	$pop2
+	.endfunc
 .Lfunc_end4:
 	.size	load_i32_with_unfolded_gep_offset, .Lfunc_end4-load_i32_with_unfolded_gep_offset
 
@@ -63,6 +68,7 @@ load_i64_with_folded_offset:
 	.result 	i64
 	i64.load	$push0=, 24($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	load_i64_with_folded_offset, .Lfunc_end5-load_i64_with_folded_offset
 
@@ -73,6 +79,7 @@ load_i64_with_folded_gep_offset:
 	.result 	i64
 	i64.load	$push0=, 24($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	load_i64_with_folded_gep_offset, .Lfunc_end6-load_i64_with_folded_gep_offset
 
@@ -85,6 +92,7 @@ load_i64_with_unfolded_gep_negative_offset:
 	i32.add 	$push1=, $0, $pop0
 	i64.load	$push2=, 0($pop1)
 	return  	$pop2
+	.endfunc
 .Lfunc_end7:
 	.size	load_i64_with_unfolded_gep_negative_offset, .Lfunc_end7-load_i64_with_unfolded_gep_negative_offset
 
@@ -97,6 +105,7 @@ load_i64_with_unfolded_offset:
 	i32.add 	$push1=, $0, $pop0
 	i64.load	$push2=, 0($pop1)
 	return  	$pop2
+	.endfunc
 .Lfunc_end8:
 	.size	load_i64_with_unfolded_offset, .Lfunc_end8-load_i64_with_unfolded_offset
 
@@ -109,6 +118,7 @@ load_i64_with_unfolded_gep_offset:
 	i32.add 	$push1=, $0, $pop0
 	i64.load	$push2=, 0($pop1)
 	return  	$pop2
+	.endfunc
 .Lfunc_end9:
 	.size	load_i64_with_unfolded_gep_offset, .Lfunc_end9-load_i64_with_unfolded_gep_offset
 
@@ -119,6 +129,7 @@ store_i32_with_folded_offset:
 	i32.const	$push0=, 0
 	i32.store	$discard=, 24($0), $pop0
 	return
+	.endfunc
 .Lfunc_end10:
 	.size	store_i32_with_folded_offset, .Lfunc_end10-store_i32_with_folded_offset
 
@@ -129,6 +140,7 @@ store_i32_with_folded_gep_offset:
 	i32.const	$push0=, 0
 	i32.store	$discard=, 24($0), $pop0
 	return
+	.endfunc
 .Lfunc_end11:
 	.size	store_i32_with_folded_gep_offset, .Lfunc_end11-store_i32_with_folded_gep_offset
 
@@ -141,6 +153,7 @@ store_i32_with_unfolded_gep_negative_offset:
 	i32.const	$push2=, 0
 	i32.store	$discard=, 0($pop1), $pop2
 	return
+	.endfunc
 .Lfunc_end12:
 	.size	store_i32_with_unfolded_gep_negative_offset, .Lfunc_end12-store_i32_with_unfolded_gep_negative_offset
 
@@ -153,6 +166,7 @@ store_i32_with_unfolded_offset:
 	i32.const	$push2=, 0
 	i32.store	$discard=, 0($pop1), $pop2
 	return
+	.endfunc
 .Lfunc_end13:
 	.size	store_i32_with_unfolded_offset, .Lfunc_end13-store_i32_with_unfolded_offset
 
@@ -165,6 +179,7 @@ store_i32_with_unfolded_gep_offset:
 	i32.const	$push2=, 0
 	i32.store	$discard=, 0($pop1), $pop2
 	return
+	.endfunc
 .Lfunc_end14:
 	.size	store_i32_with_unfolded_gep_offset, .Lfunc_end14-store_i32_with_unfolded_gep_offset
 
@@ -175,6 +190,7 @@ store_i64_with_folded_offset:
 	i64.const	$push0=, 0
 	i64.store	$discard=, 24($0), $pop0
 	return
+	.endfunc
 .Lfunc_end15:
 	.size	store_i64_with_folded_offset, .Lfunc_end15-store_i64_with_folded_offset
 
@@ -185,6 +201,7 @@ store_i64_with_folded_gep_offset:
 	i64.const	$push0=, 0
 	i64.store	$discard=, 24($0), $pop0
 	return
+	.endfunc
 .Lfunc_end16:
 	.size	store_i64_with_folded_gep_offset, .Lfunc_end16-store_i64_with_folded_gep_offset
 
@@ -197,6 +214,7 @@ store_i64_with_unfolded_gep_negative_offset:
 	i64.const	$push2=, 0
 	i64.store	$discard=, 0($pop1), $pop2
 	return
+	.endfunc
 .Lfunc_end17:
 	.size	store_i64_with_unfolded_gep_negative_offset, .Lfunc_end17-store_i64_with_unfolded_gep_negative_offset
 
@@ -209,6 +227,7 @@ store_i64_with_unfolded_offset:
 	i64.const	$push2=, 0
 	i64.store	$discard=, 0($pop1), $pop2
 	return
+	.endfunc
 .Lfunc_end18:
 	.size	store_i64_with_unfolded_offset, .Lfunc_end18-store_i64_with_unfolded_offset
 
@@ -221,6 +240,7 @@ store_i64_with_unfolded_gep_offset:
 	i64.const	$push2=, 0
 	i64.store	$discard=, 0($pop1), $pop2
 	return
+	.endfunc
 .Lfunc_end19:
 	.size	store_i64_with_unfolded_gep_offset, .Lfunc_end19-store_i64_with_unfolded_gep_offset
 
@@ -231,6 +251,7 @@ load_i32_from_numeric_address:
 	i32.const	$push0=, 0
 	i32.load	$push1=, 42($pop0)
 	return  	$pop1
+	.endfunc
 .Lfunc_end20:
 	.size	load_i32_from_numeric_address, .Lfunc_end20-load_i32_from_numeric_address
 
@@ -241,26 +262,29 @@ load_i32_from_global_address:
 	i32.const	$push0=, 0
 	i32.load	$push1=, gv($pop0)
 	return  	$pop1
+	.endfunc
 .Lfunc_end21:
 	.size	load_i32_from_global_address, .Lfunc_end21-load_i32_from_global_address
 
 	.globl	store_i32_to_numeric_address
 	.type	store_i32_to_numeric_address,@function
 store_i32_to_numeric_address:
-	.local  	i32
-	i32.const	$0=, 0
-	i32.store	$discard=, 42($0), $0
+	i32.const	$push0=, 0
+	i32.const	$push1=, 0
+	i32.store	$discard=, 42($pop0), $pop1
 	return
+	.endfunc
 .Lfunc_end22:
 	.size	store_i32_to_numeric_address, .Lfunc_end22-store_i32_to_numeric_address
 
 	.globl	store_i32_to_global_address
 	.type	store_i32_to_global_address,@function
 store_i32_to_global_address:
-	.local  	i32
-	i32.const	$0=, 0
-	i32.store	$discard=, gv($0), $0
+	i32.const	$push0=, 0
+	i32.const	$push1=, 0
+	i32.store	$discard=, gv($pop0), $pop1
 	return
+	.endfunc
 .Lfunc_end23:
 	.size	store_i32_to_global_address, .Lfunc_end23-store_i32_to_global_address
 
@@ -271,6 +295,7 @@ load_i8_s_with_folded_offset:
 	.result 	i32
 	i32.load8_s	$push0=, 24($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end24:
 	.size	load_i8_s_with_folded_offset, .Lfunc_end24-load_i8_s_with_folded_offset
 
@@ -281,6 +306,7 @@ load_i8_s_with_folded_gep_offset:
 	.result 	i32
 	i32.load8_s	$push0=, 24($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end25:
 	.size	load_i8_s_with_folded_gep_offset, .Lfunc_end25-load_i8_s_with_folded_gep_offset
 
@@ -291,6 +317,7 @@ load_i8_u_with_folded_offset:
 	.result 	i32
 	i32.load8_u	$push0=, 24($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end26:
 	.size	load_i8_u_with_folded_offset, .Lfunc_end26-load_i8_u_with_folded_offset
 
@@ -301,6 +328,7 @@ load_i8_u_with_folded_gep_offset:
 	.result 	i32
 	i32.load8_u	$push0=, 24($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end27:
 	.size	load_i8_u_with_folded_gep_offset, .Lfunc_end27-load_i8_u_with_folded_gep_offset
 
@@ -311,6 +339,7 @@ store_i8_with_folded_offset:
 	i32.const	$push0=, 0
 	i32.store8	$discard=, 24($0), $pop0
 	return
+	.endfunc
 .Lfunc_end28:
 	.size	store_i8_with_folded_offset, .Lfunc_end28-store_i8_with_folded_offset
 
@@ -321,6 +350,7 @@ store_i8_with_folded_gep_offset:
 	i32.const	$push0=, 0
 	i32.store8	$discard=, 24($0), $pop0
 	return
+	.endfunc
 .Lfunc_end29:
 	.size	store_i8_with_folded_gep_offset, .Lfunc_end29-store_i8_with_folded_gep_offset
 
@@ -338,6 +368,7 @@ aggregate_load_store:
 	i32.store	$discard=, 4($1), $3
 	i32.store	$discard=, 0($1), $2
 	return
+	.endfunc
 .Lfunc_end30:
 	.size	aggregate_load_store, .Lfunc_end30-aggregate_load_store
 
@@ -345,15 +376,29 @@ aggregate_load_store:
 	.type	aggregate_return,@function
 aggregate_return:
 	.param  	i32
-	i32.const	$push0=, 0
-	i32.store	$push1=, 12($0), $pop0
-	i32.store	$push2=, 8($0), $pop1
-	i32.store	$push3=, 4($0), $pop2
-	i32.store	$discard=, 0($0), $pop3
+	i64.const	$push0=, 0
+	i64.store	$push1=, 8($0):p2align=2, $pop0
+	i64.store	$discard=, 0($0):p2align=2, $pop1
 	return
+	.endfunc
 .Lfunc_end31:
 	.size	aggregate_return, .Lfunc_end31-aggregate_return
 
+	.globl	aggregate_return_without_merge
+	.type	aggregate_return_without_merge,@function
+aggregate_return_without_merge:
+	.param  	i32
+	i32.const	$push0=, 0
+	i32.store8	$push1=, 14($0), $pop0
+	i32.store16	$push2=, 12($0), $pop1
+	i32.store	$discard=, 8($0), $pop2
+	i64.const	$push3=, 0
+	i64.store	$discard=, 0($0), $pop3
+	return
+	.endfunc
+.Lfunc_end32:
+	.size	aggregate_return_without_merge, .Lfunc_end32-aggregate_return_without_merge
+
 	.type	gv,@object
 	.bss
 	.globl	gv
@@ -363,4 +408,3 @@ gv:
 	.size	gv, 4
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/offset.wast b/test/llvm_autogenerated/offset.wast
index 5dea41ca3..759830ffd 100644
--- a/test/llvm_autogenerated/offset.wast
+++ b/test/llvm_autogenerated/offset.wast
@@ -32,6 +32,7 @@
   (export "store_i8_with_folded_gep_offset" $store_i8_with_folded_gep_offset)
   (export "aggregate_load_store" $aggregate_load_store)
   (export "aggregate_return" $aggregate_return)
+  (export "aggregate_return_without_merge" $aggregate_return_without_merge)
   (func $load_i32_with_folded_offset (param $$0 i32) (result i32)
     (block $fake_return_waka123
       (block
@@ -311,30 +312,22 @@
     )
   )
   (func $store_i32_to_numeric_address
-    (local $$0 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$0
-          (i32.const 0)
-        )
         (i32.store offset=42 align=4
-          (get_local $$0)
-          (get_local $$0)
+          (i32.const 0)
+          (i32.const 0)
         )
         (br $fake_return_waka123)
       )
     )
   )
   (func $store_i32_to_global_address
-    (local $$0 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$0
-          (i32.const 0)
-        )
         (i32.store offset=8 align=4
-          (get_local $$0)
-          (get_local $$0)
+          (i32.const 0)
+          (i32.const 0)
         )
         (br $fake_return_waka123)
       )
@@ -452,19 +445,34 @@
   (func $aggregate_return (param $$0 i32)
     (block $fake_return_waka123
       (block
-        (i32.store align=4
+        (i64.store align=4
+          (get_local $$0)
+          (i64.store offset=8 align=4
+            (get_local $$0)
+            (i64.const 0)
+          )
+        )
+        (br $fake_return_waka123)
+      )
+    )
+  )
+  (func $aggregate_return_without_merge (param $$0 i32)
+    (block $fake_return_waka123
+      (block
+        (i32.store offset=8 align=4
           (get_local $$0)
-          (i32.store offset=4 align=4
+          (i32.store offset=12 align=16
             (get_local $$0)
-            (i32.store offset=8 align=4
+            (i32.store offset=14 align=8
               (get_local $$0)
-              (i32.store offset=12 align=4
-                (get_local $$0)
-                (i32.const 0)
-              )
+              (i32.const 0)
             )
           )
         )
+        (i64.store align=8
+          (get_local $$0)
+          (i64.const 0)
+        )
         (br $fake_return_waka123)
       )
     )
diff --git a/test/llvm_autogenerated/phi.s b/test/llvm_autogenerated/phi.s
index b29451aaf..4a3a6d522 100644
--- a/test/llvm_autogenerated/phi.s
+++ b/test/llvm_autogenerated/phi.s
@@ -14,6 +14,7 @@ test0:
 .LBB0_2:
 	end_block
 	return  	$0
+	.endfunc
 .Lfunc_end0:
 	.size	test0, .Lfunc_end0-test0
 
@@ -22,23 +23,23 @@ test0:
 test1:
 	.param  	i32
 	.result 	i32
-	.local  	i32, i32, i32, i32, i32
-	i32.const	$2=, 1
-	i32.const	$3=, 0
-	copy_local	$4=, $2
-	copy_local	$5=, $3
+	.local  	i32, i32, i32, i32
+	i32.const	$2=, 0
+	i32.const	$3=, 1
+	i32.const	$4=, 0
 .LBB1_1:
 	loop
-	copy_local	$1=, $4
-	copy_local	$4=, $3
-	i32.add 	$5=, $5, $2
-	copy_local	$3=, $1
-	i32.lt_s	$push0=, $5, $0
+	copy_local	$1=, $3
+	copy_local	$3=, $2
+	i32.const	$push1=, 1
+	i32.add 	$4=, $4, $pop1
+	copy_local	$2=, $1
+	i32.lt_s	$push0=, $4, $0
 	br_if   	$pop0, 0
 	end_loop
-	return  	$4
+	return  	$3
+	.endfunc
 .Lfunc_end1:
 	.size	test1, .Lfunc_end1-test1
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/phi.wast b/test/llvm_autogenerated/phi.wast
index 65bd501d8..c1d4d5c2c 100644
--- a/test/llvm_autogenerated/phi.wast
+++ b/test/llvm_autogenerated/phi.wast
@@ -31,47 +31,43 @@
     (local $$2 i32)
     (local $$3 i32)
     (local $$4 i32)
-    (local $$5 i32)
     (block $fake_return_waka123
       (block
         (set_local $$2
-          (i32.const 1)
+          (i32.const 0)
         )
         (set_local $$3
-          (i32.const 0)
+          (i32.const 1)
         )
         (set_local $$4
-          (get_local $$2)
-        )
-        (set_local $$5
-          (get_local $$3)
+          (i32.const 0)
         )
         (loop $label$1 $label$0
           (set_local $$1
-            (get_local $$4)
-          )
-          (set_local $$4
             (get_local $$3)
           )
-          (set_local $$5
+          (set_local $$3
+            (get_local $$2)
+          )
+          (set_local $$4
             (i32.add
-              (get_local $$5)
-              (get_local $$2)
+              (get_local $$4)
+              (i32.const 1)
             )
           )
-          (set_local $$3
+          (set_local $$2
             (get_local $$1)
           )
           (br_if
             (i32.lt_s
-              (get_local $$5)
+              (get_local $$4)
               (get_local $$0)
             )
             $label$0
           )
         )
         (br $fake_return_waka123
-          (get_local $$4)
+          (get_local $$3)
         )
       )
     )
diff --git a/test/llvm_autogenerated/reg-stackify.s b/test/llvm_autogenerated/reg-stackify.s
index 1fffec602..38ebc93f4 100644
--- a/test/llvm_autogenerated/reg-stackify.s
+++ b/test/llvm_autogenerated/reg-stackify.s
@@ -9,6 +9,7 @@ no0:
 	i32.const	$push0=, 0
 	i32.store	$discard=, 0($0), $pop0
 	return  	$1
+	.endfunc
 .Lfunc_end0:
 	.size	no0, .Lfunc_end0-no0
 
@@ -21,6 +22,7 @@ no1:
 	i32.const	$push0=, 0
 	i32.store	$discard=, 0($0), $pop0
 	return  	$1
+	.endfunc
 .Lfunc_end1:
 	.size	no1, .Lfunc_end1-no1
 
@@ -33,6 +35,7 @@ yes0:
 	i32.store	$discard=, 0($0), $pop1
 	i32.load	$push0=, 0($1)
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	yes0, .Lfunc_end2-yes0
 
@@ -43,6 +46,7 @@ yes1:
 	.result 	i32
 	i32.load	$push0=, 0($0)
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	yes1, .Lfunc_end3-yes1
 
@@ -51,24 +55,28 @@ yes1:
 stack_uses:
 	.param  	i32, i32, i32, i32
 	.result 	i32
-	.local  	i32, i32
-	i32.const	$5=, 2
-	i32.const	$4=, 1
 	block
-	i32.lt_s	$push0=, $0, $4
-	i32.lt_s	$push1=, $1, $5
-	i32.xor 	$push4=, $pop0, $pop1
-	i32.lt_s	$push2=, $2, $4
-	i32.lt_s	$push3=, $3, $5
-	i32.xor 	$push5=, $pop2, $pop3
-	i32.xor 	$push6=, $pop4, $pop5
-	i32.ne  	$push7=, $pop6, $4
-	br_if   	$pop7, 0
-	i32.const	$push8=, 0
-	return  	$pop8
+	i32.const	$push13=, 1
+	i32.lt_s	$push0=, $0, $pop13
+	i32.const	$push1=, 2
+	i32.lt_s	$push2=, $1, $pop1
+	i32.xor 	$push5=, $pop0, $pop2
+	i32.const	$push12=, 1
+	i32.lt_s	$push3=, $2, $pop12
+	i32.const	$push11=, 2
+	i32.lt_s	$push4=, $3, $pop11
+	i32.xor 	$push6=, $pop3, $pop4
+	i32.xor 	$push7=, $pop5, $pop6
+	i32.const	$push10=, 1
+	i32.ne  	$push8=, $pop7, $pop10
+	br_if   	$pop8, 0
+	i32.const	$push9=, 0
+	return  	$pop9
 .LBB4_2:
 	end_block
-	return  	$4
+	i32.const	$push14=, 1
+	return  	$pop14
+	.endfunc
 .Lfunc_end4:
 	.size	stack_uses, .Lfunc_end4-stack_uses
 
@@ -87,8 +95,24 @@ multiple_uses:
 .LBB5_3:
 	end_block
 	return
+	.endfunc
 .Lfunc_end5:
 	.size	multiple_uses, .Lfunc_end5-multiple_uses
 
+	.hidden	stackify_store_across_side_effects
+	.globl	stackify_store_across_side_effects
+	.type	stackify_store_across_side_effects,@function
+stackify_store_across_side_effects:
+	.param  	i32
+	.local  	i64
+	i64.const	$push0=, 4611686018427387904
+	i64.store	$1=, 0($0), $pop0
+	call    	evoke_side_effects@FUNCTION
+	i64.store	$discard=, 0($0), $1
+	call    	evoke_side_effects@FUNCTION
+	return
+	.endfunc
+.Lfunc_end6:
+	.size	stackify_store_across_side_effects, .Lfunc_end6-stackify_store_across_side_effects
+
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/reg-stackify.wast b/test/llvm_autogenerated/reg-stackify.wast
index 2952ef1be..3a367ac52 100644
--- a/test/llvm_autogenerated/reg-stackify.wast
+++ b/test/llvm_autogenerated/reg-stackify.wast
@@ -1,11 +1,14 @@
 (module
   (memory 0 4294967295)
+  (type $FUNCSIG$v (func))
+  (import $evoke_side_effects "env" "evoke_side_effects")
   (export "no0" $no0)
   (export "no1" $no1)
   (export "yes0" $yes0)
   (export "yes1" $yes1)
   (export "stack_uses" $stack_uses)
   (export "multiple_uses" $multiple_uses)
+  (export "stackify_store_across_side_effects" $stackify_store_across_side_effects)
   (func $no0 (param $$0 i32) (param $$1 i32) (result i32)
     (block $fake_return_waka123
       (block
@@ -69,16 +72,8 @@
     )
   )
   (func $stack_uses (param $$0 i32) (param $$1 i32) (param $$2 i32) (param $$3 i32) (result i32)
-    (local $$4 i32)
-    (local $$5 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$5
-          (i32.const 2)
-        )
-        (set_local $$4
-          (i32.const 1)
-        )
         (block $label$0
           (br_if
             (i32.ne
@@ -86,25 +81,25 @@
                 (i32.xor
                   (i32.lt_s
                     (get_local $$0)
-                    (get_local $$4)
+                    (i32.const 1)
                   )
                   (i32.lt_s
                     (get_local $$1)
-                    (get_local $$5)
+                    (i32.const 2)
                   )
                 )
                 (i32.xor
                   (i32.lt_s
                     (get_local $$2)
-                    (get_local $$4)
+                    (i32.const 1)
                   )
                   (i32.lt_s
                     (get_local $$3)
-                    (get_local $$5)
+                    (i32.const 2)
                   )
                 )
               )
-              (get_local $$4)
+              (i32.const 1)
             )
             $label$0
           )
@@ -113,7 +108,7 @@
           )
         )
         (br $fake_return_waka123
-          (get_local $$4)
+          (i32.const 1)
         )
       )
     )
@@ -151,5 +146,25 @@
       )
     )
   )
+  (func $stackify_store_across_side_effects (param $$0 i32)
+    (local $$1 i64)
+    (block $fake_return_waka123
+      (block
+        (set_local $$1
+          (i64.store align=8
+            (get_local $$0)
+            (i64.const 4611686018427387904)
+          )
+        )
+        (call_import $evoke_side_effects)
+        (i64.store align=8
+          (get_local $$0)
+          (get_local $$1)
+        )
+        (call_import $evoke_side_effects)
+        (br $fake_return_waka123)
+      )
+    )
+  )
 )
 ;; METADATA: { "asmConsts": {},"staticBump": 4 }
diff --git a/test/llvm_autogenerated/return-int32.s b/test/llvm_autogenerated/return-int32.s
index ebe7a1af2..1083e3a1d 100644
--- a/test/llvm_autogenerated/return-int32.s
+++ b/test/llvm_autogenerated/return-int32.s
@@ -6,8 +6,8 @@ return_i32:
 	.param  	i32
 	.result 	i32
 	return  	$0
+	.endfunc
 .Lfunc_end0:
 	.size	return_i32, .Lfunc_end0-return_i32
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/return-void.s b/test/llvm_autogenerated/return-void.s
index 638de33fb..2293d11fc 100644
--- a/test/llvm_autogenerated/return-void.s
+++ b/test/llvm_autogenerated/return-void.s
@@ -4,8 +4,8 @@
 	.type	return_void,@function
 return_void:
 	return
+	.endfunc
 .Lfunc_end0:
 	.size	return_void, .Lfunc_end0-return_void
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/select.s b/test/llvm_autogenerated/select.s
index fc5c7bac3..c353bd99d 100644
--- a/test/llvm_autogenerated/select.s
+++ b/test/llvm_autogenerated/select.s
@@ -7,6 +7,7 @@ select_i32_bool:
 	.result 	i32
 	i32.select	$push0=, $0, $1, $2
 	return  	$pop0
+	.endfunc
 .Lfunc_end0:
 	.size	select_i32_bool, .Lfunc_end0-select_i32_bool
 
@@ -17,6 +18,7 @@ select_i32_eq:
 	.result 	i32
 	i32.select	$push0=, $0, $2, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end1:
 	.size	select_i32_eq, .Lfunc_end1-select_i32_eq
 
@@ -27,6 +29,7 @@ select_i32_ne:
 	.result 	i32
 	i32.select	$push0=, $0, $1, $2
 	return  	$pop0
+	.endfunc
 .Lfunc_end2:
 	.size	select_i32_ne, .Lfunc_end2-select_i32_ne
 
@@ -37,6 +40,7 @@ select_i64_bool:
 	.result 	i64
 	i64.select	$push0=, $0, $1, $2
 	return  	$pop0
+	.endfunc
 .Lfunc_end3:
 	.size	select_i64_bool, .Lfunc_end3-select_i64_bool
 
@@ -47,6 +51,7 @@ select_i64_eq:
 	.result 	i64
 	i64.select	$push0=, $0, $2, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end4:
 	.size	select_i64_eq, .Lfunc_end4-select_i64_eq
 
@@ -57,6 +62,7 @@ select_i64_ne:
 	.result 	i64
 	i64.select	$push0=, $0, $1, $2
 	return  	$pop0
+	.endfunc
 .Lfunc_end5:
 	.size	select_i64_ne, .Lfunc_end5-select_i64_ne
 
@@ -67,6 +73,7 @@ select_f32_bool:
 	.result 	f32
 	f32.select	$push0=, $0, $1, $2
 	return  	$pop0
+	.endfunc
 .Lfunc_end6:
 	.size	select_f32_bool, .Lfunc_end6-select_f32_bool
 
@@ -77,6 +84,7 @@ select_f32_eq:
 	.result 	f32
 	f32.select	$push0=, $0, $2, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end7:
 	.size	select_f32_eq, .Lfunc_end7-select_f32_eq
 
@@ -87,6 +95,7 @@ select_f32_ne:
 	.result 	f32
 	f32.select	$push0=, $0, $1, $2
 	return  	$pop0
+	.endfunc
 .Lfunc_end8:
 	.size	select_f32_ne, .Lfunc_end8-select_f32_ne
 
@@ -97,6 +106,7 @@ select_f64_bool:
 	.result 	f64
 	f64.select	$push0=, $0, $1, $2
 	return  	$pop0
+	.endfunc
 .Lfunc_end9:
 	.size	select_f64_bool, .Lfunc_end9-select_f64_bool
 
@@ -107,6 +117,7 @@ select_f64_eq:
 	.result 	f64
 	f64.select	$push0=, $0, $2, $1
 	return  	$pop0
+	.endfunc
 .Lfunc_end10:
 	.size	select_f64_eq, .Lfunc_end10-select_f64_eq
 
@@ -117,8 +128,8 @@ select_f64_ne:
 	.result 	f64
 	f64.select	$push0=, $0, $1, $2
 	return  	$pop0
+	.endfunc
 .Lfunc_end11:
 	.size	select_f64_ne, .Lfunc_end11-select_f64_ne
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/signext-zeroext.s b/test/llvm_autogenerated/signext-zeroext.s
index 395166900..6a35e7461 100644
--- a/test/llvm_autogenerated/signext-zeroext.s
+++ b/test/llvm_autogenerated/signext-zeroext.s
@@ -5,11 +5,12 @@
 z2s_func:
 	.param  	i32
 	.result 	i32
-	.local  	i32
-	i32.const	$1=, 24
-	i32.shl 	$push0=, $0, $1
-	i32.shr_s	$push1=, $pop0, $1
-	return  	$pop1
+	i32.const	$push0=, 24
+	i32.shl 	$push1=, $0, $pop0
+	i32.const	$push3=, 24
+	i32.shr_s	$push2=, $pop1, $pop3
+	return  	$pop2
+	.endfunc
 .Lfunc_end0:
 	.size	z2s_func, .Lfunc_end0-z2s_func
 
@@ -21,6 +22,7 @@ s2z_func:
 	i32.const	$push0=, 255
 	i32.and 	$push1=, $0, $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end1:
 	.size	s2z_func, .Lfunc_end1-s2z_func
 
@@ -33,6 +35,7 @@ z2s_call:
 	i32.and 	$push1=, $0, $pop0
 	i32.call	$push2=, z2s_func@FUNCTION, $pop1
 	return  	$pop2
+	.endfunc
 .Lfunc_end2:
 	.size	z2s_call, .Lfunc_end2-z2s_call
 
@@ -41,16 +44,18 @@ z2s_call:
 s2z_call:
 	.param  	i32
 	.result 	i32
-	.local  	i32
-	i32.const	$1=, 24
-	i32.shl 	$push0=, $0, $1
-	i32.shr_s	$push1=, $pop0, $1
-	i32.call	$push2=, s2z_func@FUNCTION, $pop1
-	i32.shl 	$push3=, $pop2, $1
-	i32.shr_s	$push4=, $pop3, $1
-	return  	$pop4
+	i32.const	$push0=, 24
+	i32.shl 	$push1=, $0, $pop0
+	i32.const	$push8=, 24
+	i32.shr_s	$push2=, $pop1, $pop8
+	i32.call	$push3=, s2z_func@FUNCTION, $pop2
+	i32.const	$push7=, 24
+	i32.shl 	$push4=, $pop3, $pop7
+	i32.const	$push6=, 24
+	i32.shr_s	$push5=, $pop4, $pop6
+	return  	$pop5
+	.endfunc
 .Lfunc_end3:
 	.size	s2z_call, .Lfunc_end3-s2z_call
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/signext-zeroext.wast b/test/llvm_autogenerated/signext-zeroext.wast
index 2344c33be..31f6fe75c 100644
--- a/test/llvm_autogenerated/signext-zeroext.wast
+++ b/test/llvm_autogenerated/signext-zeroext.wast
@@ -5,19 +5,15 @@
   (export "z2s_call" $z2s_call)
   (export "s2z_call" $s2z_call)
   (func $z2s_func (param $$0 i32) (result i32)
-    (local $$1 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$1
-          (i32.const 24)
-        )
         (br $fake_return_waka123
           (i32.shr_s
             (i32.shl
               (get_local $$0)
-              (get_local $$1)
+              (i32.const 24)
             )
-            (get_local $$1)
+            (i32.const 24)
           )
         )
       )
@@ -50,12 +46,8 @@
     )
   )
   (func $s2z_call (param $$0 i32) (result i32)
-    (local $$1 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$1
-          (i32.const 24)
-        )
         (br $fake_return_waka123
           (i32.shr_s
             (i32.shl
@@ -63,14 +55,14 @@
                 (i32.shr_s
                   (i32.shl
                     (get_local $$0)
-                    (get_local $$1)
+                    (i32.const 24)
                   )
-                  (get_local $$1)
+                  (i32.const 24)
                 )
               )
-              (get_local $$1)
+              (i32.const 24)
             )
-            (get_local $$1)
+            (i32.const 24)
           )
         )
       )
diff --git a/test/llvm_autogenerated/store-results.s b/test/llvm_autogenerated/store-results.s
index f574cf7a7..dcb525654 100644
--- a/test/llvm_autogenerated/store-results.s
+++ b/test/llvm_autogenerated/store-results.s
@@ -8,44 +8,49 @@ single_block:
 	i32.const	$push0=, 0
 	i32.store	$push1=, 0($0), $pop0
 	return  	$pop1
+	.endfunc
 .Lfunc_end0:
 	.size	single_block, .Lfunc_end0-single_block
 
 	.globl	foo
 	.type	foo,@function
 foo:
-	.local  	i32, i32
+	.local  	i32
 	i32.const	$0=, 0
-	copy_local	$1=, $0
 .LBB1_1:
 	loop
-	i32.const	$push0=, 1
-	i32.add 	$1=, $1, $pop0
-	i32.store	$discard=, pos($0), $0
+	i32.const	$push4=, 0
+	i32.const	$push3=, 0
+	i32.store	$discard=, pos($pop4), $pop3
+	i32.const	$push2=, 1
+	i32.add 	$0=, $0, $pop2
 	i32.const	$push1=, 256
-	i32.ne  	$push2=, $1, $pop1
-	br_if   	$pop2, 0
+	i32.ne  	$push0=, $0, $pop1
+	br_if   	$pop0, 0
 	end_loop
 	return
+	.endfunc
 .Lfunc_end1:
 	.size	foo, .Lfunc_end1-foo
 
 	.globl	bar
 	.type	bar,@function
 bar:
-	.local  	i32, f32
-	f32.const	$1=, 0x0p0
-	i32.const	$0=, 0
+	.local  	f32
+	f32.const	$0=, 0x0p0
 .LBB2_1:
 	loop
-	i32.store	$discard=, pos($0), $0
-	f32.const	$push0=, 0x1p0
-	f32.add 	$1=, $1, $pop0
+	i32.const	$push4=, 0
+	i32.const	$push3=, 0
+	i32.store	$discard=, pos($pop4), $pop3
+	f32.const	$push2=, 0x1p0
+	f32.add 	$0=, $0, $pop2
 	f32.const	$push1=, 0x1p8
-	f32.ne  	$push2=, $1, $pop1
-	br_if   	$pop2, 0
+	f32.ne  	$push0=, $0, $pop1
+	br_if   	$pop0, 0
 	end_loop
 	return
+	.endfunc
 .Lfunc_end2:
 	.size	bar, .Lfunc_end2-bar
 
@@ -58,4 +63,3 @@ pos:
 	.size	pos, 12
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/store-results.wast b/test/llvm_autogenerated/store-results.wast
index 544a8c30a..17ca8effa 100644
--- a/test/llvm_autogenerated/store-results.wast
+++ b/test/llvm_autogenerated/store-results.wast
@@ -17,29 +17,25 @@
   )
   (func $foo
     (local $$0 i32)
-    (local $$1 i32)
     (block $fake_return_waka123
       (block
         (set_local $$0
           (i32.const 0)
         )
-        (set_local $$1
-          (get_local $$0)
-        )
         (loop $label$1 $label$0
-          (set_local $$1
+          (i32.store offset=8 align=4
+            (i32.const 0)
+            (i32.const 0)
+          )
+          (set_local $$0
             (i32.add
-              (get_local $$1)
+              (get_local $$0)
               (i32.const 1)
             )
           )
-          (i32.store offset=8 align=4
-            (get_local $$0)
-            (get_local $$0)
-          )
           (br_if
             (i32.ne
-              (get_local $$1)
+              (get_local $$0)
               (i32.const 256)
             )
             $label$0
@@ -50,30 +46,26 @@
     )
   )
   (func $bar
-    (local $$0 i32)
-    (local $$1 f32)
+    (local $$0 f32)
     (block $fake_return_waka123
       (block
-        (set_local $$1
-          (f32.const 0)
-        )
         (set_local $$0
-          (i32.const 0)
+          (f32.const 0)
         )
         (loop $label$1 $label$0
           (i32.store offset=8 align=4
-            (get_local $$0)
-            (get_local $$0)
+            (i32.const 0)
+            (i32.const 0)
           )
-          (set_local $$1
+          (set_local $$0
             (f32.add
-              (get_local $$1)
+              (get_local $$0)
               (f32.const 1)
             )
           )
           (br_if
             (f32.ne
-              (get_local $$1)
+              (get_local $$0)
               (f32.const 256)
             )
             $label$0
diff --git a/test/llvm_autogenerated/store-trunc.s b/test/llvm_autogenerated/store-trunc.s
index fd5e43b71..423e3ac6b 100644
--- a/test/llvm_autogenerated/store-trunc.s
+++ b/test/llvm_autogenerated/store-trunc.s
@@ -6,6 +6,7 @@ trunc_i8_i32:
 	.param  	i32, i32
 	i32.store8	$discard=, 0($0), $1
 	return
+	.endfunc
 .Lfunc_end0:
 	.size	trunc_i8_i32, .Lfunc_end0-trunc_i8_i32
 
@@ -15,6 +16,7 @@ trunc_i16_i32:
 	.param  	i32, i32
 	i32.store16	$discard=, 0($0), $1
 	return
+	.endfunc
 .Lfunc_end1:
 	.size	trunc_i16_i32, .Lfunc_end1-trunc_i16_i32
 
@@ -24,6 +26,7 @@ trunc_i8_i64:
 	.param  	i32, i64
 	i64.store8	$discard=, 0($0), $1
 	return
+	.endfunc
 .Lfunc_end2:
 	.size	trunc_i8_i64, .Lfunc_end2-trunc_i8_i64
 
@@ -33,6 +36,7 @@ trunc_i16_i64:
 	.param  	i32, i64
 	i64.store16	$discard=, 0($0), $1
 	return
+	.endfunc
 .Lfunc_end3:
 	.size	trunc_i16_i64, .Lfunc_end3-trunc_i16_i64
 
@@ -42,8 +46,8 @@ trunc_i32_i64:
 	.param  	i32, i64
 	i64.store32	$discard=, 0($0), $1
 	return
+	.endfunc
 .Lfunc_end4:
 	.size	trunc_i32_i64, .Lfunc_end4-trunc_i32_i64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/store.s b/test/llvm_autogenerated/store.s
index 9e66aea62..52371d905 100644
--- a/test/llvm_autogenerated/store.s
+++ b/test/llvm_autogenerated/store.s
@@ -6,6 +6,7 @@ sti32:
 	.param  	i32, i32
 	i32.store	$discard=, 0($0), $1
 	return
+	.endfunc
 .Lfunc_end0:
 	.size	sti32, .Lfunc_end0-sti32
 
@@ -15,6 +16,7 @@ sti64:
 	.param  	i32, i64
 	i64.store	$discard=, 0($0), $1
 	return
+	.endfunc
 .Lfunc_end1:
 	.size	sti64, .Lfunc_end1-sti64
 
@@ -24,6 +26,7 @@ stf32:
 	.param  	i32, f32
 	f32.store	$discard=, 0($0), $1
 	return
+	.endfunc
 .Lfunc_end2:
 	.size	stf32, .Lfunc_end2-stf32
 
@@ -33,8 +36,8 @@ stf64:
 	.param  	i32, f64
 	f64.store	$discard=, 0($0), $1
 	return
+	.endfunc
 .Lfunc_end3:
 	.size	stf64, .Lfunc_end3-stf64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/switch.s b/test/llvm_autogenerated/switch.s
index d64cfd92b..24438e83b 100644
--- a/test/llvm_autogenerated/switch.s
+++ b/test/llvm_autogenerated/switch.s
@@ -41,6 +41,7 @@ bar32:
 .LBB0_8:
 	end_block
 	return
+	.endfunc
 .Lfunc_end0:
 	.size	bar32, .Lfunc_end0-bar32
 
@@ -86,8 +87,8 @@ bar64:
 .LBB1_8:
 	end_block
 	return
+	.endfunc
 .Lfunc_end1:
 	.size	bar64, .Lfunc_end1-bar64
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/unreachable.s b/test/llvm_autogenerated/unreachable.s
index 83ec5b4b1..94c5dff17 100644
--- a/test/llvm_autogenerated/unreachable.s
+++ b/test/llvm_autogenerated/unreachable.s
@@ -6,6 +6,7 @@ f1:
 	.result 	i32
 	call    	abort@FUNCTION
 	unreachable
+	.endfunc
 .Lfunc_end0:
 	.size	f1, .Lfunc_end0-f1
 
@@ -14,6 +15,7 @@ f1:
 f2:
 	unreachable
 	return
+	.endfunc
 .Lfunc_end1:
 	.size	f2, .Lfunc_end1-f2
 
@@ -22,8 +24,8 @@ f2:
 f3:
 	unreachable
 	return
+	.endfunc
 .Lfunc_end2:
 	.size	f3, .Lfunc_end2-f3
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/unused-argument.s b/test/llvm_autogenerated/unused-argument.s
index 48ab32e38..00d57fba8 100644
--- a/test/llvm_autogenerated/unused-argument.s
+++ b/test/llvm_autogenerated/unused-argument.s
@@ -6,6 +6,7 @@ unused_first:
 	.param  	i32, i32
 	.result 	i32
 	return  	$1
+	.endfunc
 .Lfunc_end0:
 	.size	unused_first, .Lfunc_end0-unused_first
 
@@ -15,6 +16,7 @@ unused_second:
 	.param  	i32, i32
 	.result 	i32
 	return  	$0
+	.endfunc
 .Lfunc_end1:
 	.size	unused_second, .Lfunc_end1-unused_second
 
@@ -23,8 +25,8 @@ unused_second:
 call_something:
 	i32.call	$discard=, return_something@FUNCTION
 	return
+	.endfunc
 .Lfunc_end2:
 	.size	call_something, .Lfunc_end2-call_something
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/userstack.s b/test/llvm_autogenerated/userstack.s
index d77c7bc81..f7916e37c 100644
--- a/test/llvm_autogenerated/userstack.s
+++ b/test/llvm_autogenerated/userstack.s
@@ -17,6 +17,7 @@ alloca32:
 	i32.const	$2=, __stack_pointer
 	i32.store	$3=, 0($2), $3
 	return
+	.endfunc
 .Lfunc_end0:
 	.size	alloca32, .Lfunc_end0-alloca32
 
@@ -39,31 +40,33 @@ alloca3264:
 	i32.const	$2=, __stack_pointer
 	i32.store	$3=, 0($2), $3
 	return
+	.endfunc
 .Lfunc_end1:
 	.size	alloca3264, .Lfunc_end1-alloca3264
 
 	.globl	allocarray
 	.type	allocarray,@function
 allocarray:
-	.local  	i32, i32, i32, i32, i32, i32
+	.local  	i32, i32, i32, i32, i32
+	i32.const	$0=, __stack_pointer
+	i32.load	$0=, 0($0)
+	i32.const	$1=, 32
+	i32.sub 	$4=, $0, $1
 	i32.const	$1=, __stack_pointer
-	i32.load	$1=, 0($1)
+	i32.store	$4=, 0($1), $4
+	i32.const	$push2=, 12
+	i32.const	$3=, 12
+	i32.add 	$3=, $4, $3
+	i32.add 	$push3=, $3, $pop2
+	i32.const	$push0=, 1
+	i32.store	$push1=, 12($4), $pop0
+	i32.store	$discard=, 0($pop3), $pop1
 	i32.const	$2=, 32
-	i32.sub 	$5=, $1, $2
+	i32.add 	$4=, $4, $2
 	i32.const	$2=, __stack_pointer
-	i32.store	$5=, 0($2), $5
-	i32.const	$push0=, 1
-	i32.store	$0=, 12($5), $pop0
-	i32.const	$push1=, 4
-	i32.const	$4=, 12
-	i32.add 	$4=, $5, $4
-	i32.add 	$push2=, $4, $pop1
-	i32.store	$discard=, 0($pop2), $0
-	i32.const	$3=, 32
-	i32.add 	$5=, $5, $3
-	i32.const	$3=, __stack_pointer
-	i32.store	$5=, 0($3), $5
+	i32.store	$4=, 0($2), $4
 	return
+	.endfunc
 .Lfunc_end2:
 	.size	allocarray, .Lfunc_end2-allocarray
 
@@ -79,12 +82,13 @@ allocarray_inbounds:
 	i32.store	$3=, 0($1), $3
 	i32.const	$push0=, 1
 	i32.store	$push1=, 12($3), $pop0
-	i32.store	$discard=, 16($3), $pop1
+	i32.store	$discard=, 24($3), $pop1
 	i32.const	$2=, 32
 	i32.add 	$3=, $3, $2
 	i32.const	$2=, __stack_pointer
 	i32.store	$3=, 0($2), $3
 	return
+	.endfunc
 .Lfunc_end3:
 	.size	allocarray_inbounds, .Lfunc_end3-allocarray_inbounds
 
@@ -93,8 +97,8 @@ allocarray_inbounds:
 dynamic_alloca:
 	.param  	i32
 	return
+	.endfunc
 .Lfunc_end4:
 	.size	dynamic_alloca, .Lfunc_end4-dynamic_alloca
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/userstack.wast b/test/llvm_autogenerated/userstack.wast
index c0e032e93..df8788cfb 100644
--- a/test/llvm_autogenerated/userstack.wast
+++ b/test/llvm_autogenerated/userstack.wast
@@ -133,73 +133,69 @@
     (local $$2 i32)
     (local $$3 i32)
     (local $$4 i32)
-    (local $$5 i32)
     (block $fake_return_waka123
       (block
-        (set_local $$1
+        (set_local $$0
           (i32.const 1)
         )
-        (set_local $$1
+        (set_local $$0
           (i32.load align=4
-            (get_local $$1)
+            (get_local $$0)
           )
         )
-        (set_local $$2
+        (set_local $$1
           (i32.const 32)
         )
-        (set_local $$5
+        (set_local $$4
           (i32.sub
+            (get_local $$0)
             (get_local $$1)
-            (get_local $$2)
           )
         )
-        (set_local $$2
+        (set_local $$1
           (i32.const 1)
         )
-        (set_local $$5
+        (set_local $$4
           (i32.store align=4
-            (get_local $$2)
-            (get_local $$5)
-          )
-        )
-        (set_local $$0
-          (i32.store offset=12 align=4
-            (get_local $$5)
-            (i32.const 1)
+            (get_local $$1)
+            (get_local $$4)
           )
         )
-        (set_local $$4
+        (set_local $$3
           (i32.const 12)
         )
-        (set_local $$4
+        (set_local $$3
           (i32.add
-            (get_local $$5)
             (get_local $$4)
+            (get_local $$3)
           )
         )
         (i32.store align=4
           (i32.add
+            (get_local $$3)
+            (i32.const 12)
+          )
+          (i32.store offset=12 align=4
             (get_local $$4)
-            (i32.const 4)
+            (i32.const 1)
           )
-          (get_local $$0)
         )
-        (set_local $$3
+        (set_local $$2
           (i32.const 32)
         )
-        (set_local $$5
+        (set_local $$4
           (i32.add
-            (get_local $$5)
-            (get_local $$3)
+            (get_local $$4)
+            (get_local $$2)
           )
         )
-        (set_local $$3
+        (set_local $$2
           (i32.const 1)
         )
-        (set_local $$5
+        (set_local $$4
           (i32.store align=4
-            (get_local $$3)
-            (get_local $$5)
+            (get_local $$2)
+            (get_local $$4)
           )
         )
         (br $fake_return_waka123)
@@ -239,7 +235,7 @@
             (get_local $$3)
           )
         )
-        (i32.store offset=16 align=4
+        (i32.store offset=24 align=4
           (get_local $$3)
           (i32.store offset=12 align=4
             (get_local $$3)
diff --git a/test/llvm_autogenerated/varargs.s b/test/llvm_autogenerated/varargs.s
index feb0a2659..2d29fe3de 100644
--- a/test/llvm_autogenerated/varargs.s
+++ b/test/llvm_autogenerated/varargs.s
@@ -5,6 +5,7 @@
 end:
 	.param  	i32
 	return
+	.endfunc
 .Lfunc_end0:
 	.size	end, .Lfunc_end0-end
 
@@ -15,6 +16,7 @@ copy:
 	i32.load	$push0=, 0($1)
 	i32.store	$discard=, 0($0), $pop0
 	return
+	.endfunc
 .Lfunc_end1:
 	.size	copy, .Lfunc_end1-copy
 
@@ -30,6 +32,7 @@ arg_i8:
 	i32.store	$discard=, 0($0), $pop1
 	i32.load	$push2=, 0($1)
 	return  	$pop2
+	.endfunc
 .Lfunc_end2:
 	.size	arg_i8, .Lfunc_end2-arg_i8
 
@@ -49,6 +52,7 @@ arg_i32:
 	i32.store	$discard=, 0($0), $pop5
 	i32.load	$push6=, 0($1)
 	return  	$pop6
+	.endfunc
 .Lfunc_end3:
 	.size	arg_i32, .Lfunc_end3-arg_i32
 
@@ -56,24 +60,26 @@ arg_i32:
 	.type	arg_i128,@function
 arg_i128:
 	.param  	i32, i32
-	.local  	i32, i32, i32, i64
+	.local  	i32, i32, i64
 	i32.load	$push0=, 0($1)
 	i32.const	$push1=, 7
 	i32.add 	$push2=, $pop0, $pop1
 	i32.const	$push3=, -8
 	i32.and 	$2=, $pop2, $pop3
-	i32.const	$3=, 8
-	i32.add 	$push4=, $2, $3
-	i32.store	$4=, 0($1), $pop4
-	i64.load	$5=, 0($2)
-	i32.const	$push5=, 16
-	i32.add 	$push6=, $2, $pop5
-	i32.store	$discard=, 0($1), $pop6
-	i32.add 	$push8=, $0, $3
-	i64.load	$push7=, 0($4)
-	i64.store	$discard=, 0($pop8), $pop7
-	i64.store	$discard=, 0($0), $5
+	i32.const	$push4=, 8
+	i32.add 	$push5=, $2, $pop4
+	i32.store	$3=, 0($1), $pop5
+	i64.load	$4=, 0($2)
+	i32.const	$push6=, 16
+	i32.add 	$push7=, $2, $pop6
+	i32.store	$discard=, 0($1), $pop7
+	i32.const	$push10=, 8
+	i32.add 	$push9=, $0, $pop10
+	i64.load	$push8=, 0($3)
+	i64.store	$discard=, 0($pop9), $pop8
+	i64.store	$discard=, 0($0), $4
 	return
+	.endfunc
 .Lfunc_end4:
 	.size	arg_i128, .Lfunc_end4-arg_i128
 
@@ -82,15 +88,46 @@ arg_i128:
 caller_none:
 	call    	callee@FUNCTION
 	return
+	.endfunc
 .Lfunc_end5:
 	.size	caller_none, .Lfunc_end5-caller_none
 
 	.globl	caller_some
 	.type	caller_some,@function
 caller_some:
+	.local  	i32, i32, i32, i32, i32, i32, i32, i32, i32
+	i32.const	$5=, __stack_pointer
+	i32.load	$5=, 0($5)
+	i32.const	$6=, 16
+	i32.sub 	$8=, $5, $6
+	i32.const	$6=, __stack_pointer
+	i32.store	$8=, 0($6), $8
+	i32.const	$1=, __stack_pointer
+	i32.load	$1=, 0($1)
+	i32.const	$2=, 16
+	i32.sub 	$8=, $1, $2
+	i32.const	$2=, __stack_pointer
+	i32.store	$8=, 0($2), $8
+	i32.const	$push0=, 0
+	i32.store	$discard=, 0($8), $pop0
+	i32.const	$push1=, 8
+	i32.add 	$0=, $8, $pop1
+	i64.const	$push2=, 4611686018427387904
+	i64.store	$discard=, 0($0), $pop2
+	call    	callee@FUNCTION
+	i32.const	$3=, __stack_pointer
+	i32.load	$3=, 0($3)
+	i32.const	$4=, 16
+	i32.add 	$8=, $3, $4
+	i32.const	$4=, __stack_pointer
+	i32.store	$8=, 0($4), $8
+	i32.const	$7=, 16
+	i32.add 	$8=, $8, $7
+	i32.const	$7=, __stack_pointer
+	i32.store	$8=, 0($7), $8
 	return
+	.endfunc
 .Lfunc_end6:
 	.size	caller_some, .Lfunc_end6-caller_some
 
 
-	.section	".note.GNU-stack","",@progbits
diff --git a/test/llvm_autogenerated/varargs.wast b/test/llvm_autogenerated/varargs.wast
index aaa15d36c..15a5fff1d 100644
--- a/test/llvm_autogenerated/varargs.wast
+++ b/test/llvm_autogenerated/varargs.wast
@@ -86,8 +86,7 @@
   (func $arg_i128 (param $$0 i32) (param $$1 i32)
     (local $$2 i32)
     (local $$3 i32)
-    (local $$4 i32)
-    (local $$5 i64)
+    (local $$4 i64)
     (block $fake_return_waka123
       (block
         (set_local $$2
@@ -102,18 +101,15 @@
           )
         )
         (set_local $$3
-          (i32.const 8)
-        )
-        (set_local $$4
           (i32.store align=4
             (get_local $$1)
             (i32.add
               (get_local $$2)
-              (get_local $$3)
+              (i32.const 8)
             )
           )
         )
-        (set_local $$5
+        (set_local $$4
           (i64.load align=8
             (get_local $$2)
           )
@@ -128,15 +124,15 @@
         (i64.store align=8
           (i32.add
             (get_local $$0)
-            (get_local $$3)
+            (i32.const 8)
           )
           (i64.load align=8
-            (get_local $$4)
+            (get_local $$3)
           )
         )
         (i64.store align=8
           (get_local $$0)
-          (get_local $$5)
+          (get_local $$4)
         )
         (br $fake_return_waka123)
       )
@@ -151,8 +147,128 @@
     )
   )
   (func $caller_some
+    (local $$0 i32)
+    (local $$1 i32)
+    (local $$2 i32)
+    (local $$3 i32)
+    (local $$4 i32)
+    (local $$5 i32)
+    (local $$6 i32)
+    (local $$7 i32)
+    (local $$8 i32)
     (block $fake_return_waka123
       (block
+        (set_local $$5
+          (i32.const 1)
+        )
+        (set_local $$5
+          (i32.load align=4
+            (get_local $$5)
+          )
+        )
+        (set_local $$6
+          (i32.const 16)
+        )
+        (set_local $$8
+          (i32.sub
+            (get_local $$5)
+            (get_local $$6)
+          )
+        )
+        (set_local $$6
+          (i32.const 1)
+        )
+        (set_local $$8
+          (i32.store align=4
+            (get_local $$6)
+            (get_local $$8)
+          )
+        )
+        (set_local $$1
+          (i32.const 1)
+        )
+        (set_local $$1
+          (i32.load align=4
+            (get_local $$1)
+          )
+        )
+        (set_local $$2
+          (i32.const 16)
+        )
+        (set_local $$8
+          (i32.sub
+            (get_local $$1)
+            (get_local $$2)
+          )
+        )
+        (set_local $$2
+          (i32.const 1)
+        )
+        (set_local $$8
+          (i32.store align=4
+            (get_local $$2)
+            (get_local $$8)
+          )
+        )
+        (i32.store align=4
+          (get_local $$8)
+          (i32.const 0)
+        )
+        (set_local $$0
+          (i32.add
+            (get_local $$8)
+            (i32.const 8)
+          )
+        )
+        (i64.store align=8
+          (get_local $$0)
+          (i64.const 4611686018427387904)
+        )
+        (call_import $callee)
+        (set_local $$3
+          (i32.const 1)
+        )
+        (set_local $$3
+          (i32.load align=4
+            (get_local $$3)
+          )
+        )
+        (set_local $$4
+          (i32.const 16)
+        )
+        (set_local $$8
+          (i32.add
+            (get_local $$3)
+            (get_local $$4)
+          )
+        )
+        (set_local $$4
+          (i32.const 1)
+        )
+        (set_local $$8
+          (i32.store align=4
+            (get_local $$4)
+            (get_local $$8)
+          )
+        )
+        (set_local $$7
+          (i32.const 16)
+        )
+        (set_local $$8
+          (i32.add
+            (get_local $$8)
+            (get_local $$7)
+          )
+        )
+        (set_local $$7
+          (i32.const 1)
+        )
+        (set_local $$8
+          (i32.store align=4
+            (get_local $$7)
+            (get_local $$8)
+          )
+        )
         (br $fake_return_waka123)
       )
     )
diff --git a/test/spec b/test/spec
-Subproject 60afe7437286317f0926236cef2587dd4aa5d4d
+Subproject de27ac1b9f0cdf2b80aa4a26392d2cede1dfdfc