Merge pull request #286 from WebAssembly/asm2wasm-opts

asm2wasm opts

7 files changed, 148 insertions, 1 deletions

diff --git a/src/asm2wasm.h b/src/asm2wasm.h
index 4c37b7705..7d9b9652b 100644
--- a/src/asm2wasm.h
+++ b/src/asm2wasm.h
@@ -202,6 +202,8 @@ private:
   IString Math_ceil;
   IString Math_sqrt;
 
+  IString tempDoublePtr; // imported name of tempDoublePtr
+
   // function types. we fill in this information as we see
   // uses, in the first pass
 
@@ -483,6 +485,13 @@ void Asm2WasmBuilder::processAsm(Ref ast) {
     } else {
       assert(module[0] == NAME);
       moduleName = module[1]->getIString();
+      if (moduleName == ENV) {
+        if (imported[2] == TEMP_DOUBLE_PTR) {
+          assert(tempDoublePtr.isNull());
+          tempDoublePtr = name;
+          // we don't return here, as we can only optimize out some uses of tDP. So it remains imported
+        }
+      }
     }
     auto import = allocator.alloc<Import>();
     import->name = name;
@@ -1404,6 +1413,57 @@ Function* Asm2WasmBuilder::processFunction(Ref ast) {
       ret->type = ret->ifTrue->type;
       return ret;
     } else if (what == SEQ) {
+      // Some (x, y) patterns can be optimized, like bitcasts,
+      //  (HEAP32[tempDoublePtr >> 2] = i, Math_fround(HEAPF32[tempDoublePtr >> 2])); // i32->f32
+      //  (HEAP32[tempDoublePtr >> 2] = i, +HEAPF32[tempDoublePtr >> 2]); // i32->f32, no fround
+      //  (HEAPF32[tempDoublePtr >> 2] = f, HEAP32[tempDoublePtr >> 2] | 0); // f32->i32
+      if (ast[1][0] == ASSIGN && ast[1][2][0] == SUB && ast[1][2][1][0] == NAME && ast[1][2][2][0] == BINARY && ast[1][2][2][1] == RSHIFT &&
+          ast[1][2][2][2][0] == NAME && ast[1][2][2][2][1] == tempDoublePtr && ast[1][2][2][3][0] == NUM && ast[1][2][2][3][1]->getNumber() == 2) {
+        // (?[tempDoublePtr >> 2] = ?, ?)  so far
+        auto heap = ast[1][2][1][1]->getIString();
+        if (views.find(heap) != views.end()) {
+          AsmType writeType = views[heap].type;
+          AsmType readType = ASM_NONE;
+          Ref readValue;
+          if (ast[2][0] == BINARY && ast[2][1] == OR && ast[2][3][0] == NUM && ast[2][3][1]->getNumber() == 0) {
+            readType = ASM_INT;
+            readValue = ast[2][2];
+          } else if (ast[2][0] == UNARY_PREFIX && ast[2][1] == PLUS) {
+            readType = ASM_DOUBLE;
+            readValue = ast[2][2];
+          } else if (ast[2][0] == CALL && ast[2][1][0] == NAME && ast[2][1][1] == Math_fround) {
+            readType = ASM_FLOAT;
+            readValue = ast[2][2][0];
+          }
+          if (readType != ASM_NONE) {
+            if (readValue[0] == SUB && readValue[1][0] == NAME && readValue[2][0] == BINARY && readValue[2][1] == RSHIFT &&
+                readValue[2][2][0] == NAME && readValue[2][2][1] == tempDoublePtr && readValue[2][3][0] == NUM && readValue[2][3][1]->getNumber() == 2) {
+              // pattern looks right!
+              Ref writtenValue = ast[1][3];
+              if (writeType == ASM_INT && (readType == ASM_FLOAT || readType == ASM_DOUBLE)) {
+                auto conv = allocator.alloc<Unary>();
+                conv->op = ReinterpretInt;
+                conv->value = process(writtenValue);
+                conv->type = WasmType::f32;
+                if (readType == ASM_DOUBLE) {
+                  auto promote = allocator.alloc<Unary>();
+                  promote->op = PromoteFloat32;
+                  promote->value = conv;
+                  promote->type = WasmType::f64;
+                  return promote;
+                }
+                return conv;
+              } else if (writeType == ASM_FLOAT && readType == ASM_INT) {
+                auto conv = allocator.alloc<Unary>();
+                conv->op = ReinterpretFloat;
+                conv->value = process(writtenValue);
+                conv->type = WasmType::i32;
+                return conv;
+              }
+            }
+          }
+        }
+      }
       auto ret = allocator.alloc<Block>();
       ret->list.push_back(process(ast[1]));
       ret->list.push_back(process(ast[2]));
diff --git a/test/min.asm.js b/test/min.asm.js
index 77e54caab..ed5d186f5 100644
--- a/test/min.asm.js
+++ b/test/min.asm.js
@@ -6,6 +6,17 @@ function (global, env, buffer) {
 
   var fr = global.Math.fround;
 
+  var tDP = env.tempDoublePtr | 0;
+
+  var h8 = new global.Int8Array(buffer);
+  var h16 = new global.Int16Array(buffer);
+  var h32 = new global.Int32Array(buffer);
+  var hU8 = new global.Uint8Array(buffer);
+  var hU16 = new global.Uint16Array(buffer);
+  var hU32 = new global.Uint32Array(buffer);
+  var hF32 = new global.Float32Array(buffer);
+  var hF64 = new global.Float64Array(buffer);
+
   function floats(f) {
     f = fr(f);
     var t = fr(0);
@@ -18,6 +29,13 @@ function (global, env, buffer) {
     n = fr(-(c[k >> 2] = p, fr(g[k >> 2])));
     return n;
   }
+  function bitcasts(i, f) {
+    i = i | 0;
+    f = Math_fround(f);
+    (h32[tDP >> 2] = i, fr(hF32[tDP >> 2])); // i32->f32
+    (h32[tDP >> 2] = i, +hF32[tDP >> 2]); // i32->f32, no fround
+    (hF32[tDP >> 2] = f, h32[tDP >> 2] | 0); // f32->i32
+  }
 
   return { floats: floats };
 }
diff --git a/test/min.fromasm b/test/min.fromasm
index 839fd167c..564980a96 100644
--- a/test/min.fromasm
+++ b/test/min.fromasm
@@ -30,4 +30,17 @@
       (get_local $n)
     )
   )
+  (func $bitcasts (param $i i32) (param $f f32)
+    (f32.reinterpret/i32
+      (get_local $i)
+    )
+    (f64.promote/f32
+      (f32.reinterpret/i32
+        (get_local $i)
+      )
+    )
+    (i32.reinterpret/f32
+      (get_local $f)
+    )
+  )
 )
diff --git a/test/min.fromasm.imprecise b/test/min.fromasm.imprecise
index 839fd167c..564980a96 100644
--- a/test/min.fromasm.imprecise
+++ b/test/min.fromasm.imprecise
@@ -30,4 +30,17 @@
       (get_local $n)
     )
   )
+  (func $bitcasts (param $i i32) (param $f f32)
+    (f32.reinterpret/i32
+      (get_local $i)
+    )
+    (f64.promote/f32
+      (f32.reinterpret/i32
+        (get_local $i)
+      )
+    )
+    (i32.reinterpret/f32
+      (get_local $f)
+    )
+  )
 )
diff --git a/test/unit.asm.js b/test/unit.asm.js
index 5bb84bdcd..cd432da5c 100644
--- a/test/unit.asm.js
+++ b/test/unit.asm.js
@@ -1,4 +1,4 @@
-function asm() {
+function asm(global, env, buffer) {
   "use asm";
 
   var t = global.NaN, u = global.Infinity;
@@ -7,11 +7,21 @@ function asm() {
   var Math_fround = global.Math.fround;
   var Math_abs = global.Math.abs;
   var Math_ceil = global.Math.ceil;
+  var tempDoublePtr = env.tempDoublePtr | 0;
 
   var abort = env.abort;
   var print = env.print;
   var h = env.h;
 
+  var HEAP8 = new global.Int8Array(buffer);
+  var HEAP16 = new global.Int16Array(buffer);
+  var HEAP32 = new global.Int32Array(buffer);
+  var HEAPU8 = new global.Uint8Array(buffer);
+  var HEAPU16 = new global.Uint16Array(buffer);
+  var HEAPU32 = new global.Uint32Array(buffer);
+  var HEAPF32 = new global.Float32Array(buffer);
+  var HEAPF64 = new global.Float64Array(buffer);
+
   function big_negative() {
     var temp = 0.0;
     temp = +-2147483648;
@@ -190,6 +200,13 @@ function asm() {
       print(2);
     }
   }
+  function bitcasts(i, f) {
+    i = i | 0;
+    f = Math_fround(f);
+    (HEAP32[tempDoublePtr >> 2] = i, Math_fround(HEAPF32[tempDoublePtr >> 2])); // i32->f32
+    (HEAP32[tempDoublePtr >> 2] = i, +HEAPF32[tempDoublePtr >> 2]); // i32->f32, no fround
+    (HEAPF32[tempDoublePtr >> 2] = f, HEAP32[tempDoublePtr >> 2] | 0); // f32->i32
+  }
 
   function z() {
   }
diff --git a/test/unit.fromasm b/test/unit.fromasm
index 3db25c78b..d7fce8b8f 100644
--- a/test/unit.fromasm
+++ b/test/unit.fromasm
@@ -516,6 +516,19 @@
       (br $while-in$1)
     )
   )
+  (func $bitcasts (param $i i32) (param $f f32)
+    (f32.reinterpret/i32
+      (get_local $i)
+    )
+    (f64.promote/f32
+      (f32.reinterpret/i32
+        (get_local $i)
+      )
+    )
+    (i32.reinterpret/f32
+      (get_local $f)
+    )
+  )
   (func $z
     (nop)
   )
diff --git a/test/unit.fromasm.imprecise b/test/unit.fromasm.imprecise
index ef9ccaa5d..6383d6e87 100644
--- a/test/unit.fromasm.imprecise
+++ b/test/unit.fromasm.imprecise
@@ -512,6 +512,19 @@
       (br $while-in$1)
     )
   )
+  (func $bitcasts (param $i i32) (param $f f32)
+    (f32.reinterpret/i32
+      (get_local $i)
+    )
+    (f64.promote/f32
+      (f32.reinterpret/i32
+        (get_local $i)
+      )
+    )
+    (i32.reinterpret/f32
+      (get_local $f)
+    )
+  )
   (func $z
     (nop)
   )