summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--src/passes/CodeFolding.cpp25
-rw-r--r--src/passes/pass.cpp2
-rw-r--r--test/emcc_O2_hello_world.fromasm26
-rw-r--r--test/emcc_O2_hello_world.fromasm.clamp26
-rw-r--r--test/emcc_O2_hello_world.fromasm.imprecise26
-rw-r--r--test/emcc_hello_world.fromasm15
-rw-r--r--test/emcc_hello_world.fromasm.clamp15
-rw-r--r--test/emcc_hello_world.fromasm.imprecise15
-rw-r--r--test/memorygrowth.fromasm26
-rw-r--r--test/memorygrowth.fromasm.clamp26
-rw-r--r--test/memorygrowth.fromasm.imprecise26
-rw-r--r--test/passes/remove-unused-names_code-folding.txt31
-rw-r--r--test/passes/remove-unused-names_code-folding.wast18
13 files changed, 174 insertions, 103 deletions
diff --git a/src/passes/CodeFolding.cpp b/src/passes/CodeFolding.cpp
index c2ad07595..b639c7681 100644
--- a/src/passes/CodeFolding.cpp
+++ b/src/passes/CodeFolding.cpp
@@ -15,7 +15,8 @@
*/
//
-// Folds duplicate code together, saving space.
+// Folds duplicate code together, saving space (and possibly phis in
+// the wasm VM, which can save time).
//
// We fold tails of code where they merge and moving the code
// to the merge point is helpful. There are two cases here: (1) expressions,
@@ -198,8 +199,6 @@ struct CodeFolding : public WalkerPass<ControlFlowWalker<CodeFolding>> {
void visitIf(If* curr) {
if (!curr->ifFalse) return;
// if both sides are identical, this is easy to fold
- // (except if the condition is unreachable and we return a value, then we can't just replace
- // outselves with a drop
if (ExpressionAnalyzer::equal(curr->ifTrue, curr->ifFalse)) {
Builder builder(*getModule());
// remove if (4 bytes), remove one arm, add drop (1), add block (3),
@@ -216,6 +215,26 @@ struct CodeFolding : public WalkerPass<ControlFlowWalker<CodeFolding>> {
// if both are blocks, look for a tail we can merge
auto* left = curr->ifTrue->dynCast<Block>();
auto* right = curr->ifFalse->dynCast<Block>();
+ // If one is a block and the other isn't, and the non-block is a tail
+ // of the other, we can fold that - for our convenience, we just add
+ // a block and run the rest of the optimization mormally.
+ auto maybeAddBlock = [this](Block* block, Expression*& other) -> Block* {
+ // if other is a suffix of the block, wrap it in a block
+ if (block->list.empty() ||
+ !ExpressionAnalyzer::equal(other, block->list.back())) {
+ return nullptr;
+ }
+ // do it, assign to the out param `other`, and return the block
+ Builder builder(*getModule());
+ auto* ret = builder.makeBlock(other);
+ other = ret;
+ return ret;
+ };
+ if (left && !right) {
+ right = maybeAddBlock(left, curr->ifFalse);
+ } else if (!left && right) {
+ left = maybeAddBlock(right, curr->ifTrue);
+ }
// we need nameless blocks, as if there is a name, someone might branch
// to the end, skipping the code we want to merge
if (left && right &&
diff --git a/src/passes/pass.cpp b/src/passes/pass.cpp
index 38a21fae8..0ec5864e0 100644
--- a/src/passes/pass.cpp
+++ b/src/passes/pass.cpp
@@ -154,7 +154,7 @@ void PassRunner::addDefaultFunctionOptimizationPasses() {
add("simplify-locals");
add("vacuum"); // previous pass creates garbage
add("reorder-locals");
- if (options.shrinkLevel >= 1) {
+ if (options.optimizeLevel >= 3 || options.shrinkLevel >= 1) {
add("code-folding");
}
add("merge-blocks"); // makes remove-unused-brs more effective
diff --git a/test/emcc_O2_hello_world.fromasm b/test/emcc_O2_hello_world.fromasm
index b046c9fe9..25b5a1be8 100644
--- a/test/emcc_O2_hello_world.fromasm
+++ b/test/emcc_O2_hello_world.fromasm
@@ -3097,15 +3097,16 @@
(i32.eq
(tee_local $7
(block $label$break$L257 (result i32)
- (if (result i32)
- (i32.and
- (i32.load
- (i32.const 620)
+ (if
+ (i32.eqz
+ (i32.and
+ (i32.load
+ (i32.const 620)
+ )
+ (i32.const 4)
)
- (i32.const 4)
)
- (i32.const 190)
- (block (result i32)
+ (block
(block $label$break$L259
(if
(tee_local $10
@@ -3478,9 +3479,9 @@
(i32.const 4)
)
)
- (i32.const 190)
)
)
+ (i32.const 190)
)
)
(i32.const 190)
@@ -7894,13 +7895,13 @@
)
)
)
- (set_local $5
- (if (result i32)
+ (block
+ (if
(i32.eq
(get_local $3)
(i32.const 2)
)
- (block (result i32)
+ (block
(i32.store
(get_local $8)
(i32.add
@@ -7913,8 +7914,9 @@
(set_local $3
(i32.const 2)
)
- (get_local $14)
)
+ )
+ (set_local $5
(get_local $14)
)
)
diff --git a/test/emcc_O2_hello_world.fromasm.clamp b/test/emcc_O2_hello_world.fromasm.clamp
index b046c9fe9..25b5a1be8 100644
--- a/test/emcc_O2_hello_world.fromasm.clamp
+++ b/test/emcc_O2_hello_world.fromasm.clamp
@@ -3097,15 +3097,16 @@
(i32.eq
(tee_local $7
(block $label$break$L257 (result i32)
- (if (result i32)
- (i32.and
- (i32.load
- (i32.const 620)
+ (if
+ (i32.eqz
+ (i32.and
+ (i32.load
+ (i32.const 620)
+ )
+ (i32.const 4)
)
- (i32.const 4)
)
- (i32.const 190)
- (block (result i32)
+ (block
(block $label$break$L259
(if
(tee_local $10
@@ -3478,9 +3479,9 @@
(i32.const 4)
)
)
- (i32.const 190)
)
)
+ (i32.const 190)
)
)
(i32.const 190)
@@ -7894,13 +7895,13 @@
)
)
)
- (set_local $5
- (if (result i32)
+ (block
+ (if
(i32.eq
(get_local $3)
(i32.const 2)
)
- (block (result i32)
+ (block
(i32.store
(get_local $8)
(i32.add
@@ -7913,8 +7914,9 @@
(set_local $3
(i32.const 2)
)
- (get_local $14)
)
+ )
+ (set_local $5
(get_local $14)
)
)
diff --git a/test/emcc_O2_hello_world.fromasm.imprecise b/test/emcc_O2_hello_world.fromasm.imprecise
index ff0fa1fe3..3db79a820 100644
--- a/test/emcc_O2_hello_world.fromasm.imprecise
+++ b/test/emcc_O2_hello_world.fromasm.imprecise
@@ -3096,15 +3096,16 @@
(i32.eq
(tee_local $7
(block $label$break$L257 (result i32)
- (if (result i32)
- (i32.and
- (i32.load
- (i32.const 620)
+ (if
+ (i32.eqz
+ (i32.and
+ (i32.load
+ (i32.const 620)
+ )
+ (i32.const 4)
)
- (i32.const 4)
)
- (i32.const 190)
- (block (result i32)
+ (block
(block $label$break$L259
(if
(tee_local $10
@@ -3477,9 +3478,9 @@
(i32.const 4)
)
)
- (i32.const 190)
)
)
+ (i32.const 190)
)
)
(i32.const 190)
@@ -7893,13 +7894,13 @@
)
)
)
- (set_local $5
- (if (result i32)
+ (block
+ (if
(i32.eq
(get_local $3)
(i32.const 2)
)
- (block (result i32)
+ (block
(i32.store
(get_local $8)
(i32.add
@@ -7912,8 +7913,9 @@
(set_local $3
(i32.const 2)
)
- (get_local $14)
)
+ )
+ (set_local $5
(get_local $14)
)
)
diff --git a/test/emcc_hello_world.fromasm b/test/emcc_hello_world.fromasm
index 16af1a160..0af69c9fe 100644
--- a/test/emcc_hello_world.fromasm
+++ b/test/emcc_hello_world.fromasm
@@ -579,17 +579,14 @@
(get_local $0)
)
)
- (set_local $0
- (if (result i32)
- (tee_local $2
- (i32.eqz
- (i32.const 0)
- )
- )
- (get_local $1)
- (get_local $1)
+ (set_local $2
+ (i32.eqz
+ (i32.const 0)
)
)
+ (set_local $0
+ (get_local $1)
+ )
)
(block
(set_local $0
diff --git a/test/emcc_hello_world.fromasm.clamp b/test/emcc_hello_world.fromasm.clamp
index 3a6147692..a72cae0f6 100644
--- a/test/emcc_hello_world.fromasm.clamp
+++ b/test/emcc_hello_world.fromasm.clamp
@@ -577,17 +577,14 @@
(get_local $0)
)
)
- (set_local $0
- (if (result i32)
- (tee_local $2
- (i32.eqz
- (i32.const 0)
- )
- )
- (get_local $1)
- (get_local $1)
+ (set_local $2
+ (i32.eqz
+ (i32.const 0)
)
)
+ (set_local $0
+ (get_local $1)
+ )
)
(block
(set_local $0
diff --git a/test/emcc_hello_world.fromasm.imprecise b/test/emcc_hello_world.fromasm.imprecise
index dae544f1b..70f3b96a3 100644
--- a/test/emcc_hello_world.fromasm.imprecise
+++ b/test/emcc_hello_world.fromasm.imprecise
@@ -576,17 +576,14 @@
(get_local $0)
)
)
- (set_local $0
- (if (result i32)
- (tee_local $2
- (i32.eqz
- (i32.const 0)
- )
- )
- (get_local $1)
- (get_local $1)
+ (set_local $2
+ (i32.eqz
+ (i32.const 0)
)
)
+ (set_local $0
+ (get_local $1)
+ )
)
(block
(set_local $0
diff --git a/test/memorygrowth.fromasm b/test/memorygrowth.fromasm
index 031e1b8ce..7d5380cd0 100644
--- a/test/memorygrowth.fromasm
+++ b/test/memorygrowth.fromasm
@@ -3120,15 +3120,16 @@
(i32.eq
(tee_local $7
(block $label$break$b (result i32)
- (if (result i32)
- (i32.and
- (i32.load
- (i32.const 1652)
+ (if
+ (i32.eqz
+ (i32.and
+ (i32.load
+ (i32.const 1652)
+ )
+ (i32.const 4)
)
- (i32.const 4)
)
- (i32.const 188)
- (block (result i32)
+ (block
(block $label$break$c
(if
(tee_local $18
@@ -3500,9 +3501,9 @@
(i32.const 4)
)
)
- (i32.const 188)
)
)
+ (i32.const 188)
)
)
(i32.const 188)
@@ -7982,13 +7983,13 @@
)
)
)
- (set_local $5
- (if (result i32)
+ (block
+ (if
(i32.eq
(get_local $3)
(i32.const 2)
)
- (block (result i32)
+ (block
(i32.store
(get_local $8)
(i32.add
@@ -8001,8 +8002,9 @@
(set_local $3
(i32.const 2)
)
- (get_local $14)
)
+ )
+ (set_local $5
(get_local $14)
)
)
diff --git a/test/memorygrowth.fromasm.clamp b/test/memorygrowth.fromasm.clamp
index 031e1b8ce..7d5380cd0 100644
--- a/test/memorygrowth.fromasm.clamp
+++ b/test/memorygrowth.fromasm.clamp
@@ -3120,15 +3120,16 @@
(i32.eq
(tee_local $7
(block $label$break$b (result i32)
- (if (result i32)
- (i32.and
- (i32.load
- (i32.const 1652)
+ (if
+ (i32.eqz
+ (i32.and
+ (i32.load
+ (i32.const 1652)
+ )
+ (i32.const 4)
)
- (i32.const 4)
)
- (i32.const 188)
- (block (result i32)
+ (block
(block $label$break$c
(if
(tee_local $18
@@ -3500,9 +3501,9 @@
(i32.const 4)
)
)
- (i32.const 188)
)
)
+ (i32.const 188)
)
)
(i32.const 188)
@@ -7982,13 +7983,13 @@
)
)
)
- (set_local $5
- (if (result i32)
+ (block
+ (if
(i32.eq
(get_local $3)
(i32.const 2)
)
- (block (result i32)
+ (block
(i32.store
(get_local $8)
(i32.add
@@ -8001,8 +8002,9 @@
(set_local $3
(i32.const 2)
)
- (get_local $14)
)
+ )
+ (set_local $5
(get_local $14)
)
)
diff --git a/test/memorygrowth.fromasm.imprecise b/test/memorygrowth.fromasm.imprecise
index 062d52a2b..97c650686 100644
--- a/test/memorygrowth.fromasm.imprecise
+++ b/test/memorygrowth.fromasm.imprecise
@@ -3118,15 +3118,16 @@
(i32.eq
(tee_local $7
(block $label$break$b (result i32)
- (if (result i32)
- (i32.and
- (i32.load
- (i32.const 1652)
+ (if
+ (i32.eqz
+ (i32.and
+ (i32.load
+ (i32.const 1652)
+ )
+ (i32.const 4)
)
- (i32.const 4)
)
- (i32.const 188)
- (block (result i32)
+ (block
(block $label$break$c
(if
(tee_local $18
@@ -3498,9 +3499,9 @@
(i32.const 4)
)
)
- (i32.const 188)
)
)
+ (i32.const 188)
)
)
(i32.const 188)
@@ -7980,13 +7981,13 @@
)
)
)
- (set_local $5
- (if (result i32)
+ (block
+ (if
(i32.eq
(get_local $3)
(i32.const 2)
)
- (block (result i32)
+ (block
(i32.store
(get_local $8)
(i32.add
@@ -7999,8 +8000,9 @@
(set_local $3
(i32.const 2)
)
- (get_local $14)
)
+ )
+ (set_local $5
(get_local $14)
)
)
diff --git a/test/passes/remove-unused-names_code-folding.txt b/test/passes/remove-unused-names_code-folding.txt
index 5f5b30e7e..a37cbefbd 100644
--- a/test/passes/remove-unused-names_code-folding.txt
+++ b/test/passes/remove-unused-names_code-folding.txt
@@ -2,6 +2,7 @@
(type $0 (func))
(type $1 (func (param i32 i32) (result i32)))
(type $2 (func (result i32)))
+ (type $3 (func (param i32) (result i32)))
(func $ifs (; 0 ;) (type $0)
(if
(i32.const 0)
@@ -1677,4 +1678,34 @@
(i32.const 3)
)
)
+ (func $if-suffix (; 35 ;) (type $3) (param $x i32) (result i32)
+ (block
+ (if
+ (get_local $x)
+ (block
+ )
+ (drop
+ (call $if-suffix
+ (i32.const -1)
+ )
+ )
+ )
+ (set_local $x
+ (i32.const 1)
+ )
+ )
+ (block (result i32)
+ (if
+ (get_local $x)
+ (block
+ )
+ (drop
+ (call $if-suffix
+ (i32.const -2)
+ )
+ )
+ )
+ (i32.const 2)
+ )
+ )
)
diff --git a/test/passes/remove-unused-names_code-folding.wast b/test/passes/remove-unused-names_code-folding.wast
index 35d95ba90..1472c7871 100644
--- a/test/passes/remove-unused-names_code-folding.wast
+++ b/test/passes/remove-unused-names_code-folding.wast
@@ -1173,4 +1173,22 @@
)
(drop (i32.const 3))
)
+ (func $if-suffix (param $x i32) (result i32)
+ (if
+ (get_local $x)
+ (set_local $x (i32.const 1))
+ (block
+ (drop (call $if-suffix (i32.const -1)))
+ (set_local $x (i32.const 1))
+ )
+ )
+ (if (result i32)
+ (get_local $x)
+ (i32.const 2)
+ (block (result i32)
+ (drop (call $if-suffix (i32.const -2)))
+ (i32.const 2)
+ )
+ )
+ )
)
generated by cgit v1.2.3 (git 2.39.1) at 2025-08-11 13:13:30 +0000