Reapply "[WebAssembly] Constant fold wasm.dot" #153070

badumbatish · 2025-08-11T19:08:30Z

In #149619, for the test of @dot_follow_modulo_spec_2, constant folding the addition of two i32 1073741824 causes an overflow from 2^32 to -2^32=-2147483648, which triggers the UB sanitizer. This PR reapplies the previous PR, explicitly casting the addition operand to int64_t first before performing the addition before producing a int32 number via Constant *C = get(cast<IntegerType>(Ty->getScalarType()), V, isSigned)

llvmbot · 2025-08-11T19:09:02Z

@llvm/pr-subscribers-llvm-transforms

Author: Jasmine Tang (badumbatish)

Changes

In #149619, for the test of @dot_follow_modulo_spec_2, constant folding the addition of two i32 1073741824 causes an overflow from 2^32 to -2^32=-2147483648, which triggers the UB sanitizer. This PR reapplies the previous PR, explicitly allowing the wrap around.

Full diff: https://github.com/llvm/llvm-project/pull/153070.diff

2 Files Affected:

(modified) llvm/lib/Analysis/ConstantFolding.cpp (+25)
(added) llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/dot.ll (+56)

diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp
index 9c1c2c6e60f02..bf291809b07a2 100644
--- a/llvm/lib/Analysis/ConstantFolding.cpp
+++ b/llvm/lib/Analysis/ConstantFolding.cpp
@@ -1657,6 +1657,7 @@ bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {
   case Intrinsic::aarch64_sve_convert_from_svbool:
   case Intrinsic::wasm_alltrue:
   case Intrinsic::wasm_anytrue:
+  case Intrinsic::wasm_dot:
   // WebAssembly float semantics are always known
   case Intrinsic::wasm_trunc_signed:
   case Intrinsic::wasm_trunc_unsigned:
@@ -3826,6 +3827,30 @@ static Constant *ConstantFoldFixedVectorCall(
     }
     return ConstantVector::get(Result);
   }
+  case Intrinsic::wasm_dot: {
+    unsigned NumElements =
+        cast<FixedVectorType>(Operands[0]->getType())->getNumElements();
+
+    assert(NumElements == 8 && Result.size() == 4 &&
+           "wasm dot takes i16x8 and produces i32x4");
+    assert(Ty->isIntegerTy());
+    int32_t MulVector[8];
+
+    for (unsigned I = 0; I < NumElements; ++I) {
+      ConstantInt *Elt0 =
+          cast<ConstantInt>(Operands[0]->getAggregateElement(I));
+      ConstantInt *Elt1 =
+          cast<ConstantInt>(Operands[1]->getAggregateElement(I));
+
+      MulVector[I] = Elt0->getSExtValue() * Elt1->getSExtValue();
+    }
+    for (unsigned I = 0; I < Result.size(); I++) {
+      int64_t IAdd = (int64_t)MulVector[I * 2] + MulVector[I * 2 + 1];
+      Result[I] = ConstantInt::get(Ty, IAdd);
+    }
+
+    return ConstantVector::get(Result);
+  }
   default:
     break;
   }
diff --git a/llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/dot.ll b/llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/dot.ll
new file mode 100644
index 0000000000000..b537b7bccf861
--- /dev/null
+++ b/llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/dot.ll
@@ -0,0 +1,56 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+
+; RUN: opt -passes=instsimplify -S < %s | FileCheck %s
+
+; Test that intrinsics wasm dot call are constant folded
+
+target triple = "wasm32-unknown-unknown"
+
+
+define <4 x i32> @dot_zero() {
+; CHECK-LABEL: define <4 x i32> @dot_zero() {
+; CHECK-NEXT:    ret <4 x i32> zeroinitializer
+;
+  %res = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> zeroinitializer, <8 x i16> zeroinitializer)
+  ret <4 x i32> %res
+}
+
+; a               =   1    2    3    4    5    6    7    8
+; b               =   1    2    3    4    5    6    7    8
+; k1|k2 = a * b   =   1    4    9   16   25   36   49   64
+; k1 + k2         =   (1+4) |  (9 + 16) | (25 + 36)  | (49 + 64)
+; result          =    5    |   25    |   61    |   113
+define <4 x i32> @dot_nonzero() {
+; CHECK-LABEL: define <4 x i32> @dot_nonzero() {
+; CHECK-NEXT:    ret <4 x i32> <i32 5, i32 25, i32 61, i32 113>
+;
+  %res = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> <i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7, i16 8>, <8 x i16> <i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7, i16 8>)
+  ret <4 x i32> %res
+}
+
+define <4 x i32> @dot_doubly_negative() {
+; CHECK-LABEL: define <4 x i32> @dot_doubly_negative() {
+; CHECK-NEXT:    ret <4 x i32> splat (i32 2)
+;
+  %res = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>, <8 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>)
+  ret <4 x i32> %res
+}
+
+; Tests that i16 max signed values fit in i32
+define <4 x i32> @dot_follow_modulo_spec_1() {
+; CHECK-LABEL: define <4 x i32> @dot_follow_modulo_spec_1() {
+; CHECK-NEXT:    ret <4 x i32> <i32 2147352578, i32 0, i32 0, i32 0>
+;
+  %res = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> <i16 32767, i16 32767, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, <8 x i16> <i16 32767, i16 32767, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>)
+  ret <4 x i32> %res
+}
+
+; Tests that i16 min signed values fit in i32
+define <4 x i32> @dot_follow_modulo_spec_2() {
+; CHECK-LABEL: define <4 x i32> @dot_follow_modulo_spec_2() {
+; CHECK-NEXT:    ret <4 x i32> <i32 -2147483648, i32 0, i32 0, i32 0>
+;
+  %res = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> <i16 -32768, i16 -32768, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, <8 x i16> <i16 -32768, i16 -32768, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>)
+  ret <4 x i32> %res
+}
+

llvmbot · 2025-08-11T19:09:03Z

@llvm/pr-subscribers-llvm-analysis

Author: Jasmine Tang (badumbatish)

Changes

In #149619, for the test of @dot_follow_modulo_spec_2, constant folding the addition of two i32 1073741824 causes an overflow from 2^32 to -2^32=-2147483648, which triggers the UB sanitizer. This PR reapplies the previous PR, explicitly allowing the wrap around.

Full diff: https://github.com/llvm/llvm-project/pull/153070.diff

2 Files Affected:

(modified) llvm/lib/Analysis/ConstantFolding.cpp (+25)
(added) llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/dot.ll (+56)

diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp
index 9c1c2c6e60f02..bf291809b07a2 100644
--- a/llvm/lib/Analysis/ConstantFolding.cpp
+++ b/llvm/lib/Analysis/ConstantFolding.cpp
@@ -1657,6 +1657,7 @@ bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {
   case Intrinsic::aarch64_sve_convert_from_svbool:
   case Intrinsic::wasm_alltrue:
   case Intrinsic::wasm_anytrue:
+  case Intrinsic::wasm_dot:
   // WebAssembly float semantics are always known
   case Intrinsic::wasm_trunc_signed:
   case Intrinsic::wasm_trunc_unsigned:
@@ -3826,6 +3827,30 @@ static Constant *ConstantFoldFixedVectorCall(
     }
     return ConstantVector::get(Result);
   }
+  case Intrinsic::wasm_dot: {
+    unsigned NumElements =
+        cast<FixedVectorType>(Operands[0]->getType())->getNumElements();
+
+    assert(NumElements == 8 && Result.size() == 4 &&
+           "wasm dot takes i16x8 and produces i32x4");
+    assert(Ty->isIntegerTy());
+    int32_t MulVector[8];
+
+    for (unsigned I = 0; I < NumElements; ++I) {
+      ConstantInt *Elt0 =
+          cast<ConstantInt>(Operands[0]->getAggregateElement(I));
+      ConstantInt *Elt1 =
+          cast<ConstantInt>(Operands[1]->getAggregateElement(I));
+
+      MulVector[I] = Elt0->getSExtValue() * Elt1->getSExtValue();
+    }
+    for (unsigned I = 0; I < Result.size(); I++) {
+      int64_t IAdd = (int64_t)MulVector[I * 2] + MulVector[I * 2 + 1];
+      Result[I] = ConstantInt::get(Ty, IAdd);
+    }
+
+    return ConstantVector::get(Result);
+  }
   default:
     break;
   }
diff --git a/llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/dot.ll b/llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/dot.ll
new file mode 100644
index 0000000000000..b537b7bccf861
--- /dev/null
+++ b/llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/dot.ll
@@ -0,0 +1,56 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+
+; RUN: opt -passes=instsimplify -S < %s | FileCheck %s
+
+; Test that intrinsics wasm dot call are constant folded
+
+target triple = "wasm32-unknown-unknown"
+
+
+define <4 x i32> @dot_zero() {
+; CHECK-LABEL: define <4 x i32> @dot_zero() {
+; CHECK-NEXT:    ret <4 x i32> zeroinitializer
+;
+  %res = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> zeroinitializer, <8 x i16> zeroinitializer)
+  ret <4 x i32> %res
+}
+
+; a               =   1    2    3    4    5    6    7    8
+; b               =   1    2    3    4    5    6    7    8
+; k1|k2 = a * b   =   1    4    9   16   25   36   49   64
+; k1 + k2         =   (1+4) |  (9 + 16) | (25 + 36)  | (49 + 64)
+; result          =    5    |   25    |   61    |   113
+define <4 x i32> @dot_nonzero() {
+; CHECK-LABEL: define <4 x i32> @dot_nonzero() {
+; CHECK-NEXT:    ret <4 x i32> <i32 5, i32 25, i32 61, i32 113>
+;
+  %res = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> <i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7, i16 8>, <8 x i16> <i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7, i16 8>)
+  ret <4 x i32> %res
+}
+
+define <4 x i32> @dot_doubly_negative() {
+; CHECK-LABEL: define <4 x i32> @dot_doubly_negative() {
+; CHECK-NEXT:    ret <4 x i32> splat (i32 2)
+;
+  %res = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>, <8 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>)
+  ret <4 x i32> %res
+}
+
+; Tests that i16 max signed values fit in i32
+define <4 x i32> @dot_follow_modulo_spec_1() {
+; CHECK-LABEL: define <4 x i32> @dot_follow_modulo_spec_1() {
+; CHECK-NEXT:    ret <4 x i32> <i32 2147352578, i32 0, i32 0, i32 0>
+;
+  %res = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> <i16 32767, i16 32767, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, <8 x i16> <i16 32767, i16 32767, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>)
+  ret <4 x i32> %res
+}
+
+; Tests that i16 min signed values fit in i32
+define <4 x i32> @dot_follow_modulo_spec_2() {
+; CHECK-LABEL: define <4 x i32> @dot_follow_modulo_spec_2() {
+; CHECK-NEXT:    ret <4 x i32> <i32 -2147483648, i32 0, i32 0, i32 0>
+;
+  %res = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> <i16 -32768, i16 -32768, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, <8 x i16> <i16 -32768, i16 -32768, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>)
+  ret <4 x i32> %res
+}
+

lukel97

LGTM. Just want to check that you were able to run the tests again with UB sanitizer enabled and it didn't trigger?

llvm/lib/Analysis/ConstantFolding.cpp

badumbatish · 2025-08-14T17:33:12Z

Just want to check that you were able to run the tests again with UB sanitizer enabled and it didn't trigger?

yep no more trigger, sanitizer checks all of WebAssembly in

./build/bin/llvm-lit llvm/test/CodeGen/WebAssembly -v

badumbatish added 8 commits July 18, 2025 16:36

[WebAssembly] Precommit test for constant folding dot

5563f46

[WebAssembly] Constant fold dot operation

50ca839

Merge branch 'main' into wasm_fold_dot

3287cd6

Addresses specs questions and added test to reflect

fa8c096

Addresses nit, added negative test case

10d7dc9

Address nits and performance issues

0271d8d

Addresses PR reviews

67492fa

UB-proof wrap around of adding constant

a151fdc

badumbatish requested review from lukel97 and tlively August 11, 2025 19:08

llvmbot added llvm:instcombine Covers the InstCombine, InstSimplify and AggressiveInstCombine passes llvm:analysis Includes value tracking, cost tables and constant folding llvm:transforms labels Aug 11, 2025

lukel97 approved these changes Aug 14, 2025

View reviewed changes

llvm/lib/Analysis/ConstantFolding.cpp Outdated Show resolved Hide resolved

lukel97 changed the title ~~[WebAssembly] Reapply #149619~~ Reapply "[WebAssembly] Constant fold wasm.dot" Aug 14, 2025

tlively approved these changes Aug 14, 2025

View reviewed changes

Clarify casting of MulVector[...]

751de4c

badumbatish merged commit 10d9e7b into llvm:main Aug 15, 2025
9 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Reapply "[WebAssembly] Constant fold wasm.dot" #153070

Reapply "[WebAssembly] Constant fold wasm.dot" #153070

badumbatish commented Aug 11, 2025 •

edited

Loading

Uh oh!

llvmbot commented Aug 11, 2025

Uh oh!

llvmbot commented Aug 11, 2025

Uh oh!

lukel97 left a comment

Uh oh!

Uh oh!

badumbatish commented Aug 14, 2025 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Reapply "[WebAssembly] Constant fold wasm.dot" #153070

Reapply "[WebAssembly] Constant fold wasm.dot" #153070

Conversation

badumbatish commented Aug 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Aug 11, 2025

Uh oh!

llvmbot commented Aug 11, 2025

Uh oh!

lukel97 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

badumbatish commented Aug 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

badumbatish commented Aug 11, 2025 •

edited

Loading

badumbatish commented Aug 14, 2025 •

edited

Loading