[WebAssembly] Constant fold dot operation

Author: badumbatish

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 +3827,36 @@ static Constant *ConstantFoldFixedVectorCall(
     }
     return ConstantVector::get(Result);
   }
+  case Intrinsic::wasm_dot: {
+    unsigned NumElements =
+        cast<FixedVectorType>(Operands[0]->getType())->getNumElements();
+
+    assert(NumElements == 8 && NumElements / 2 == Result.size() &&
+           "wasm dot takes i16x8 and produce i32x4");
+    assert(Ty->isIntegerTy());
+    SmallVector<APInt, 8> MulVector;
+
+    for (unsigned I = 0; I < NumElements; ++I) {
+      ConstantInt *Elt0 =
+          cast<ConstantInt>(Operands[0]->getAggregateElement(I));
+      ConstantInt *Elt1 =
+          cast<ConstantInt>(Operands[1]->getAggregateElement(I));
+
+      // sext 32 first, according to specs
+      APInt IMul = Elt0->getValue().sext(32) * Elt1->getValue().sext(32);
+
+      // TODO: imul in specs includes a modulo operation
+      // Is this performed automatically via trunc = true in APInt creation of *
+      MulVector.push_back(IMul);
+    }
+    for (unsigned I = 0; I < Result.size(); ++I) {
+      // Same case as with imul
+      APInt IAdd = MulVector[I] + MulVector[I + Result.size()];
+      Result[I] = ConstantInt::get(Ty, IAdd);
+    }
+
+    return ConstantVector::get(Result);
+  }
   default:
     break;
   }

llvm/test/Transforms/InstSimplify/ConstProp/WebAssembly/dot.ll

@@ -9,26 +9,28 @@ target triple = "wasm32-unknown-unknown"
 
 define <4 x i32> @dot_zero() {
 ; CHECK-LABEL: define <4 x i32> @dot_zero() {
-; CHECK-NEXT:    [[RES:%.*]] = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> zeroinitializer, <8 x i16> zeroinitializer)
-; CHECK-NEXT:    ret <4 x i32> [[RES]]
+; 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+25) |  (4+36) | (9+49)  | (16+64)
+; result          =    26    |   40    |   58    |   80
 define <4 x i32> @dot_nonzero() {
 ; CHECK-LABEL: define <4 x i32> @dot_nonzero() {
-; CHECK-NEXT:    [[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>)
-; CHECK-NEXT:    ret <4 x i32> [[RES]]
+; CHECK-NEXT:    ret <4 x i32> <i32 26, i32 40, i32 58, i32 80>
 ;
   %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:    [[RES:%.*]] = tail call <4 x i32> @llvm.wasm.dot(<8 x i16> splat (i16 -1), <8 x i16> splat (i16 -1))
-; CHECK-NEXT:    ret <4 x i32> [[RES]]
+; 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