merge add and fadd

Author: farzonl

llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp

@@ -73,35 +73,38 @@ static bool isIntrinsicExpansion(Function &F) {
   }
   return false;
 }
+static Value *expandVecReduceAdd(CallInst *Orig, Intrinsic::ID IntrinsicId) {
+  assert(IntrinsicId == Intrinsic::vector_reduce_add ||
+         IntrinsicId == Intrinsic::vector_reduce_fadd);
 
-static Value *expandVecReduceFAdd(CallInst *Orig) {
-  // Note: vector_reduce_fadd first argument is a starting value
-  // Our use doesn't need it, so ignoring argument zero.
-  Value *X = Orig->getOperand(1);
   IRBuilder<> Builder(Orig);
+  bool IsFAdd = (IntrinsicId == Intrinsic::vector_reduce_fadd);
+
+  // Define the addition operation based on the intrinsic ID.
+  auto AddOp = [&Builder, IsFAdd](Value *Sum, Value *Elt) {
+    return IsFAdd ? Builder.CreateFAdd(Sum, Elt) : Builder.CreateAdd(Sum, Elt);
+  };
+
+  Value *X = Orig->getOperand(IsFAdd ? 1 : 0);
   Type *Ty = X->getType();
   auto *XVec = dyn_cast<FixedVectorType>(Ty);
   unsigned XVecSize = XVec->getNumElements();
   Value *Sum = Builder.CreateExtractElement(X, static_cast<uint64_t>(0));
-  for (unsigned I = 1; I < XVecSize; I++) {
-    Value *Elt = Builder.CreateExtractElement(X, I);
-    Sum = Builder.CreateFAdd(Sum, Elt);
-  }
-  return Sum;
-}
 
-static Value *expandVecReduceAdd(CallInst *Orig) {
-  Value *X = Orig->getOperand(0);
-  IRBuilder<> Builder(Orig);
-  Type *Ty = X->getType();
-  auto *XVec = dyn_cast<FixedVectorType>(Ty);
-  unsigned XVecSize = XVec->getNumElements();
+  // Handle the initial start value for floating-point addition.
+  if (IsFAdd) {
+    llvm::Constant *StartValue =
+        llvm::dyn_cast<llvm::Constant>(Orig->getOperand(0));
+    if (StartValue && !StartValue->isZeroValue())
+      Sum = Builder.CreateFAdd(Sum, StartValue);
+  }
 
-  Value *Sum = Builder.CreateExtractElement(X, static_cast<uint64_t>(0));
+  // Accumulate the remaining vector elements.
   for (unsigned I = 1; I < XVecSize; I++) {
     Value *Elt = Builder.CreateExtractElement(X, I);
-    Sum = Builder.CreateAdd(Sum, Elt);
+    Sum = AddOp(Sum, Elt);
   }
+
   return Sum;
 }
 
@@ -614,10 +617,8 @@ static bool expandIntrinsic(Function &F, CallInst *Orig) {
     Result = expandRadiansIntrinsic(Orig);
     break;
   case Intrinsic::vector_reduce_add:
-    Result = expandVecReduceAdd(Orig);
-    break;
   case Intrinsic::vector_reduce_fadd:
-    Result = expandVecReduceFAdd(Orig);
+    Result = expandVecReduceAdd(Orig, IntrinsicId);
     break;
   }
   if (Result) {

llvm/test/CodeGen/DirectX/vector_reduce_add.ll

@@ -17,6 +17,21 @@ entry:
   ret half %rdx.fadd
 }
 
+define noundef half @test_length_half2_start1(<2 x half> noundef %p0) {
+; CHECK-LABEL: define noundef half @test_length_half2_start1(
+; CHECK-SAME: <2 x half> noundef [[P0:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP0:%.*]] = extractelement <2 x half> [[P0]], i64 0
+; CHECK-NEXT:    [[TMP1:%.*]] = fadd half [[TMP0]], 0xH0001
+; CHECK-NEXT:    [[TMP2:%.*]] = extractelement <2 x half> [[P0]], i64 1
+; CHECK-NEXT:    [[TMP3:%.*]] = fadd half [[TMP1]], [[TMP2]]
+; CHECK-NEXT:    ret half [[TMP3]]
+;
+entry:
+  %rdx.fadd = call half @llvm.vector.reduce.fadd.v2f16(half 0xH0001, <2 x half> %p0)
+  ret half %rdx.fadd
+}
+
 define noundef half @test_length_half3(<3 x half> noundef %p0) {
 ; CHECK-LABEL: define noundef half @test_length_half3(
 ; CHECK-SAME: <3 x half> noundef [[P0:%.*]]) {
@@ -81,6 +96,23 @@ entry:
   ret float %rdx.fadd
 }
 
+define noundef float @test_length_float3_start1(<3 x float> noundef %p0) {
+; CHECK-LABEL: define noundef float @test_length_float3_start1(
+; CHECK-SAME: <3 x float> noundef [[P0:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP0:%.*]] = extractelement <3 x float> [[P0]], i64 0
+; CHECK-NEXT:    [[TMP1:%.*]] = fadd float [[TMP0]], 1.000000e+00
+; CHECK-NEXT:    [[TMP2:%.*]] = extractelement <3 x float> [[P0]], i64 1
+; CHECK-NEXT:    [[TMP3:%.*]] = fadd float [[TMP1]], [[TMP2]]
+; CHECK-NEXT:    [[TMP4:%.*]] = extractelement <3 x float> [[P0]], i64 2
+; CHECK-NEXT:    [[TMP5:%.*]] = fadd float [[TMP3]], [[TMP4]]
+; CHECK-NEXT:    ret float [[TMP5]]
+;
+entry:
+  %rdx.fadd = call float @llvm.vector.reduce.fadd.v3f32(float 1.000000e+00, <3 x float> %p0)
+  ret float %rdx.fadd
+}
+
 define noundef float @test_length_float4(<4 x float> noundef %p0) {
 ; CHECK-LABEL: define noundef float @test_length_float4(
 ; CHECK-SAME: <4 x float> noundef [[P0:%.*]]) {
@@ -147,6 +179,25 @@ entry:
   ret double %rdx.fadd
 }
 
+define noundef double @test_length_double4_start1(<4 x double> noundef %p0) {
+; CHECK-LABEL: define noundef double @test_length_double4_start1(
+; CHECK-SAME: <4 x double> noundef [[P0:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP0:%.*]] = extractelement <4 x double> [[P0]], i64 0
+; CHECK-NEXT:    [[TMP1:%.*]] = fadd double [[TMP0]], 1.000000e+00
+; CHECK-NEXT:    [[TMP2:%.*]] = extractelement <4 x double> [[P0]], i64 1
+; CHECK-NEXT:    [[TMP3:%.*]] = fadd double [[TMP1]], [[TMP2]]
+; CHECK-NEXT:    [[TMP4:%.*]] = extractelement <4 x double> [[P0]], i64 2
+; CHECK-NEXT:    [[TMP5:%.*]] = fadd double [[TMP3]], [[TMP4]]
+; CHECK-NEXT:    [[TMP6:%.*]] = extractelement <4 x double> [[P0]], i64 3
+; CHECK-NEXT:    [[TMP7:%.*]] = fadd double [[TMP5]], [[TMP6]]
+; CHECK-NEXT:    ret double [[TMP7]]
+;
+entry:
+  %rdx.fadd = call double @llvm.vector.reduce.fadd.v4f64(double 1.000000e+00, <4 x double> %p0)
+  ret double %rdx.fadd
+}
+
 define noundef i16 @test_length_short2(<2 x i16> noundef %p0) {
 ; CHECK-LABEL: define noundef i16 @test_length_short2(
 ; CHECK-SAME: <2 x i16> noundef [[P0:%.*]]) {