[LLVM][InstCombine] Preserve vector types when shrinking FP constants. (llvm#163598)

While my objective is to make the shrinkfp path safe for ConstantFP
based splats I discovered the following issues also affect
ConstantVector based splats:

1. PreferBFloat is not set for bfloat vectors.
2. getMinimumFPType() returns a scalar type for vector constants where
getSplatValue() is successful.

Author: paulwalker-arm

llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp

@@ -1643,33 +1643,46 @@ Instruction *InstCombinerImpl::visitSExt(SExtInst &Sext) {
 
 /// Return a Constant* for the specified floating-point constant if it fits
 /// in the specified FP type without changing its value.
-static bool fitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
+static bool fitsInFPType(APFloat F, const fltSemantics &Sem) {
   bool losesInfo;
-  APFloat F = CFP->getValueAPF();
   (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
   return !losesInfo;
 }
 
-static Type *shrinkFPConstant(ConstantFP *CFP, bool PreferBFloat) {
-  if (CFP->getType() == Type::getPPC_FP128Ty(CFP->getContext()))
-    return nullptr;  // No constant folding of this.
+static Type *shrinkFPConstant(LLVMContext &Ctx, const APFloat &F,
+                              bool PreferBFloat) {
   // See if the value can be truncated to bfloat and then reextended.
-  if (PreferBFloat && fitsInFPType(CFP, APFloat::BFloat()))
-    return Type::getBFloatTy(CFP->getContext());
+  if (PreferBFloat && fitsInFPType(F, APFloat::BFloat()))
+    return Type::getBFloatTy(Ctx);
   // See if the value can be truncated to half and then reextended.
-  if (!PreferBFloat && fitsInFPType(CFP, APFloat::IEEEhalf()))
-    return Type::getHalfTy(CFP->getContext());
+  if (!PreferBFloat && fitsInFPType(F, APFloat::IEEEhalf()))
+    return Type::getHalfTy(Ctx);
   // See if the value can be truncated to float and then reextended.
-  if (fitsInFPType(CFP, APFloat::IEEEsingle()))
-    return Type::getFloatTy(CFP->getContext());
-  if (CFP->getType()->isDoubleTy())
-    return nullptr;  // Won't shrink.
-  if (fitsInFPType(CFP, APFloat::IEEEdouble()))
-    return Type::getDoubleTy(CFP->getContext());
+  if (fitsInFPType(F, APFloat::IEEEsingle()))
+    return Type::getFloatTy(Ctx);
+  if (&F.getSemantics() == &APFloat::IEEEdouble())
+    return nullptr; // Won't shrink.
+  // See if the value can be truncated to double and then reextended.
+  if (fitsInFPType(F, APFloat::IEEEdouble()))
+    return Type::getDoubleTy(Ctx);
   // Don't try to shrink to various long double types.
   return nullptr;
 }
 
+static Type *shrinkFPConstant(ConstantFP *CFP, bool PreferBFloat) {
+  Type *Ty = CFP->getType();
+  if (Ty->getScalarType()->isPPC_FP128Ty())
+    return nullptr; // No constant folding of this.
+
+  Type *ShrinkTy =
+      shrinkFPConstant(CFP->getContext(), CFP->getValueAPF(), PreferBFloat);
+  if (ShrinkTy)
+    if (auto *VecTy = dyn_cast<VectorType>(Ty))
+      ShrinkTy = VectorType::get(ShrinkTy, VecTy);
+
+  return ShrinkTy;
+}
+
 // Determine if this is a vector of ConstantFPs and if so, return the minimal
 // type we can safely truncate all elements to.
 static Type *shrinkFPConstantVector(Value *V, bool PreferBFloat) {
@@ -1720,10 +1733,10 @@ static Type *getMinimumFPType(Value *V, bool PreferBFloat) {
 
   // Try to shrink scalable and fixed splat vectors.
   if (auto *FPC = dyn_cast<Constant>(V))
-    if (isa<VectorType>(V->getType()))
+    if (auto *VTy = dyn_cast<VectorType>(V->getType()))
       if (auto *Splat = dyn_cast_or_null<ConstantFP>(FPC->getSplatValue()))
         if (Type *T = shrinkFPConstant(Splat, PreferBFloat))
-          return T;
+          return VectorType::get(T, VTy);
 
   // Try to shrink a vector of FP constants. This returns nullptr on scalable
   // vectors
@@ -1796,10 +1809,9 @@ Instruction *InstCombinerImpl::visitFPTrunc(FPTruncInst &FPT) {
   Type *Ty = FPT.getType();
   auto *BO = dyn_cast<BinaryOperator>(FPT.getOperand(0));
   if (BO && BO->hasOneUse()) {
-    Type *LHSMinType =
-        getMinimumFPType(BO->getOperand(0), /*PreferBFloat=*/Ty->isBFloatTy());
-    Type *RHSMinType =
-        getMinimumFPType(BO->getOperand(1), /*PreferBFloat=*/Ty->isBFloatTy());
+    bool PreferBFloat = Ty->getScalarType()->isBFloatTy();
+    Type *LHSMinType = getMinimumFPType(BO->getOperand(0), PreferBFloat);
+    Type *RHSMinType = getMinimumFPType(BO->getOperand(1), PreferBFloat);
     unsigned OpWidth = BO->getType()->getFPMantissaWidth();
     unsigned LHSWidth = LHSMinType->getFPMantissaWidth();
     unsigned RHSWidth = RHSMinType->getFPMantissaWidth();

llvm/test/Transforms/InstCombine/fpextend.ll

@@ -1,5 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+; RUN: opt < %s -passes=instcombine -use-constant-fp-for-fixed-length-splat -S | FileCheck %s
 
 define float @test(float %x) nounwind  {
 ; CHECK-LABEL: @test(
@@ -449,6 +450,28 @@ define bfloat @bf16_frem(bfloat %x) {
   ret bfloat %t3
 }
 
+define <4 x bfloat> @v4bf16_frem_x_const(<4 x bfloat> %x) {
+; CHECK-LABEL: @v4bf16_frem_x_const(
+; CHECK-NEXT:    [[TMP1:%.*]] = frem <4 x bfloat> [[X:%.*]], splat (bfloat 0xR40C9)
+; CHECK-NEXT:    ret <4 x bfloat> [[TMP1]]
+;
+  %t1 = fpext <4 x bfloat> %x to <4 x float>
+  %t2 = frem <4 x float> %t1, splat(float 6.281250e+00)
+  %t3 = fptrunc <4 x float> %t2 to <4 x bfloat>
+  ret <4 x bfloat> %t3
+}
+
+define <4 x bfloat> @v4bf16_frem_const_x(<4 x bfloat> %x) {
+; CHECK-LABEL: @v4bf16_frem_const_x(
+; CHECK-NEXT:    [[TMP1:%.*]] = frem <4 x bfloat> splat (bfloat 0xR40C9), [[X:%.*]]
+; CHECK-NEXT:    ret <4 x bfloat> [[TMP1]]
+;
+  %t1 = fpext <4 x bfloat> %x to <4 x float>
+  %t2 = frem <4 x float> splat(float 6.281250e+00), %t1
+  %t3 = fptrunc <4 x float> %t2 to <4 x bfloat>
+  ret <4 x bfloat> %t3
+}
+
 define <4 x float> @v4f32_fadd(<4 x float> %a) {
 ; CHECK-LABEL: @v4f32_fadd(
 ; CHECK-NEXT:    [[TMP1:%.*]] = fadd <4 x float> [[A:%.*]], splat (float -1.000000e+00)
@@ -459,3 +482,16 @@ define <4 x float> @v4f32_fadd(<4 x float> %a) {
   %5 = fptrunc <4 x double> %4 to <4 x float>
   ret <4 x float> %5
 }
+
+define <4 x float> @v4f32_fadd_const_not_shrinkable(<4 x float> %a) {
+; CHECK-LABEL: @v4f32_fadd_const_not_shrinkable(
+; CHECK-NEXT:    [[TMP1:%.*]] = fpext <4 x float> [[A:%.*]] to <4 x double>
+; CHECK-NEXT:    [[TMP2:%.*]] = fadd <4 x double> [[TMP1]], splat (double -1.000000e+100)
+; CHECK-NEXT:    [[TMP3:%.*]] = fptrunc <4 x double> [[TMP2]] to <4 x float>
+; CHECK-NEXT:    ret <4 x float> [[TMP3]]
+;
+  %2 = fpext <4 x float> %a to <4 x double>
+  %4 = fadd <4 x double> %2, splat (double -1.000000e+100)
+  %5 = fptrunc <4 x double> %4 to <4 x float>
+  ret <4 x float> %5
+}