[InstCombine][SSE] Add MOVMSK constant folding (PR27982)

[RKSimon]

Summary

This patch adds support for folding undef/zero/constant inputs to MOVMSK instructions.

The SSE/AVX versions can be fully folded, but the MMX version can only handle undef inputs.

[ab]

LGTM, with one nit.

lib/Transforms/InstCombine/InstCombineCalls.cpp

348

i -> I ?
Or, since I is confusing: EI/EE? AI/AE?

lib/Transforms/InstCombine/InstCombineCalls.cpp

if (ShiftLeft)
     return Builder.CreateShl(Vec, ShiftVec);
 
   if (LogicalShift)
     return Builder.CreateLShr(Vec, ShiftVec);
 
   return Builder.CreateAShr(Vec, ShiftVec);
 }
 
+static Value *simplifyX86movmsk(const IntrinsicInst &II,
+                                InstCombiner::BuilderTy &Builder) {
+  Value *Arg = II.getArgOperand(0);
+  Type *ResTy = II.getType();
+  Type *ArgTy = Arg->getType();
+
+  // movmsk(undef) -> zero as we must ensure the upper bits are zero.
+  if (isa<UndefValue>(Arg))
+    return Constant::getNullValue(ResTy);
+
+  // We can't easily peek through x86_mmx types.
+  if (!ArgTy->isVectorTy())
+    return nullptr;
+
+  auto *C = dyn_cast<Constant>(Arg);
+  if (!C)
+    return nullptr;
+
+  // Extract signbits of the vector input and pack into integer result.
+  APInt Result(ResTy->getPrimitiveSizeInBits(), 0);
+  for (unsigned i = 0, E = ArgTy->getVectorNumElements(); i != E; ++i) {
+    auto *COp = C->getAggregateElement(i);
+    if (!COp)
+      return nullptr;
+    if (isa<UndefValue>(COp))
+      continue;
+
+    auto *CInt = dyn_cast<ConstantInt>(COp);
+    auto *CFp = dyn_cast<ConstantFP>(COp);
+    if (!CInt && !CFp)
+      return nullptr;
+
+    if ((CInt && CInt->isNegative()) || (CFp && CFp->isNegative()))
+      Result.setBit(i);
+  }
+
+  return Constant::getIntegerValue(ResTy, Result);
+}
+
 static Value *simplifyX86insertps(const IntrinsicInst &II,
                                   InstCombiner::BuilderTy &Builder) {
   auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2));
   if (!CInt)
     return nullptr;
 
   VectorType *VecTy = cast<VectorType>(II.getType());
   assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
case Intrinsic::x86_sse2_cvttsd2si64: {
     unsigned VWidth = Arg->getType()->getVectorNumElements();
     if (Value *V = SimplifyDemandedVectorEltsLow(Arg, VWidth, 1)) {
       II->setArgOperand(0, V);
       return II;
     }
     break;
   }
 
+  case Intrinsic::x86_mmx_pmovmskb:
+  case Intrinsic::x86_sse_movmsk_ps:
+  case Intrinsic::x86_sse2_movmsk_pd:
+  case Intrinsic::x86_sse2_pmovmskb_128:
+  case Intrinsic::x86_avx_movmsk_pd_256:
+  case Intrinsic::x86_avx_movmsk_ps_256:
+  case Intrinsic::x86_avx2_pmovmskb: {
+    if (Value *V = simplifyX86movmsk(*II, *Builder))
+      return replaceInstUsesWith(*II, V);
+    break;
+  }
+
   case Intrinsic::x86_sse_comieq_ss:
   case Intrinsic::x86_sse_comige_ss:
   case Intrinsic::x86_sse_comigt_ss:
   case Intrinsic::x86_sse_comile_ss:
   case Intrinsic::x86_sse_comilt_ss:
   case Intrinsic::x86_sse_comineq_ss:
   case Intrinsic::x86_sse_ucomieq_ss:
   case Intrinsic::x86_sse_ucomige_ss:

test/Transforms/InstCombine/x86-movmsk.ll

 ;
   %1 = call i32 @llvm.x86.avx.movmsk.pd.256(<4 x double> %a0)
   %2 = and i32 %1, -16
   ret i32 %2
 }
 
 ; llvm.x86.avx2.pmovmskb uses the whole of the 32-bit register.
 
+;
+; Constant Folding (UNDEF -> ZERO)
+;
+
+define i32 @undef_x86_mmx_pmovmskb() {
+; CHECK-LABEL: @undef_x86_mmx_pmovmskb(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.mmx.pmovmskb(x86_mmx undef)
+  ret i32 %1
+}
+
+define i32 @undef_x86_sse_movmsk_ps() {
+; CHECK-LABEL: @undef_x86_sse_movmsk_ps(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.sse.movmsk.ps(<4 x float> undef)
+  ret i32 %1
+}
+
+define i32 @undef_x86_sse2_movmsk_pd() {
+; CHECK-LABEL: @undef_x86_sse2_movmsk_pd(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.sse2.movmsk.pd(<2 x double> undef)
+  ret i32 %1
+}
+
+define i32 @undef_x86_sse2_pmovmskb_128() {
+; CHECK-LABEL: @undef_x86_sse2_pmovmskb_128(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.sse2.pmovmskb.128(<16 x i8> undef)
+  ret i32 %1
+}
+
+define i32 @undef_x86_avx_movmsk_ps_256() {
+; CHECK-LABEL: @undef_x86_avx_movmsk_ps_256(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.avx.movmsk.ps.256(<8 x float> undef)
+  ret i32 %1
+}
+
+define i32 @undef_x86_avx_movmsk_pd_256() {
+; CHECK-LABEL: @undef_x86_avx_movmsk_pd_256(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.avx.movmsk.pd.256(<4 x double> undef)
+  ret i32 %1
+}
+
+define i32 @undef_x86_avx2_pmovmskb() {
+; CHECK-LABEL: @undef_x86_avx2_pmovmskb(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.avx2.pmovmskb(<32 x i8> undef)
+  ret i32 %1
+}
+
+;
+; Constant Folding (ZERO -> ZERO)
+;
+
+define i32 @zero_x86_mmx_pmovmskb() {
+; CHECK-LABEL: @zero_x86_mmx_pmovmskb(
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.x86.mmx.pmovmskb(x86_mmx bitcast (<1 x i64> zeroinitializer to x86_mmx))
+; CHECK-NEXT:    ret i32 [[TMP1]]
+;
+  %1 = bitcast <1 x i64> zeroinitializer to x86_mmx
+  %2 = call i32 @llvm.x86.mmx.pmovmskb(x86_mmx %1)
+  ret i32 %2
+}
+
+define i32 @zero_x86_sse_movmsk_ps() {
+; CHECK-LABEL: @zero_x86_sse_movmsk_ps(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.sse.movmsk.ps(<4 x float> zeroinitializer)
+  ret i32 %1
+}
+
+define i32 @zero_x86_sse2_movmsk_pd() {
+; CHECK-LABEL: @zero_x86_sse2_movmsk_pd(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.sse2.movmsk.pd(<2 x double> zeroinitializer)
+  ret i32 %1
+}
+
+define i32 @zero_x86_sse2_pmovmskb_128() {
+; CHECK-LABEL: @zero_x86_sse2_pmovmskb_128(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.sse2.pmovmskb.128(<16 x i8> zeroinitializer)
+  ret i32 %1
+}
+
+define i32 @zero_x86_avx_movmsk_ps_256() {
+; CHECK-LABEL: @zero_x86_avx_movmsk_ps_256(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.avx.movmsk.ps.256(<8 x float> zeroinitializer)
+  ret i32 %1
+}
+
+define i32 @zero_x86_avx_movmsk_pd_256() {
+; CHECK-LABEL: @zero_x86_avx_movmsk_pd_256(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.avx.movmsk.pd.256(<4 x double> zeroinitializer)
+  ret i32 %1
+}
+
+define i32 @zero_x86_avx2_pmovmskb() {
+; CHECK-LABEL: @zero_x86_avx2_pmovmskb(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.avx2.pmovmskb(<32 x i8> zeroinitializer)
+  ret i32 %1
+}
+
+;
+; Constant Folding
+;
+
+define i32 @fold_x86_mmx_pmovmskb() {
+; CHECK-LABEL: @fold_x86_mmx_pmovmskb(
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.x86.mmx.pmovmskb(x86_mmx bitcast (<8 x i8> <i8 0, i8 -1, i8 -1, i8 127, i8 -127, i8 63, i8 64, i8 0> to x86_mmx))
+; CHECK-NEXT:    ret i32 [[TMP1]]
+;
+  %1 = bitcast <8 x i8> <i8 0, i8 255, i8 -1, i8 127, i8 -127, i8 63, i8 64, i8 256> to x86_mmx
+  %2 = call i32 @llvm.x86.mmx.pmovmskb(x86_mmx %1)
+  ret i32 %2
+}
+
+define i32 @fold_x86_sse_movmsk_ps() {
+; CHECK-LABEL: @fold_x86_sse_movmsk_ps(
+; CHECK-NEXT:    ret i32 10
+;
+  %1 = call i32 @llvm.x86.sse.movmsk.ps(<4 x float> <float 1.0, float -1.0, float 100.0, float -200.0>)
+  ret i32 %1
+}
+
+define i32 @fold_x86_sse2_movmsk_pd() {
+; CHECK-LABEL: @fold_x86_sse2_movmsk_pd(
+; CHECK-NEXT:    ret i32 2
+;
+  %1 = call i32 @llvm.x86.sse2.movmsk.pd(<2 x double> <double 1.0, double -1.0>)
+  ret i32 %1
+}
+
+define i32 @fold_x86_sse2_pmovmskb_128() {
+; CHECK-LABEL: @fold_x86_sse2_pmovmskb_128(
+; CHECK-NEXT:    ret i32 5654
+;
+  %1 = call i32 @llvm.x86.sse2.pmovmskb.128(<16 x i8> <i8 0, i8 255, i8 -1, i8 127, i8 -127, i8 63, i8 64, i8 256, i8 0, i8 255, i8 -1, i8 127, i8 -127, i8 63, i8 64, i8 256>)
+  ret i32 %1
+}
+
+define i32 @fold_x86_avx_movmsk_ps_256() {
+; CHECK-LABEL: @fold_x86_avx_movmsk_ps_256(
+; CHECK-NEXT:    ret i32 170
+;
+  %1 = call i32 @llvm.x86.avx.movmsk.ps.256(<8 x float> <float 1.0, float -1.0, float 100.0, float -200.0, float +0.0, float -0.0, float 100000.0, float -5000000.0>)
+  ret i32 %1
+}
+
+define i32 @fold_x86_avx_movmsk_pd_256() {
+; CHECK-LABEL: @fold_x86_avx_movmsk_pd_256(
+; CHECK-NEXT:    ret i32 10
+;
+  %1 = call i32 @llvm.x86.avx.movmsk.pd.256(<4 x double> <double 1.0, double -1.0, double 100.0, double -200.0>)
+  ret i32 %1
+}
+
+define i32 @fold_x86_avx2_pmovmskb() {
+; CHECK-LABEL: @fold_x86_avx2_pmovmskb(
+; CHECK-NEXT:    ret i32 370546176
+;
+  %1 = call i32 @llvm.x86.avx2.pmovmskb(<32 x i8> <i8 0, i8 undef, i8 undef, i8 undef, i8 undef, i8 undef, i8 undef, i8 undef, i8 0, i8 255, i8 -1, i8 127, i8 -127, i8 63, i8 64, i8 256, i8 0, i8 255, i8 -1, i8 127, i8 -127, i8 63, i8 64, i8 256, i8 0, i8 255, i8 -1, i8 127, i8 -127, i8 63, i8 64, i8 256>)
+  ret i32 %1
+}
+
 declare i32 @llvm.x86.mmx.pmovmskb(x86_mmx)
 
 declare i32 @llvm.x86.sse.movmsk.ps(<4 x float>)
 declare i32 @llvm.x86.sse2.movmsk.pd(<2 x double>)
 declare i32 @llvm.x86.sse2.pmovmskb.128(<16 x i8>)
 
 declare i32 @llvm.x86.avx.movmsk.ps.256(<8 x float>)
 declare i32 @llvm.x86.avx.movmsk.pd.256(<4 x double>)
 declare i32 @llvm.x86.avx2.pmovmskb(<32 x i8>)