[InstCombine][X86] Add DemandedElts support for PMULDQ/PMULUDQ instructions

[RKSimon]

Summary

PMULDQ/PMULUDQ vXi64 instructions only use the even numbered v2Xi32 input elements which SimplifyDemandedVectorElts should try and use.

We can't do much with the undef demanded elts - we should probably only support the (mul X, undef -> 0) pattern the same as regular integer multiplies. I can add support for it if you guys want but I can't see it being used by real world code. Same with constant folding support.

[craig.topper]

lib/Transforms/InstCombine/InstCombineCalls.cpp

1999

Can we just call SimplifyDemandedVectorElts on the Intrinsic node here with an all 1s demanded mask. This allows reuse of the code in InstCombineSimplifyDemanded. See the code at line 1879 in this file in this diff.

lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

1459

Why is this code formatted differently than the same code for ArgOperand 0?

[RKSimon]

Updated based on Craig's feedback

[RKSimon]

lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

1459

Because for some reason I used clang-format on one and kept to the formatting for the rest of the file for the other.....

[craig.topper]

LGTM

lib/Transforms/InstCombine/InstCombineCalls.cpp

Instruction *InstCombiner::visitCallInst(CallInst &CI) {
   case Intrinsic::x86_avx512_psrlv_q_512:
   case Intrinsic::x86_avx512_psrlv_w_128:
   case Intrinsic::x86_avx512_psrlv_w_256:
   case Intrinsic::x86_avx512_psrlv_w_512:
     if (Value *V = simplifyX86varShift(*II, *Builder))
       return replaceInstUsesWith(*II, V);
     break;
 
+  case Intrinsic::x86_sse2_pmulu_dq:
+  case Intrinsic::x86_sse41_pmuldq:
+  case Intrinsic::x86_avx2_pmul_dq:
+  case Intrinsic::x86_avx2_pmulu_dq: {
+    unsigned VWidth = II->getType()->getVectorNumElements();
+    APInt UndefElts(VWidth, 0);
+    APInt DemandedElts = APInt::getAllOnesValue(VWidth);
+    if (Value *V = SimplifyDemandedVectorElts(II, DemandedElts, UndefElts)) {
+      if (V != II)
+        return replaceInstUsesWith(*II, V);
+      return II;
+    }
+    break;
+  }
+
   case Intrinsic::x86_sse41_insertps:
     if (Value *V = simplifyX86insertps(*II, *Builder))
       return replaceInstUsesWith(*II, V);
     break;
 
   case Intrinsic::x86_sse4a_extrq: {
     Value *Op0 = II->getArgOperand(0);
     Value *Op1 = II->getArgOperand(1);

lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

case Intrinsic::x86_avx512_mask3_vfnmsub_sd:
 
       // Lower element is undefined if all three lower elements are undefined.
       // Consider things like undef&0.  The result is known zero, not undef.
       if (!UndefElts2[0] || !UndefElts3[0])
         UndefElts.clearBit(0);
 
       break;
 
+    case Intrinsic::x86_sse2_pmulu_dq:
+    case Intrinsic::x86_sse41_pmuldq:
+    case Intrinsic::x86_avx2_pmul_dq:
+    case Intrinsic::x86_avx2_pmulu_dq: {
+      Value *Op0 = II->getArgOperand(0);
+      Value *Op1 = II->getArgOperand(1);
+      unsigned InnerVWidth = Op0->getType()->getVectorNumElements();
+      assert((VWidth * 2) == InnerVWidth && "Unexpected input size");
+
+      APInt InnerDemandedElts(InnerVWidth, 0);
+      for (unsigned i = 0; i != VWidth; ++i)
+        if (DemandedElts[i])
+          InnerDemandedElts.setBit(i * 2);
+
+      UndefElts2 = APInt(InnerVWidth, 0);
+      TmpV = SimplifyDemandedVectorElts(Op0, InnerDemandedElts, UndefElts2,
+                                        Depth + 1);
+      if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
+
+      UndefElts3 = APInt(InnerVWidth, 0);
+      TmpV = SimplifyDemandedVectorElts(Op1, InnerDemandedElts, UndefElts3,
+                                        Depth + 1);
+      if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
+
+      break;
+    }
+
     // SSE4A instructions leave the upper 64-bits of the 128-bit result
     // in an undefined state.
     case Intrinsic::x86_sse4a_extrq:
     case Intrinsic::x86_sse4a_extrqi:
     case Intrinsic::x86_sse4a_insertq:
     case Intrinsic::x86_sse4a_insertqi:
       UndefElts |= APInt::getHighBitsSet(VWidth, VWidth / 2);
       break;
     }
     break;
   }
   }
   return MadeChange ? I : nullptr;
 }

test/Transforms/InstCombine/x86-muldq.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -instcombine -S | FileCheck %s
 
 ;
 ; PMULUDQ/PMULDQ - only the even elements (0, 2, 4, 6) of the vXi32 inputs are required.
 ;
 
 define <2 x i64> @test_demanded_elts_pmuludq_128(<4 x i32> %a0, <4 x i32> %a1) {
 ; CHECK-LABEL: @test_demanded_elts_pmuludq_128(
-; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x i32> %a0, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 2, i32 2>
-; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <4 x i32> %a1, <4 x i32> undef, <4 x i32> <i32 1, i32 1, i32 3, i32 3>
-; CHECK-NEXT:    [[TMP3:%.*]] = call <2 x i64> @llvm.x86.sse2.pmulu.dq(<4 x i32> [[TMP1]], <4 x i32> [[TMP2]])
-; CHECK-NEXT:    [[TMP4:%.*]] = shufflevector <2 x i64> [[TMP3]], <2 x i64> undef, <2 x i32> zeroinitializer
-; CHECK-NEXT:    ret <2 x i64> [[TMP4]]
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x i32> %a1, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
+; CHECK-NEXT:    [[TMP2:%.*]] = call <2 x i64> @llvm.x86.sse2.pmulu.dq(<4 x i32> %a0, <4 x i32> [[TMP1]])
+; CHECK-NEXT:    [[TMP3:%.*]] = shufflevector <2 x i64> [[TMP2]], <2 x i64> undef, <2 x i32> zeroinitializer
+; CHECK-NEXT:    ret <2 x i64> [[TMP3]]
 ;
   %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 2, i32 2>
   %2 = shufflevector <4 x i32> %a1, <4 x i32> undef, <4 x i32> <i32 1, i32 1, i32 3, i32 3>
   %3 = call <2 x i64> @llvm.x86.sse2.pmulu.dq(<4 x i32> %1, <4 x i32> %2)
   %4 = shufflevector <2 x i64> %3, <2 x i64> undef, <2 x i32> zeroinitializer
   ret <2 x i64> %4
 }
 
 define <4 x i64> @test_demanded_elts_pmuludq_256(<8 x i32> %a0, <8 x i32> %a1) {
 ; CHECK-LABEL: @test_demanded_elts_pmuludq_256(
-; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <8 x i32> %a0, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 2, i32 2, i32 4, i32 4, i32 6, i32 6>
-; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <8 x i32> %a1, <8 x i32> undef, <8 x i32> <i32 1, i32 1, i32 3, i32 3, i32 5, i32 5, i32 7, i32 7>
-; CHECK-NEXT:    [[TMP3:%.*]] = call <4 x i64> @llvm.x86.avx2.pmulu.dq(<8 x i32> [[TMP1]], <8 x i32> [[TMP2]])
-; CHECK-NEXT:    ret <4 x i64> [[TMP3]]
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <8 x i32> %a1, <8 x i32> undef, <8 x i32> <i32 1, i32 undef, i32 3, i32 undef, i32 5, i32 undef, i32 7, i32 undef>
+; CHECK-NEXT:    [[TMP2:%.*]] = call <4 x i64> @llvm.x86.avx2.pmulu.dq(<8 x i32> %a0, <8 x i32> [[TMP1]])
+; CHECK-NEXT:    ret <4 x i64> [[TMP2]]
 ;
   %1 = shufflevector <8 x i32> %a0, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 2, i32 2, i32 4, i32 4, i32 6, i32 6>
   %2 = shufflevector <8 x i32> %a1, <8 x i32> undef, <8 x i32> <i32 1, i32 1, i32 3, i32 3, i32 5, i32 5, i32 7, i32 7>
   %3 = call <4 x i64> @llvm.x86.avx2.pmulu.dq(<8 x i32> %1, <8 x i32> %2)
   ret <4 x i64> %3
 }
 
 define <2 x i64> @test_demanded_elts_pmuldq_128(<4 x i32> %a0, <4 x i32> %a1) {
 ; CHECK-LABEL: @test_demanded_elts_pmuldq_128(
-; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x i32> %a0, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 2, i32 2>
-; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <4 x i32> %a1, <4 x i32> undef, <4 x i32> <i32 1, i32 1, i32 3, i32 3>
-; CHECK-NEXT:    [[TMP3:%.*]] = call <2 x i64> @llvm.x86.sse41.pmuldq(<4 x i32> [[TMP1]], <4 x i32> [[TMP2]])
-; CHECK-NEXT:    ret <2 x i64> [[TMP3]]
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x i32> %a1, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 3, i32 undef>
+; CHECK-NEXT:    [[TMP2:%.*]] = call <2 x i64> @llvm.x86.sse41.pmuldq(<4 x i32> %a0, <4 x i32> [[TMP1]])
+; CHECK-NEXT:    ret <2 x i64> [[TMP2]]
 ;
   %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 2, i32 2>
   %2 = shufflevector <4 x i32> %a1, <4 x i32> undef, <4 x i32> <i32 1, i32 1, i32 3, i32 3>
   %3 = call <2 x i64> @llvm.x86.sse41.pmuldq(<4 x i32> %1, <4 x i32> %2)
   ret <2 x i64> %3
 }
 
 define <4 x i64> @test_demanded_elts_pmuluq_256(<8 x i32> %a0, <8 x i32> %a1) {
 ; CHECK-LABEL: @test_demanded_elts_pmuluq_256(
-; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <8 x i32> %a0, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 2, i32 2, i32 4, i32 4, i32 6, i32 6>
-; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <8 x i32> %a1, <8 x i32> undef, <8 x i32> <i32 1, i32 1, i32 3, i32 3, i32 5, i32 5, i32 7, i32 7>
-; CHECK-NEXT:    [[TMP3:%.*]] = call <4 x i64> @llvm.x86.avx2.pmul.dq(<8 x i32> [[TMP1]], <8 x i32> [[TMP2]])
-; CHECK-NEXT:    [[TMP4:%.*]] = shufflevector <4 x i64> [[TMP3]], <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 3, i32 3>
-; CHECK-NEXT:    ret <4 x i64> [[TMP4]]
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <8 x i32> %a1, <8 x i32> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 7, i32 undef>
+; CHECK-NEXT:    [[TMP2:%.*]] = call <4 x i64> @llvm.x86.avx2.pmul.dq(<8 x i32> %a0, <8 x i32> [[TMP1]])
+; CHECK-NEXT:    [[TMP3:%.*]] = shufflevector <4 x i64> [[TMP2]], <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 3, i32 3>
+; CHECK-NEXT:    ret <4 x i64> [[TMP3]]
 ;
   %1 = shufflevector <8 x i32> %a0, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 2, i32 2, i32 4, i32 4, i32 6, i32 6>
   %2 = shufflevector <8 x i32> %a1, <8 x i32> undef, <8 x i32> <i32 1, i32 1, i32 3, i32 3, i32 5, i32 5, i32 7, i32 7>
   %3 = call <4 x i64> @llvm.x86.avx2.pmul.dq(<8 x i32> %1, <8 x i32> %2)
   %4 = shufflevector <4 x i64> %3, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 3, i32 3>
   ret <4 x i64> %4
 }
 
 declare <2 x i64> @llvm.x86.sse2.pmulu.dq(<4 x i32>, <4 x i32>) nounwind readnone
 declare <2 x i64> @llvm.x86.sse41.pmuldq(<4 x i32>, <4 x i32>) nounwind readnone
 
 declare <4 x i64> @llvm.x86.avx2.pmulu.dq(<8 x i32>, <8 x i32>) nounwind readnone
 declare <4 x i64> @llvm.x86.avx2.pmul.dq(<8 x i32>, <8 x i32>) nounwind readnone