[InstCombine][SSE] Add DemandedElts support for PSHUFB instructions

[RKSimon]

Summary

Simplify a pshufb shuffle mask based on the elements of the mask that are actually demanded.

[craig.topper]

LGTM

lib/Transforms/InstCombine/InstCombineCalls.cpp

if (auto *ConstantMask = dyn_cast<ConstantDataVector>(Mask)) {
       Constant *NewSelector = getNegativeIsTrueBoolVec(ConstantMask);
       return SelectInst::Create(NewSelector, Op1, Op0, "blendv");
     }
     break;
   }
 
   case Intrinsic::x86_ssse3_pshuf_b_128:
   case Intrinsic::x86_avx2_pshuf_b:
-  case Intrinsic::x86_avx512_pshuf_b_512:
+  case Intrinsic::x86_avx512_pshuf_b_512: {
     if (Value *V = simplifyX86pshufb(*II, *Builder))
       return replaceInstUsesWith(*II, V);
+
+    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_avx_vpermilvar_ps:
   case Intrinsic::x86_avx_vpermilvar_ps_256:
   case Intrinsic::x86_avx512_vpermilvar_ps_512:
   case Intrinsic::x86_avx_vpermilvar_pd:
   case Intrinsic::x86_avx_vpermilvar_pd_256:
   case Intrinsic::x86_avx512_vpermilvar_pd_512:
     if (Value *V = simplifyX86vpermilvar(*II, *Builder))

lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

case Intrinsic::x86_avx512_pmulu_dq_512: {
       UndefElts3 = APInt(InnerVWidth, 0);
       TmpV = SimplifyDemandedVectorElts(Op1, InnerDemandedElts, UndefElts3,
                                         Depth + 1);
       if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
 
       break;
     }
 
+    case Intrinsic::x86_ssse3_pshuf_b_128:
+    case Intrinsic::x86_avx2_pshuf_b:
+    case Intrinsic::x86_avx512_pshuf_b_512: {
+      Value *Op1 = II->getArgOperand(1);
+      TmpV = SimplifyDemandedVectorElts(Op1, DemandedElts, UndefElts,
+                                        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-pshufb.ll

 ; CHECK-LABEL: @fold_with_allundef_elts_avx512(
 ; CHECK-NEXT:    ret <64 x i8> undef
 ;
   %1 = tail call <64 x i8> @llvm.x86.avx512.pshuf.b.512(<64 x i8> %InVec, <64 x i8> undef)
   ret <64 x i8> %1
 }
 
 ; Demanded elts tests.
-; FIXME: Missed opportunities to pass demanded elts through the pshufb shuffle mask
 
 define <16 x i8> @demanded_elts_insertion(<16 x i8> %InVec, <16 x i8> %BaseMask, i8 %M0, i8 %M15) {
 ; CHECK-LABEL: @demanded_elts_insertion(
-; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <16 x i8> %BaseMask, i8 %M0, i32 0
-; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <16 x i8> [[TMP1]], i8 %M15, i32 15
-; CHECK-NEXT:    [[TMP3:%.*]] = tail call <16 x i8> @llvm.x86.ssse3.pshuf.b.128(<16 x i8> %InVec, <16 x i8> [[TMP2]])
-; CHECK-NEXT:    [[TMP4:%.*]] = shufflevector <16 x i8> [[TMP3]], <16 x i8> undef, <16 x i32> <i32 undef, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 undef>
-; CHECK-NEXT:    ret <16 x i8> [[TMP4]]
+; CHECK-NEXT:    [[TMP1:%.*]] = tail call <16 x i8> @llvm.x86.ssse3.pshuf.b.128(<16 x i8> %InVec, <16 x i8> %BaseMask)
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <16 x i8> [[TMP1]], <16 x i8> undef, <16 x i32> <i32 undef, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 undef>
+; CHECK-NEXT:    ret <16 x i8> [[TMP2]]
 ;
   %1 = insertelement <16 x i8> %BaseMask, i8 %M0, i32 0
   %2 = insertelement <16 x i8> %1, i8 %M15, i32 15
   %3 = tail call <16 x i8> @llvm.x86.ssse3.pshuf.b.128(<16 x i8> %InVec, <16 x i8> %2)
   %4 = shufflevector <16 x i8> %3, <16 x i8> undef, <16 x i32> <i32 undef, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 undef>
   ret <16 x i8> %4
 }
 
 define <32 x i8> @demanded_elts_insertion_avx2(<32 x i8> %InVec, <32 x i8> %BaseMask, i8 %M0, i8 %M22) {
 ; CHECK-LABEL: @demanded_elts_insertion_avx2(
 ; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <32 x i8> %BaseMask, i8 %M0, i32 0
-; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <32 x i8> [[TMP1]], i8 %M22, i32 22
-; CHECK-NEXT:    [[TMP3:%.*]] = tail call <32 x i8> @llvm.x86.avx2.pshuf.b(<32 x i8> %InVec, <32 x i8> [[TMP2]])
-; CHECK-NEXT:    ret <32 x i8> [[TMP3]]
+; CHECK-NEXT:    [[TMP2:%.*]] = tail call <32 x i8> @llvm.x86.avx2.pshuf.b(<32 x i8> %InVec, <32 x i8> [[TMP1]])
+; CHECK-NEXT:    ret <32 x i8> [[TMP2]]
 ;
   %1 = insertelement <32 x i8> %BaseMask, i8 %M0, i32 0
   %2 = insertelement <32 x i8> %1, i8 %M22, i32 22
   %3 = tail call <32 x i8> @llvm.x86.avx2.pshuf.b(<32 x i8> %InVec, <32 x i8> %2)
   %4 = shufflevector <32 x i8> %3, <32 x i8> undef, <32 x i32> <i32 undef, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 undef, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
   ret <32 x i8> %4
 }
 
 define <64 x i8> @demanded_elts_insertion_avx512(<64 x i8> %InVec, <64 x i8> %BaseMask, i8 %M0, i8 %M30) {
 ; CHECK-LABEL: @demanded_elts_insertion_avx512(
-; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <64 x i8> %BaseMask, i8 %M0, i32 0
-; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <64 x i8> [[TMP1]], i8 %M30, i32 30
-; CHECK-NEXT:    [[TMP3:%.*]] = tail call <64 x i8> @llvm.x86.avx512.pshuf.b.512(<64 x i8> %InVec, <64 x i8> [[TMP2]])
-; CHECK-NEXT:    [[TMP4:%.*]] = shufflevector <64 x i8> [[TMP3]], <64 x i8> undef, <64 x i32> zeroinitializer
-; CHECK-NEXT:    ret <64 x i8> [[TMP4]]
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <64 x i8> undef, i8 %M0, i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = tail call <64 x i8> @llvm.x86.avx512.pshuf.b.512(<64 x i8> %InVec, <64 x i8> [[TMP1]])
+; CHECK-NEXT:    [[TMP3:%.*]] = shufflevector <64 x i8> [[TMP2]], <64 x i8> undef, <64 x i32> zeroinitializer
+; CHECK-NEXT:    ret <64 x i8> [[TMP3]]
 ;
   %1 = insertelement <64 x i8> %BaseMask, i8 %M0, i32 0
   %2 = insertelement <64 x i8> %1, i8 %M30, i32 30
   %3 = tail call <64 x i8> @llvm.x86.avx512.pshuf.b.512(<64 x i8> %InVec, <64 x i8> %2)
   %4 = shufflevector <64 x i8> %3, <64 x i8> undef, <64 x i32> zeroinitializer
   ret <64 x i8> %4
 }
 
 declare <16 x i8> @llvm.x86.ssse3.pshuf.b.128(<16 x i8>, <16 x i8>)
 declare <32 x i8> @llvm.x86.avx2.pshuf.b(<32 x i8>, <32 x i8>)
 declare <64 x i8> @llvm.x86.avx512.pshuf.b.512(<64 x i8>, <64 x i8>)