[llvm] [X86] Fold blend(pshufb(x,m1),pshufb(y,m2)) -> blend(pshufb(x,blend(m1,m2)),pshufb(y,blend(m1,m2))) to reduce constant pool (PR #98466)

via llvm-commits llvm-commits at lists.llvm.org
Thu Jul 11 04:31:49 PDT 2024


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

<details>
<summary>Changes</summary>

Share PSHUFB masks where we have no overlap in used elements.

Fixes #<!-- -->98346

---

Patch is 1.24 MiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/98466.diff


23 Files Affected:

- (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+53-17) 
- (modified) llvm/test/CodeGen/X86/oddshuffles.ll (+8-6) 
- (modified) llvm/test/CodeGen/X86/shuffle-strided-with-offset-512.ll (+3-2) 
- (modified) llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll (+6-4) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-2.ll (+48-51) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-4.ll (+656-682) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-5.ll (+64-52) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-6.ll (+2267-2301) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-7.ll (+2568-2537) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-2.ll (+47-37) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-5.ll (+77-65) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-6.ll (+132-100) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-7.ll (+204-188) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-3.ll (+16-12) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-4.ll (+16-12) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-5.ll (+48-38) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-6.ll (+130-98) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-7.ll (+12-9) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-store-i16-stride-8.ll (+17-13) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-6.ll (+44-33) 
- (modified) llvm/test/CodeGen/X86/vector-interleaved-store-i8-stride-8.ll (+217-214) 
- (modified) llvm/test/CodeGen/X86/vector-shuffle-256-v16.ll (+12-9) 
- (modified) llvm/test/CodeGen/X86/vector-shuffle-256-v32.ll (+24-16) 


``````````diff
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index ea916b778d86d..92e2c75892754 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -41016,23 +41016,59 @@ static SDValue combineTargetShuffle(SDValue N, const SDLoc &DL,
   case X86ISD::BLENDI: {
     SDValue N0 = N.getOperand(0);
     SDValue N1 = N.getOperand(1);
-
-    // blend(bitcast(x),bitcast(y)) -> bitcast(blend(x,y)) to narrower types.
-    // TODO: Handle MVT::v16i16 repeated blend mask.
-    if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
-        N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
-      MVT SrcVT = N0.getOperand(0).getSimpleValueType();
-      if ((VT.getScalarSizeInBits() % SrcVT.getScalarSizeInBits()) == 0 &&
-          SrcVT.getScalarSizeInBits() >= 32) {
-        unsigned Size = VT.getVectorNumElements();
-        unsigned NewSize = SrcVT.getVectorNumElements();
-        APInt BlendMask = N.getConstantOperandAPInt(2).zextOrTrunc(Size);
-        APInt NewBlendMask = APIntOps::ScaleBitMask(BlendMask, NewSize);
-        return DAG.getBitcast(
-            VT, DAG.getNode(X86ISD::BLENDI, DL, SrcVT, N0.getOperand(0),
-                            N1.getOperand(0),
-                            DAG.getTargetConstant(NewBlendMask.getZExtValue(),
-                                                  DL, MVT::i8)));
+    unsigned EltBits = VT.getScalarSizeInBits();
+
+    if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST) {
+      // blend(bitcast(x),bitcast(y)) -> bitcast(blend(x,y)) to narrower types.
+      // TODO: Handle MVT::v16i16 repeated blend mask.
+      if (N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
+        MVT SrcVT = N0.getOperand(0).getSimpleValueType();
+        unsigned SrcBits = SrcVT.getScalarSizeInBits();
+        if ((EltBits % SrcBits) == 0 && SrcBits >= 32) {
+          unsigned Size = VT.getVectorNumElements();
+          unsigned NewSize = SrcVT.getVectorNumElements();
+          APInt BlendMask = N.getConstantOperandAPInt(2).zextOrTrunc(Size);
+          APInt NewBlendMask = APIntOps::ScaleBitMask(BlendMask, NewSize);
+          return DAG.getBitcast(
+              VT, DAG.getNode(X86ISD::BLENDI, DL, SrcVT, N0.getOperand(0),
+                              N1.getOperand(0),
+                              DAG.getTargetConstant(NewBlendMask.getZExtValue(),
+                                                    DL, MVT::i8)));
+        }
+      }
+      // Share PSHUFB masks:
+      // blend(pshufb(x,m1),pshufb(y,m2))
+      // --> m3 = blend(m1,m2)
+      //     blend(pshufb(x,m3),pshufb(y,m3))
+      if (N0.hasOneUse() && N1.hasOneUse()) {
+        SmallVector<int> Mask, ByteMask;
+        SmallVector<SDValue> Ops;
+        SDValue LHS = peekThroughOneUseBitcasts(N0);
+        SDValue RHS = peekThroughOneUseBitcasts(N1);
+        if (LHS.getOpcode() == X86ISD::PSHUFB &&
+            RHS.getOpcode() == X86ISD::PSHUFB &&
+            LHS.getOperand(1) != RHS.getOperand(1) &&
+            (LHS.getOperand(1).hasOneUse() || RHS.getOperand(1).hasOneUse()) &&
+            getTargetShuffleMask(N, false, Ops, Mask)) {
+          assert(Ops.size() == 2 && LHS == peekThroughOneUseBitcasts(Ops[0]) &&
+                 RHS == peekThroughOneUseBitcasts(Ops[1]) &&
+                 "BLENDI decode mismatch");
+          MVT ShufVT = LHS.getSimpleValueType();
+          SDValue MaskLHS = LHS.getOperand(1);
+          SDValue MaskRHS = RHS.getOperand(1);
+          llvm::narrowShuffleMaskElts(EltBits / 8, Mask, ByteMask);
+          if (SDValue NewMask = combineX86ShufflesConstants(
+                  ShufVT, {MaskLHS, MaskRHS}, ByteMask, true, DAG, DL,
+                  Subtarget)) {
+            SDValue NewLHS = DAG.getNode(X86ISD::PSHUFB, DL, ShufVT,
+                                         LHS.getOperand(0), NewMask);
+            SDValue NewRHS = DAG.getNode(X86ISD::PSHUFB, DL, ShufVT,
+                                         RHS.getOperand(0), NewMask);
+            return DAG.getNode(X86ISD::BLENDI, DL, VT,
+                               DAG.getBitcast(VT, NewLHS),
+                               DAG.getBitcast(VT, NewRHS), N.getOperand(2));
+          }
+        }
       }
     }
     return SDValue();
diff --git a/llvm/test/CodeGen/X86/oddshuffles.ll b/llvm/test/CodeGen/X86/oddshuffles.ll
index d3a3b1e980db0..b40b2c82843cc 100644
--- a/llvm/test/CodeGen/X86/oddshuffles.ll
+++ b/llvm/test/CodeGen/X86/oddshuffles.ll
@@ -1294,10 +1294,11 @@ define void @interleave_24i16_in(ptr %p, ptr %q1, ptr %q2, ptr %q3) nounwind {
 ; AVX2-SLOW-NEXT:    vmovdqu (%rdx), %xmm1
 ; AVX2-SLOW-NEXT:    vmovdqu (%rcx), %xmm2
 ; AVX2-SLOW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm3
-; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm4 = ymm3[0,1,u,u,6,7,2,3,u,u,8,9,4,5,u,u,16,17,u,u,22,23,18,19,u,u,24,25,20,21,u,u]
+; AVX2-SLOW-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,0,1,6,7,2,3,2,3,8,9,4,5,4,5,16,17,6,7,22,23,18,19,8,9,24,25,20,21,10,11]
+; AVX2-SLOW-NEXT:    vpshufb %ymm4, %ymm3, %ymm5
 ; AVX2-SLOW-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1]
-; AVX2-SLOW-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,0,1,u,u,u,u,2,3,u,u,u,u,4,5,u,u,22,23,u,u,u,u,24,25,u,u,u,u,26,27]
-; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6],ymm3[7],ymm4[8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14],ymm3[15]
+; AVX2-SLOW-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
+; AVX2-SLOW-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14],ymm3[15]
 ; AVX2-SLOW-NEXT:    vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,1,1,0,2]
 ; AVX2-SLOW-NEXT:    vpermd %ymm2, %ymm4, %ymm4
 ; AVX2-SLOW-NEXT:    vpmovsxbw {{.*#+}} ymm5 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
@@ -1339,10 +1340,11 @@ define void @interleave_24i16_in(ptr %p, ptr %q1, ptr %q2, ptr %q3) nounwind {
 ; AVX2-FAST-PERLANE-NEXT:    vmovdqu (%rdx), %xmm1
 ; AVX2-FAST-PERLANE-NEXT:    vmovdqu (%rcx), %xmm2
 ; AVX2-FAST-PERLANE-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm3
-; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm4 = ymm3[0,1,u,u,6,7,2,3,u,u,8,9,4,5,u,u,16,17,u,u,22,23,18,19,u,u,24,25,20,21,u,u]
+; AVX2-FAST-PERLANE-NEXT:    vmovdqa {{.*#+}} ymm4 = [0,1,0,1,6,7,2,3,2,3,8,9,4,5,4,5,16,17,6,7,22,23,18,19,8,9,24,25,20,21,10,11]
+; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm4, %ymm3, %ymm5
 ; AVX2-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1]
-; AVX2-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm3 = ymm3[u,u,0,1,u,u,u,u,2,3,u,u,u,u,4,5,u,u,22,23,u,u,u,u,24,25,u,u,u,u,26,27]
-; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6],ymm3[7],ymm4[8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14],ymm3[15]
+; AVX2-FAST-PERLANE-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
+; AVX2-FAST-PERLANE-NEXT:    vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14],ymm3[15]
 ; AVX2-FAST-PERLANE-NEXT:    vpmovsxbd {{.*#+}} ymm4 = [0,0,0,0,1,1,0,2]
 ; AVX2-FAST-PERLANE-NEXT:    vpermd %ymm2, %ymm4, %ymm4
 ; AVX2-FAST-PERLANE-NEXT:    vpmovsxbw {{.*#+}} ymm5 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
diff --git a/llvm/test/CodeGen/X86/shuffle-strided-with-offset-512.ll b/llvm/test/CodeGen/X86/shuffle-strided-with-offset-512.ll
index e94f51233256c..45842d4148a8b 100644
--- a/llvm/test/CodeGen/X86/shuffle-strided-with-offset-512.ll
+++ b/llvm/test/CodeGen/X86/shuffle-strided-with-offset-512.ll
@@ -12,8 +12,9 @@ define void @shuffle_v64i8_to_v32i8_1(ptr %L, ptr %S) nounwind {
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512F-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u,17,19,21,23,25,27,29,31]
-; AVX512F-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u,17,19,21,23,25,27,29,31,u,u,u,u,u,u,u,u]
+; AVX512F-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [1,3,5,7,9,11,13,15,1,3,5,7,9,11,13,15,1,3,5,7,9,11,13,15,1,3,5,7,9,11,13,15]
+; AVX512F-NEXT:    vpshufb %ymm2, %ymm1, %ymm1
+; AVX512F-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
 ; AVX512F-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
 ; AVX512F-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX512F-NEXT:    vmovdqa %ymm0, (%rsi)
diff --git a/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll b/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
index 95e249984e184..cf0820aac3262 100644
--- a/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
+++ b/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
@@ -20,8 +20,9 @@ define void @shuffle_v64i8_to_v32i8(ptr %L, ptr %S) nounwind {
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512F-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512F-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30]
-; AVX512F-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30,u,u,u,u,u,u,u,u]
+; AVX512F-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14]
+; AVX512F-NEXT:    vpshufb %ymm2, %ymm1, %ymm1
+; AVX512F-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
 ; AVX512F-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
 ; AVX512F-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX512F-NEXT:    vmovdqa %ymm0, (%rsi)
@@ -44,8 +45,9 @@ define void @shuffle_v64i8_to_v32i8(ptr %L, ptr %S) nounwind {
 ; AVX512VL-FAST-PERLANE:       # %bb.0:
 ; AVX512VL-FAST-PERLANE-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX512VL-FAST-PERLANE-NEXT:    vmovdqa 32(%rdi), %ymm1
-; AVX512VL-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30]
-; AVX512VL-FAST-PERLANE-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30,u,u,u,u,u,u,u,u]
+; AVX512VL-FAST-PERLANE-NEXT:    vpbroadcastq {{.*#+}} ymm2 = [0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14,0,2,4,6,8,10,12,14]
+; AVX512VL-FAST-PERLANE-NEXT:    vpshufb %ymm2, %ymm1, %ymm1
+; AVX512VL-FAST-PERLANE-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
 ; AVX512VL-FAST-PERLANE-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
 ; AVX512VL-FAST-PERLANE-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX512VL-FAST-PERLANE-NEXT:    vmovdqa %ymm0, (%rsi)
diff --git a/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-2.ll b/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-2.ll
index 3bc97f71f04fb..00e43df15deea 100644
--- a/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-2.ll
+++ b/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-2.ll
@@ -488,8 +488,9 @@ define void @load_i16_stride2_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) no
 ; AVX2-NEXT:    vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
 ; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm3[0,2],ymm2[0,2],ymm3[4,6],ymm2[4,6]
 ; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]
-; AVX2-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31]
-; AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31,u,u,u,u,u,u,u,u]
+; AVX2-NEXT:    vmovdqa {{.*#+}} ymm3 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
+; AVX2-NEXT:    vpshufb %ymm3, %ymm1, %ymm1
+; AVX2-NEXT:    vpshufb %ymm3, %ymm0, %ymm0
 ; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
 ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vmovaps %ymm2, (%rsi)
@@ -506,8 +507,9 @@ define void @load_i16_stride2_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) no
 ; AVX2-FP-NEXT:    vpshufb %ymm2, %ymm0, %ymm2
 ; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm2 = ymm2[0,2],ymm3[0,2],ymm2[4,6],ymm3[4,6]
 ; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]
-; AVX2-FP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31]
-; AVX2-FP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31,u,u,u,u,u,u,u,u]
+; AVX2-FP-NEXT:    vmovdqa {{.*#+}} ymm3 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
+; AVX2-FP-NEXT:    vpshufb %ymm3, %ymm1, %ymm1
+; AVX2-FP-NEXT:    vpshufb %ymm3, %ymm0, %ymm0
 ; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
 ; AVX2-FP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-FP-NEXT:    vmovaps %ymm2, (%rsi)
@@ -524,8 +526,9 @@ define void @load_i16_stride2_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) no
 ; AVX2-FCP-NEXT:    vpshufb %ymm2, %ymm0, %ymm2
 ; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm2 = ymm2[0,2],ymm3[0,2],ymm2[4,6],ymm3[4,6]
 ; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]
-; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31]
-; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31,u,u,u,u,u,u,u,u]
+; AVX2-FCP-NEXT:    vmovdqa {{.*#+}} ymm3 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
+; AVX2-FCP-NEXT:    vpshufb %ymm3, %ymm1, %ymm1
+; AVX2-FCP-NEXT:    vpshufb %ymm3, %ymm0, %ymm0
 ; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
 ; AVX2-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rsi)
@@ -736,14 +739,13 @@ define void @load_i16_stride2_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) no
 ; AVX2-NEXT:    vpshufhw {{.*#+}} ymm6 = ymm6[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
 ; AVX2-NEXT:    vshufps {{.*#+}} ymm5 = ymm6[0,2],ymm5[0,2],ymm6[4,6],ymm5[4,6]
 ; AVX2-NEXT:    vpermpd {{.*#+}} ymm5 = ymm5[0,2,1,3]
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31]
+; AVX2-NEXT:    vmovdqa {{.*#+}} ymm6 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
 ; AVX2-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm7 = [2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31,u,u,u,u,u,u,u,u]
-; AVX2-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
+; AVX2-NEXT:    vpshufb %ymm6, %ymm2, %ymm2
 ; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]
 ; AVX2-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,1,3]
 ; AVX2-NEXT:    vpshufb %ymm6, %ymm1, %ymm1
-; AVX2-NEXT:    vpshufb %ymm7, %ymm0, %ymm0
+; AVX2-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
 ; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
 ; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-NEXT:    vmovaps %ymm5, (%rsi)
@@ -768,14 +770,13 @@ define void @load_i16_stride2_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) no
 ; AVX2-FP-NEXT:    vpshufb %ymm4, %ymm0, %ymm4
 ; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[0,2],ymm6[0,2],ymm4[4,6],ymm6[4,6]
 ; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm4[0,2,1,3]
-; AVX2-FP-NEXT:    vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31]
+; AVX2-FP-NEXT:    vmovdqa {{.*#+}} ymm6 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
 ; AVX2-FP-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX2-FP-NEXT:    vmovdqa {{.*#+}} ymm7 = [2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31,u,u,u,u,u,u,u,u]
-; AVX2-FP-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
+; AVX2-FP-NEXT:    vpshufb %ymm6, %ymm2, %ymm2
 ; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]
 ; AVX2-FP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,1,3]
 ; AVX2-FP-NEXT:    vpshufb %ymm6, %ymm1, %ymm1
-; AVX2-FP-NEXT:    vpshufb %ymm7, %ymm0, %ymm0
+; AVX2-FP-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
 ; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
 ; AVX2-FP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-FP-NEXT:    vmovaps %ymm4, (%rsi)
@@ -800,14 +801,13 @@ define void @load_i16_stride2_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) no
 ; AVX2-FCP-NEXT:    vpshufb %ymm4, %ymm0, %ymm4
 ; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[0,2],ymm6[0,2],ymm4[4,6],ymm6[4,6]
 ; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm4[0,2,1,3]
-; AVX2-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31]
+; AVX2-FCP-NEXT:    vmovdqa {{.*#+}} ymm6 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]
 ; AVX2-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm3
-; AVX2-FCP-NEXT:    vmovdqa {{.*#+}} ymm7 = [2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31,u,u,u,u,u,u,u,u]
-; AVX2-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
+; AVX2-FCP-NEXT:    vpshufb %ymm6, %ymm2, %ymm2
 ; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]
 ; AVX2-FCP-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,1,3]
 ; AVX2-FCP-NEXT:    vpshufb %ymm6, %ymm1, %ymm1
-; AVX2-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm0
+; AVX2-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm0
 ; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
 ; AVX2-FCP-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; AVX2-FCP-NEXT:    vmovaps %ymm4, (%rsi)
@@ -1180,20 +1180,20 @@ define void @load_i16_stride2_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) no
 ; AVX2-NEXT:    vmovdqa 224(%rdi), %ymm6
 ; AVX2-NEXT:    vmovdqa (%rdi), %ymm3
 ; AVX2-NEXT:    vmovdqa 32(%rdi), %ymm5
-; AVX2-NEXT:    vmovdqa 64(%rdi), %ymm8
-; AVX2-NEXT:    vmovdqa 96(%rdi), %ymm9
-; AVX2-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm9[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX2-NEXT:    vmovdqa 64(%rdi), %ymm7
+; AVX2-NEXT:    vmovdqa 96(%rdi), %ymm8
+; AVX2-NEXT:    vpshuflw {{.*#+}} ymm2 = ymm8[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
 ; AVX2-NEXT:    vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
-; AVX2-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm8[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX2-NEXT:    vpshufhw {{.*#+}} ymm7 = ymm7[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
-; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm7[0,2],ymm2[0,2],ymm7[4,6],ymm2[4,6]
+; AVX2-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm7[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX2-NEXT:    vpshufhw {{.*#+}} ymm9 = ymm9[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
+; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm9[0,2],ymm2[0,2],ymm9[4,6],ymm2[4,6]
 ; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]
-; AVX2-NEXT:    vpshuflw {{.*#+}} ymm7 = ymm6[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
-; AVX2-NEXT:    vpshufhw {{.*#+}} ymm7 = ymm7[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
+; AVX2-NEXT:    vpshuflw {{.*#+}} ymm9 = ymm6[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
+; AVX2-NEXT:    vpshufhw {{.*#+}} ymm9 = ymm9[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
 ; AVX2-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm4[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
 ; AVX2-NEXT:    vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
-; AVX2-NEXT:    vshufps {{.*#+}} ymm7 = ymm10[0,2],ymm7[0,2],ymm10[4,6],ymm7[4,6]
-; AVX2-NEXT:    vpermpd {{.*#+}} ymm7 = ymm7[0,2,1,3]
+; AVX2-NEXT:    vshufps {{.*#+}} ymm9 = ymm10[0,2],ymm9[0,2],ymm10[4,6],ymm9[4,6]
+; AVX2-NEXT:    vpermpd {{.*#+}} ymm9 = ymm9[0,2,1,3]
 ; AVX2-NEXT:    vpshuflw {{.*#+}} ymm10 = ymm5[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
 ; AVX2-NEXT:    vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
 ; AVX2-NEXT:    vpshuflw {{.*#+}} ymm11 = ymm3[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]
@@ -1206,32 +1206,31 @@ define void @load_i16_stride2_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) no
 ; AVX2-NEXT:    vpshufhw {{.*#+}} ymm12 = ymm12[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]
 ; AVX2-NEXT:    vshufps {{.*#+}} ymm11 = ymm12[0,2],ymm11[0,2],ymm12[4,6],ymm11[4,6]
 ; AVX2-NEXT:    vpermpd {{.*#+}} ymm11 = ymm11[0,2,1,3]
-; AVX2-NEXT:    vmovdqa {{.*#+}} ymm12 = [u,u,u,u,u,u,u,u,2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u,18,19,22,23,26,27,30,31]
-; AVX2-NEXT:    vpshufb %ymm12, %ymm9, %ymm9
-; AVX2-NEXT:    vmovd...
[truncated]

``````````

</details>


https://github.com/llvm/llvm-project/pull/98466


More information about the llvm-commits mailing list