[X86] matchBinaryPermuteShuffle - match AVX512 "cross lane" SHLDQ/SRLDQ style patterns using VALIGN

[RKSimon]

Very similar to what we do in lowerShuffleAsVALIGN

I've updated isTargetShuffleEquivalent to correctly handle SM_SentinelZero in the expected shuffle mask, but it only allows an exact match (or the test mask was undef) - it can't be used to match zero elements with MaskedVectorIsZero.

Noticed while working on #140516

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

Changes

Very similar to what we do in lowerShuffleAsVALIGN

I've updated isTargetShuffleEquivalent to correctly handle SM_SentinelZero in the expected shuffle mask, but it only allows an exact match (or the test mask was undef) - it can't be used to match zero elements with MaskedVectorIsZero.

Noticed while working on #140516

---

Full diff: https://github.com/llvm/llvm-project/pull/140538.diff

2 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+37-5)
• (modified) llvm/test/CodeGen/X86/vector-shuffle-combining-avx512f.ll (+2-5)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 3b6b0d7b86c9c..6f58910e55add 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -10096,7 +10096,9 @@ static bool isTargetShuffleEquivalent(MVT VT, ArrayRef<int> Mask,
   if (Size != (int)ExpectedMask.size())
     return false;
   assert(llvm::all_of(ExpectedMask,
-                      [Size](int M) { return isInRange(M, 0, 2 * Size); }) &&
+                      [Size](int M) {
+                        return M == SM_SentinelZero || (M, 0, 2 * Size);
+                      }) &&
          "Illegal target shuffle mask");
 
   // Check for out-of-range target shuffle mask indices.
@@ -10119,6 +10121,9 @@ static bool isTargetShuffleEquivalent(MVT VT, ArrayRef<int> Mask,
     int ExpectedIdx = ExpectedMask[i];
     if (MaskIdx == SM_SentinelUndef || MaskIdx == ExpectedIdx)
       continue;
+    // If we failed to match an expected SM_SentinelZero then early out.
+    if (ExpectedIdx < 0)
+      return false;
     if (MaskIdx == SM_SentinelZero) {
       // If we need this expected index to be a zero element, then update the
       // relevant zero mask and perform the known bits at the end to minimize
@@ -39594,18 +39599,45 @@ static bool matchBinaryPermuteShuffle(
       ((MaskVT.is128BitVector() && Subtarget.hasVLX()) ||
        (MaskVT.is256BitVector() && Subtarget.hasVLX()) ||
        (MaskVT.is512BitVector() && Subtarget.hasAVX512()))) {
+    MVT AlignVT = MVT::getVectorVT(MVT::getIntegerVT(EltSizeInBits),
+                                   MaskVT.getSizeInBits() / EltSizeInBits);
     if (!isAnyZero(Mask)) {
       int Rotation = matchShuffleAsElementRotate(V1, V2, Mask);
       if (0 < Rotation) {
         Shuffle = X86ISD::VALIGN;
-        if (EltSizeInBits == 64)
-          ShuffleVT = MVT::getVectorVT(MVT::i64, MaskVT.getSizeInBits() / 64);
-        else
-          ShuffleVT = MVT::getVectorVT(MVT::i32, MaskVT.getSizeInBits() / 32);
+        ShuffleVT = AlignVT;
         PermuteImm = Rotation;
         return true;
       }
     }
+    // See if we can use VALIGN as a cross-lane version of VSHLDQ/VSRLDQ.
+    unsigned ZeroLo = Zeroable.countr_one();
+    unsigned ZeroHi = Zeroable.countl_one();
+    assert((ZeroLo + ZeroHi) < NumMaskElts && "Zeroable shuffle detected");
+    if (ZeroLo) {
+      SmallVector<int, 16> ShiftMask(NumMaskElts, SM_SentinelZero);
+      std::iota(ShiftMask.begin() + ZeroLo, ShiftMask.end(), 0);
+      if (isTargetShuffleEquivalent(MaskVT, Mask, ShiftMask, DAG, V1)) {
+        V1 = V1;
+        V2 = getZeroVector(AlignVT, Subtarget, DAG, DL);
+        Shuffle = X86ISD::VALIGN;
+        ShuffleVT = AlignVT;
+        PermuteImm = NumMaskElts - ZeroLo;
+        return true;
+      }
+    }
+    if (ZeroHi) {
+      SmallVector<int, 16> ShiftMask(NumMaskElts, SM_SentinelZero);
+      std::iota(ShiftMask.begin(), ShiftMask.begin() + NumMaskElts - ZeroHi, ZeroHi);
+      if (isTargetShuffleEquivalent(MaskVT, Mask, ShiftMask, DAG, V1)) {
+        V2 = V1;
+        V1 = getZeroVector(AlignVT, Subtarget, DAG, DL);
+        Shuffle = X86ISD::VALIGN;
+        ShuffleVT = AlignVT;
+        PermuteImm = ZeroHi;
+        return true;
+      }
+    }
   }
 
   // Attempt to match against PALIGNR byte rotate.
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512f.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512f.ll
index b3b90b5f51501..68967c2ce6536 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512f.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512f.ll
@@ -812,10 +812,8 @@ define <8 x i64> @combine_vpermt2var_8i64_as_valignq(<8 x i64> %x0, <8 x i64> %x
 define <8 x i64> @combine_vpermt2var_8i64_as_valignq_zero(<8 x i64> %x0) {
 ; CHECK-LABEL: combine_vpermt2var_8i64_as_valignq_zero:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpmovsxbq {{.*#+}} zmm2 = [15,0,1,2,3,4,5,6]
 ; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpermt2q %zmm0, %zmm2, %zmm1
-; CHECK-NEXT:    vmovdqa64 %zmm1, %zmm0
+; CHECK-NEXT:    valignq {{.*#+}} zmm0 = zmm0[7],zmm1[0,1,2,3,4,5,6]
 ; CHECK-NEXT:    ret{{[l|q]}}
   %res0 = call <8 x i64> @llvm.x86.avx512.maskz.vpermt2var.q.512(<8 x i64> <i64 15, i64 0, i64 1, i64 2, i64 3, i64 4, i64 5, i64 6>, <8 x i64> zeroinitializer, <8 x i64> %x0, i8 -1)
   ret <8 x i64> %res0
@@ -825,8 +823,7 @@ define <8 x i64> @combine_vpermt2var_8i64_as_zero_valignq(<8 x i64> %x0) {
 ; CHECK-LABEL: combine_vpermt2var_8i64_as_zero_valignq:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpmovsxbq {{.*#+}} zmm2 = [15,0,1,2,3,4,5,6]
-; CHECK-NEXT:    vpermt2q %zmm1, %zmm2, %zmm0
+; CHECK-NEXT:    valignq {{.*#+}} zmm0 = zmm1[7],zmm0[0,1,2,3,4,5,6]
 ; CHECK-NEXT:    ret{{[l|q]}}
   %res0 = call <8 x i64> @llvm.x86.avx512.maskz.vpermt2var.q.512(<8 x i64> <i64 15, i64 0, i64 1, i64 2, i64 3, i64 4, i64 5, i64 6>, <8 x i64> %x0, <8 x i64> zeroinitializer, i8 -1)
   ret <8 x i64> %res0

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[phoebewang]

LGTM.

[kazutakahirata]

@RKSimon I've landed ad80f73 to fix a warning from this PR. Thanks!