[X86] combineX86ShuffleChain - always prefer VPERMQ/PD for unary subvector shuffles on AVX2+ targets (#134849)

When combining 2 x 128-bit subvectors, don't assume that if the node is
already a X86ISD::VPERM2X128 node then there's nothing to do.

Fix issue where if we'd somehow combined to X86ISD::VPERM2X128
(typically if the 2 operands had then simplified to a common operand),
we can't canonicalise back to X86ISD::VPERMI on AVX2+ targets.

This matches the v4f64/v4i64 shuffle lowering preference for VPERMQ/PD
over VPERM2F128/I128.

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -39827,7 +39827,6 @@ static SDValue combineX86ShuffleChain(
 
     // If we're inserting the low subvector, an insert-subvector 'concat'
     // pattern is quicker than VPERM2X128.
-    // TODO: Add AVX2 support instead of VPERMQ/VPERMPD.
     if (BaseMask[0] == 0 && (BaseMask[1] == 0 || BaseMask[1] == 2) &&
         !Subtarget.hasAVX2()) {
       if (Depth == 0 && RootOpc == ISD::INSERT_SUBVECTOR)
@@ -39838,15 +39837,15 @@ static SDValue combineX86ShuffleChain(
       return insertSubVector(Lo, Hi, NumRootElts / 2, DAG, DL, 128);
     }
 
-    if (Depth == 0 && RootOpc == X86ISD::VPERM2X128)
-      return SDValue(); // Nothing to do!
-
-    // If we have AVX2, prefer to use VPERMQ/VPERMPD for unary shuffles unless
-    // we need to use the zeroing feature.
+    // Don't lower to VPERM2X128 here if we have AVX2+, prefer to use
+    // VPERMQ/VPERMPD for unary shuffles unless we need to use the zeroing
+    // feature.
     // Prefer blends for sequential shuffles unless we are optimizing for size.
     if (UnaryShuffle &&
         !(Subtarget.hasAVX2() && isUndefOrInRange(Mask, 0, 2)) &&
         (OptForSize || !isSequentialOrUndefOrZeroInRange(Mask, 0, 2, 0))) {
+      if (Depth == 0 && RootOpc == X86ISD::VPERM2X128)
+        return SDValue(); // Nothing to do!
       unsigned PermMask = 0;
       PermMask |= ((Mask[0] < 0 ? 0x8 : (Mask[0] & 1)) << 0);
       PermMask |= ((Mask[1] < 0 ? 0x8 : (Mask[1] & 1)) << 4);
@@ -39864,6 +39863,8 @@ static SDValue combineX86ShuffleChain(
              "Unexpected shuffle sentinel value");
       // Prefer blends to X86ISD::VPERM2X128.
       if (!((Mask[0] == 0 && Mask[1] == 3) || (Mask[0] == 2 && Mask[1] == 1))) {
+        if (Depth == 0 && RootOpc == X86ISD::VPERM2X128)
+          return SDValue(); // Nothing to do!
         unsigned PermMask = 0;
         PermMask |= ((Mask[0] & 3) << 0);
         PermMask |= ((Mask[1] & 3) << 4);

llvm/test/CodeGen/X86/matrix-multiply.ll

@@ -131,7 +131,7 @@ define <4 x double> @test_mul2x2_f64(<4 x double> %a0, <4 x double> %a1) nounwin
 ; AVX2-LABEL: test_mul2x2_f64:
 ; AVX2:       # %bb.0: # %entry
 ; AVX2-NEXT:    vshufpd {{.*#+}} ymm2 = ymm1[1,1,3,3]
-; AVX2-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm0[2,3,2,3]
 ; AVX2-NEXT:    vmulpd %ymm2, %ymm3, %ymm2
 ; AVX2-NEXT:    vmovddup {{.*#+}} ymm1 = ymm1[0,0,2,2]
 ; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,1]
@@ -142,7 +142,7 @@ define <4 x double> @test_mul2x2_f64(<4 x double> %a0, <4 x double> %a1) nounwin
 ; AVX512-LABEL: test_mul2x2_f64:
 ; AVX512:       # %bb.0: # %entry
 ; AVX512-NEXT:    vshufpd {{.*#+}} ymm2 = ymm1[1,1,3,3]
-; AVX512-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
+; AVX512-NEXT:    vpermpd {{.*#+}} ymm3 = ymm0[2,3,2,3]
 ; AVX512-NEXT:    vmulpd %ymm2, %ymm3, %ymm2
 ; AVX512-NEXT:    vmovddup {{.*#+}} ymm1 = ymm1[0,0,2,2]
 ; AVX512-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,1,0,1]