[X86] lowerShuffleWithUndefHalf - don't split vXi8 unary shuffles if the 128-bit source lanes are already in place (#122919)

Allows us to use PSHUFB to shuffle the lanes, and then perform a sub-lane permutation down to the lower half

Fixes #116815

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -15672,12 +15672,16 @@ static SDValue lowerShuffleWithUndefHalf(const SDLoc &DL, MVT VT, SDValue V1,
             (!isSingleSHUFPSMask(HalfMask) ||
              Subtarget.hasFastVariableCrossLaneShuffle()))
           return SDValue();
-        // If this is a unary shuffle (assume that the 2nd operand is
+        // If this is an unary shuffle (assume that the 2nd operand is
         // canonicalized to undef), then we can use vpermpd. Otherwise, we
         // are better off extracting the upper half of 1 operand and using a
         // narrow shuffle.
         if (EltWidth == 64 && V2.isUndef())
           return SDValue();
+        // If this is an unary vXi8 shuffle with inplace halves, then perform as
+        // full width pshufb, and then merge.
+        if (EltWidth == 8 && HalfIdx1 == 0 && HalfIdx2 == 1)
+          return SDValue();
       }
       // AVX512 has efficient cross-lane shuffles for all legal 512-bit types.
       if (Subtarget.hasAVX512() && VT.is512BitVector())

llvm/test/CodeGen/X86/shuffle-vs-trunc-128.ll

@@ -699,16 +699,13 @@ define <16 x i8> @evenelts_v32i16_shuffle_v16i16_to_v16i8(<32 x i16> %n2) nounwi
 ;
 ; AVX2-LABEL: evenelts_v32i16_shuffle_v16i16_to_v16i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
-; AVX2-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
-; AVX2-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; AVX2-NEXT:    vmovd {{.*#+}} xmm3 = [0,4,8,12,0,0,0,0,0,0,0,0,0,0,0,0]
-; AVX2-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
-; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
-; AVX2-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
+; AVX2-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm3
+; AVX2-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
+; AVX2-NEXT:    vpshufb %ymm2, %ymm1, %ymm1
+; AVX2-NEXT:    vpmovsxbd {{.*#+}} xmm2 = [0,0,0,4]
+; AVX2-NEXT:    vpermd %ymm1, %ymm2, %ymm1
 ; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
@@ -783,16 +780,13 @@ define <16 x i8> @oddelts_v32i16_shuffle_v16i16_to_v16i8(<32 x i16> %n2) nounwin
 ;
 ; AVX2-LABEL: oddelts_v32i16_shuffle_v16i16_to_v16i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
-; AVX2-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
-; AVX2-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
-; AVX2-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
-; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm2
-; AVX2-NEXT:    vmovd {{.*#+}} xmm3 = [2,6,10,14,0,0,0,0,0,0,0,0,0,0,0,0]
-; AVX2-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
-; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
-; AVX2-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
+; AVX2-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm3
+; AVX2-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
+; AVX2-NEXT:    vpshufb %ymm2, %ymm1, %ymm1
+; AVX2-NEXT:    vpmovsxbd {{.*#+}} xmm2 = [0,0,0,4]
+; AVX2-NEXT:    vpermd %ymm1, %ymm2, %ymm1
 ; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq

llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll

@@ -275,53 +275,45 @@ define <16 x i8> @trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_
 define <16 x i8> @trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_57_62(<64 x i8> %x) {
 ; AVX512F-LABEL: trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_57_62:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512F-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX512F-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,1,5,9,14,u,u,u,u,u,u,u,u]
-; AVX512F-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[u,u,u,u,1,5,9,13,u,u,u,u,u,u,u,u]
-; AVX512F-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
-; AVX512F-NEXT:    vpsrld $8, %ymm0, %ymm0
-; AVX512F-NEXT:    vpmovdb %zmm0, %xmm0
-; AVX512F-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
+; AVX512F-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,4,16,21]
+; AVX512F-NEXT:    vextracti64x4 $1, %zmm0, %ymm2
+; AVX512F-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,30,u,u,u,u,u,u,u,u]
+; AVX512F-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
+; AVX512F-NEXT:    vpermt2d %zmm2, %zmm1, %zmm0
+; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_57_62:
 ; AVX512VL:       # %bb.0:
-; AVX512VL-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512VL-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX512VL-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,1,5,9,14,u,u,u,u,u,u,u,u]
-; AVX512VL-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[u,u,u,u,1,5,9,13,u,u,u,u,u,u,u,u]
-; AVX512VL-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
-; AVX512VL-NEXT:    vpsrld $8, %ymm0, %ymm0
-; AVX512VL-NEXT:    vpmovdb %ymm0, %xmm0
-; AVX512VL-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
+; AVX512VL-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,4,8,13]
+; AVX512VL-NEXT:    vextracti64x4 $1, %zmm0, %ymm2
+; AVX512VL-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,30,u,u,u,u,u,u,u,u]
+; AVX512VL-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
+; AVX512VL-NEXT:    vpermt2d %ymm2, %ymm1, %ymm0
+; AVX512VL-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
 ; AVX512VL-NEXT:    vzeroupper
 ; AVX512VL-NEXT:    retq
 ;
 ; AVX512BW-LABEL: trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_57_62:
 ; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512BW-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,1,5,9,14,u,u,u,u,u,u,u,u]
-; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[u,u,u,u,1,5,9,13,u,u,u,u,u,u,u,u]
-; AVX512BW-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
-; AVX512BW-NEXT:    vpsrld $8, %ymm0, %ymm0
-; AVX512BW-NEXT:    vpmovdb %zmm0, %xmm0
-; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
+; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,4,16,21]
+; AVX512BW-NEXT:    vextracti64x4 $1, %zmm0, %ymm2
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,30,u,u,u,u,u,u,u,u]
+; AVX512BW-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
+; AVX512BW-NEXT:    vpermt2d %zmm2, %zmm1, %zmm0
+; AVX512BW-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
 ;
 ; AVX512BWVL-LABEL: trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_57_62:
 ; AVX512BWVL:       # %bb.0:
-; AVX512BWVL-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
-; AVX512BWVL-NEXT:    vextracti128 $1, %ymm1, %xmm2
-; AVX512BWVL-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,1,5,9,14,u,u,u,u,u,u,u,u]
-; AVX512BWVL-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[u,u,u,u,1,5,9,13,u,u,u,u,u,u,u,u]
-; AVX512BWVL-NEXT:    vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
-; AVX512BWVL-NEXT:    vpsrld $8, %ymm0, %ymm0
-; AVX512BWVL-NEXT:    vpmovdb %ymm0, %xmm0
-; AVX512BWVL-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
+; AVX512BWVL-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,4,8,13]
+; AVX512BWVL-NEXT:    vextracti64x4 $1, %zmm0, %ymm2
+; AVX512BWVL-NEXT:    vpshufb {{.*#+}} ymm2 = ymm2[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,30,u,u,u,u,u,u,u,u]
+; AVX512BWVL-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[1,5,9,13,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
+; AVX512BWVL-NEXT:    vpermt2d %ymm2, %ymm1, %ymm0
+; AVX512BWVL-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
 ; AVX512BWVL-NEXT:    vzeroupper
 ; AVX512BWVL-NEXT:    retq
 ;

llvm/test/CodeGen/X86/trunc-vector-width.ll

@@ -4,14 +4,16 @@
 define void @test(ptr %a0) #0 {
 ; CHECK-LABEL: test:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vmovdqu (%rdi), %xmm0
-; CHECK-NEXT:    vpblendd {{.*#+}} xmm0 = mem[0],xmm0[1,2,3]
-; CHECK-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[4,4,5,5,0,0,1,1,u,u,u,u,u,u,u,u]
-; CHECK-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; CHECK-NEXT:    vpternlogq $15, %xmm0, %xmm0, %xmm0
+; CHECK-NEXT:    vmovdqu (%rdi), %ymm0
+; CHECK-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,4,0,0]
+; CHECK-NEXT:    vpblendd {{.*#+}} ymm0 = mem[0],ymm0[1,2,3,4,5,6,7]
+; CHECK-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,4,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
+; CHECK-NEXT:    vpermd %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpternlogq {{.*#+}} xmm0 = ~xmm0
 ; CHECK-NEXT:    vpextrb $1, %xmm0, (%rax)
 ; CHECK-NEXT:    vpextrb $4, %xmm0, (%rax)
 ; CHECK-NEXT:    vpextrb $8, %xmm0, (%rax)
+; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %load = load <64 x i8>, ptr %a0, align 1
   %shuf = shufflevector <64 x i8> %load, <64 x i8> undef, <16 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 28, i32 32, i32 36, i32 40, i32 44, i32 48, i32 52, i32 56, i32 60>