[llvm] [X86] Use GFNI for vXi8 shifts/rotates (PR #89115)

[via llvm-commits]

================
@@ -262,84 +208,51 @@ declare <64 x i8> @llvm.fshl.v64i8(<64 x i8>, <64 x i8>, <64 x i8>)
 define <64 x i8> @splatconstant_fshr_v64i8(<64 x i8> %a, <64 x i8> %b) nounwind {
 ; GFNISSE-LABEL: splatconstant_fshr_v64i8:
 ; GFNISSE:       # %bb.0:
-; GFNISSE-NEXT:    psrlw $2, %xmm4
-; GFNISSE-NEXT:    movdqa {{.*#+}} xmm8 = [192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192]
-; GFNISSE-NEXT:    movdqa %xmm8, %xmm9
-; GFNISSE-NEXT:    pandn %xmm4, %xmm9
-; GFNISSE-NEXT:    psllw $6, %xmm0
-; GFNISSE-NEXT:    pand %xmm8, %xmm0
-; GFNISSE-NEXT:    por %xmm9, %xmm0
-; GFNISSE-NEXT:    psrlw $2, %xmm5
-; GFNISSE-NEXT:    movdqa %xmm8, %xmm4
-; GFNISSE-NEXT:    pandn %xmm5, %xmm4
-; GFNISSE-NEXT:    psllw $6, %xmm1
-; GFNISSE-NEXT:    pand %xmm8, %xmm1
-; GFNISSE-NEXT:    por %xmm4, %xmm1
-; GFNISSE-NEXT:    psrlw $2, %xmm6
-; GFNISSE-NEXT:    movdqa %xmm8, %xmm4
-; GFNISSE-NEXT:    pandn %xmm6, %xmm4
-; GFNISSE-NEXT:    psllw $6, %xmm2
-; GFNISSE-NEXT:    pand %xmm8, %xmm2
-; GFNISSE-NEXT:    por %xmm4, %xmm2
-; GFNISSE-NEXT:    psrlw $2, %xmm7
-; GFNISSE-NEXT:    psllw $6, %xmm3
-; GFNISSE-NEXT:    pand %xmm8, %xmm3
-; GFNISSE-NEXT:    pandn %xmm7, %xmm8
-; GFNISSE-NEXT:    por %xmm8, %xmm3
+; GFNISSE-NEXT:    movdqa {{.*#+}} xmm8 = [290499906672525312,290499906672525312]
+; GFNISSE-NEXT:    gf2p8affineqb $0, %xmm8, %xmm4
+; GFNISSE-NEXT:    pmovsxwq {{.*#+}} xmm9 = [258,258]
+; GFNISSE-NEXT:    gf2p8affineqb $0, %xmm9, %xmm0
+; GFNISSE-NEXT:    por %xmm4, %xmm0
+; GFNISSE-NEXT:    gf2p8affineqb $0, %xmm8, %xmm5
+; GFNISSE-NEXT:    gf2p8affineqb $0, %xmm9, %xmm1
+; GFNISSE-NEXT:    por %xmm5, %xmm1
+; GFNISSE-NEXT:    gf2p8affineqb $0, %xmm8, %xmm6
+; GFNISSE-NEXT:    gf2p8affineqb $0, %xmm9, %xmm2
+; GFNISSE-NEXT:    por %xmm6, %xmm2
+; GFNISSE-NEXT:    gf2p8affineqb $0, %xmm8, %xmm7
+; GFNISSE-NEXT:    gf2p8affineqb $0, %xmm9, %xmm3
+; GFNISSE-NEXT:    por %xmm7, %xmm3
 ; GFNISSE-NEXT:    retq
 ;
 ; GFNIAVX1-LABEL: splatconstant_fshr_v64i8:
 ; GFNIAVX1:       # %bb.0:
-; GFNIAVX1-NEXT:    vextractf128 $1, %ymm2, %xmm4
-; GFNIAVX1-NEXT:    vpsrlw $2, %xmm4, %xmm4
-; GFNIAVX1-NEXT:    vbroadcastss {{.*#+}} xmm5 = [63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63]
-; GFNIAVX1-NEXT:    vpand %xmm5, %xmm4, %xmm4
-; GFNIAVX1-NEXT:    vpsrlw $2, %xmm2, %xmm2
-; GFNIAVX1-NEXT:    vpand %xmm5, %xmm2, %xmm2
-; GFNIAVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm2, %ymm2
-; GFNIAVX1-NEXT:    vextractf128 $1, %ymm0, %xmm4
-; GFNIAVX1-NEXT:    vpsllw $6, %xmm4, %xmm4
-; GFNIAVX1-NEXT:    vbroadcastss {{.*#+}} xmm6 = [192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192]
-; GFNIAVX1-NEXT:    vpand %xmm6, %xmm4, %xmm4
-; GFNIAVX1-NEXT:    vpsllw $6, %xmm0, %xmm0
-; GFNIAVX1-NEXT:    vpand %xmm6, %xmm0, %xmm0
-; GFNIAVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm0
+; GFNIAVX1-NEXT:    vbroadcastsd {{.*#+}} ymm4 = [0,0,128,64,32,16,8,4,0,0,128,64,32,16,8,4,0,0,128,64,32,16,8,4,0,0,128,64,32,16,8,4]
+; GFNIAVX1-NEXT:    vgf2p8affineqb $0, %ymm4, %ymm2, %ymm2
+; GFNIAVX1-NEXT:    vbroadcastsd {{.*#+}} ymm5 = [2,1,0,0,0,0,0,0,2,1,0,0,0,0,0,0,2,1,0,0,0,0,0,0,2,1,0,0,0,0,0,0]
+; GFNIAVX1-NEXT:    vgf2p8affineqb $0, %ymm5, %ymm0, %ymm0
 ; GFNIAVX1-NEXT:    vorps %ymm2, %ymm0, %ymm0
-; GFNIAVX1-NEXT:    vextractf128 $1, %ymm3, %xmm2
-; GFNIAVX1-NEXT:    vpsrlw $2, %xmm2, %xmm2
-; GFNIAVX1-NEXT:    vpand %xmm5, %xmm2, %xmm2
-; GFNIAVX1-NEXT:    vpsrlw $2, %xmm3, %xmm3
-; GFNIAVX1-NEXT:    vpand %xmm5, %xmm3, %xmm3
-; GFNIAVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
-; GFNIAVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
-; GFNIAVX1-NEXT:    vpsllw $6, %xmm3, %xmm3
-; GFNIAVX1-NEXT:    vpand %xmm6, %xmm3, %xmm3
-; GFNIAVX1-NEXT:    vpsllw $6, %xmm1, %xmm1
-; GFNIAVX1-NEXT:    vpand %xmm6, %xmm1, %xmm1
-; GFNIAVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
+; GFNIAVX1-NEXT:    vgf2p8affineqb $0, %ymm4, %ymm3, %ymm2
+; GFNIAVX1-NEXT:    vgf2p8affineqb $0, %ymm5, %ymm1, %ymm1
 ; GFNIAVX1-NEXT:    vorps %ymm2, %ymm1, %ymm1
 ; GFNIAVX1-NEXT:    retq
 ;
 ; GFNIAVX2-LABEL: splatconstant_fshr_v64i8:
 ; GFNIAVX2:       # %bb.0:
-; GFNIAVX2-NEXT:    vpsrlw $2, %ymm2, %ymm2
-; GFNIAVX2-NEXT:    vpbroadcastb {{.*#+}} ymm4 = [192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192,192]
-; GFNIAVX2-NEXT:    vpandn %ymm2, %ymm4, %ymm2
-; GFNIAVX2-NEXT:    vpsllw $6, %ymm0, %ymm0
-; GFNIAVX2-NEXT:    vpand %ymm4, %ymm0, %ymm0
+; GFNIAVX2-NEXT:    vpbroadcastq {{.*#+}} ymm4 = [290499906672525312,290499906672525312,290499906672525312,290499906672525312]
+; GFNIAVX2-NEXT:    vgf2p8affineqb $0, %ymm4, %ymm2, %ymm2
+; GFNIAVX2-NEXT:    vpbroadcastq {{.*#+}} ymm5 = [258,258,258,258]
+; GFNIAVX2-NEXT:    vgf2p8affineqb $0, %ymm5, %ymm0, %ymm0
 ; GFNIAVX2-NEXT:    vpor %ymm2, %ymm0, %ymm0
-; GFNIAVX2-NEXT:    vpsrlw $2, %ymm3, %ymm2
-; GFNIAVX2-NEXT:    vpandn %ymm2, %ymm4, %ymm2
-; GFNIAVX2-NEXT:    vpsllw $6, %ymm1, %ymm1
-; GFNIAVX2-NEXT:    vpand %ymm4, %ymm1, %ymm1
+; GFNIAVX2-NEXT:    vgf2p8affineqb $0, %ymm4, %ymm3, %ymm2
+; GFNIAVX2-NEXT:    vgf2p8affineqb $0, %ymm5, %ymm1, %ymm1
 ; GFNIAVX2-NEXT:    vpor %ymm2, %ymm1, %ymm1
 ; GFNIAVX2-NEXT:    retq
 ;
 ; GFNIAVX512-LABEL: splatconstant_fshr_v64i8:
 ; GFNIAVX512:       # %bb.0:
-; GFNIAVX512-NEXT:    vpsllw $6, %zmm0, %zmm2
-; GFNIAVX512-NEXT:    vpsrlw $2, %zmm1, %zmm0
-; GFNIAVX512-NEXT:    vpternlogd $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm2, %zmm0
----------------
goldsteinn wrote:

likewise

https://github.com/llvm/llvm-project/pull/89115