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

[Simon Pilgrim via llvm-commits]

================
@@ -41,25 +37,23 @@ declare <16 x i8> @llvm.fshl.v16i8(<16 x i8>, <16 x i8>, <16 x i8>)
 define <16 x i8> @splatconstant_fshr_v16i8(<16 x i8> %a, <16 x i8> %b) nounwind {
 ; GFNISSE-LABEL: splatconstant_fshr_v16i8:
 ; GFNISSE:       # %bb.0:
-; GFNISSE-NEXT:    psrlw $7, %xmm1
-; GFNISSE-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; GFNISSE-NEXT:    gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; GFNISSE-NEXT:    paddb %xmm0, %xmm0
 ; GFNISSE-NEXT:    por %xmm1, %xmm0
 ; GFNISSE-NEXT:    retq
 ;
 ; GFNIAVX1OR2-LABEL: splatconstant_fshr_v16i8:
 ; GFNIAVX1OR2:       # %bb.0:
-; GFNIAVX1OR2-NEXT:    vpsrlw $7, %xmm1, %xmm1
-; GFNIAVX1OR2-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
+; GFNIAVX1OR2-NEXT:    vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
 ; GFNIAVX1OR2-NEXT:    vpaddb %xmm0, %xmm0, %xmm0
 ; GFNIAVX1OR2-NEXT:    vpor %xmm1, %xmm0, %xmm0
 ; GFNIAVX1OR2-NEXT:    retq
 ;
 ; GFNIAVX512-LABEL: splatconstant_fshr_v16i8:
 ; GFNIAVX512:       # %bb.0:
-; GFNIAVX512-NEXT:    vpsrlw $7, %xmm1, %xmm1
+; GFNIAVX512-NEXT:    vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to2}, %xmm1, %xmm1
 ; GFNIAVX512-NEXT:    vpaddb %xmm0, %xmm0, %xmm0
-; GFNIAVX512-NEXT:    vpternlogd $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm1, %xmm0
----------------
RKSimon wrote:

This looks to be be about the same perf, but I'll see if we can keep the existing codegen as shifts+bitops are much more likely to fold further: https://llvm.godbolt.org/z/ozrbPPv1s

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