[llvm] [X86] LowerShiftByScalarImmediate - vXi8 shl(X,2) - prefer PADDB+PADDB pair over PSLLW+PAND (PR #186095)

[Simon Pilgrim via llvm-commits]

================
@@ -447,11 +447,13 @@ define <32 x i8> @var_funnnel_v32i8(<32 x i8> %x, <32 x i8> %amt) nounwind {
 ; AVX512F-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (m32bcst & (zmm3 ^ zmm2))
 ; AVX512F-NEXT:    vpsllw $5, %ymm1, %ymm1
 ; AVX512F-NEXT:    vpblendvb %ymm1, %ymm3, %ymm0, %ymm0
-; AVX512F-NEXT:    vpsllw $2, %ymm0, %ymm2
-; AVX512F-NEXT:    vpsrlw $6, %ymm0, %ymm3
-; AVX512F-NEXT:    vpternlogd {{.*#+}} zmm3 = zmm3 ^ (m32bcst & (zmm3 ^ zmm2))
+; AVX512F-NEXT:    vpsrlw $6, %ymm0, %ymm2
+; AVX512F-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm2, %ymm2
+; AVX512F-NEXT:    vpaddb %ymm0, %ymm0, %ymm3
+; AVX512F-NEXT:    vpaddb %ymm3, %ymm3, %ymm3
+; AVX512F-NEXT:    vpor %ymm2, %ymm3, %ymm2
----------------
RKSimon wrote:

This is the repeated mask issue I mentioned on #189986 - I can add custom rotate folding for non-AVX512VL targets to get the VPTERNLOG back if you think it necessary, but its pretty niche.

https://github.com/llvm/llvm-project/pull/186095