[llvm] [X86] Fold EXTRACT_SUBVECTOR(ONEUSE(EXTRACT_SUBVECTOR(V,C1))),C2) - EXTRACT_SUBVECTOR(V,C1+C2) (PR #111685)

[Simon Pilgrim via llvm-commits]

================
@@ -4565,17 +4557,30 @@ define void @vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3(ptr %i
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 ;
-; AVX512BW-LABEL: vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3:
-; AVX512BW:       # %bb.0:
-; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm0
-; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [0,13,14,15,0,1,2,3]
-; AVX512BW-NEXT:    vpaddb (%rsi), %zmm0, %zmm0
-; AVX512BW-NEXT:    vpermd %zmm0, %zmm1, %zmm1
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
-; AVX512BW-NEXT:    vpaddb (%rdx), %zmm0, %zmm0
-; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rcx)
-; AVX512BW-NEXT:    vzeroupper
-; AVX512BW-NEXT:    retq
+; AVX512BW-SLOW-LABEL: vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3:
+; AVX512BW-SLOW:       # %bb.0:
+; AVX512BW-SLOW-NEXT:    vmovdqa64 (%rdi), %zmm0
+; AVX512BW-SLOW-NEXT:    vpaddb (%rsi), %zmm0, %zmm0
+; AVX512BW-SLOW-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
+; AVX512BW-SLOW-NEXT:    vpermq {{.*#+}} ymm2 = ymm0[0,1,0,1]
+; AVX512BW-SLOW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1,2,3],ymm2[4,5,6,7]
----------------
RKSimon wrote:

I don't believe so - as I explained in the summary, we've reduced the length of the shuffle chain by an instruction, so it now falls below the threshold that we use for targets with slow cross-lane-variable shuffles. The fast version below is unchanged.

https://github.com/llvm/llvm-project/pull/111685