@@ -2011,15 +2011,14 @@ define void @vec384_i8_widen_to_i24_factor3_broadcast_to_v16i24_factor16(ptr %in
2011
2011
;
2012
2012
; AVX2-LABEL: vec384_i8_widen_to_i24_factor3_broadcast_to_v16i24_factor16:
2013
2013
; AVX2: # %bb.0:
2014
- ; AVX2-NEXT: vpbroadcastb (%rdi), %xmm0
2015
- ; AVX2-NEXT: vmovdqa (%rdi), % xmm1
2016
- ; AVX2-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm1[0],zero,zero,xmm1[0],zero,zero,xmm1[0],zero,zero,xmm1[0],zero,zero,xmm1[0],zero
2017
- ; AVX2-NEXT: vinserti128 $1, %xmm2 , %ymm0, %ymm0
2014
+ ; AVX2-NEXT: vmovdqa (%rdi), %xmm0
2015
+ ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm0[0],zero,zero,xmm0[0],zero,zero,xmm0[0],zero,zero,xmm0[0],zero,zero,xmm0[0],zero,zero
2016
+ ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
2017
+ ; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip) , %ymm0, %ymm0
2018
2018
; AVX2-NEXT: vperm2i128 {{.*#+}} ymm2 = mem[2,3],ymm0[2,3]
2019
2019
; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,255,255,0,255,255,0,255,255,0,255,255,0,255,255,0,0,255,255,0,255,255,0,255,255,0,255,255,0,255,255,0]
2020
2020
; AVX2-NEXT: # ymm3 = mem[0,1,0,1]
2021
2021
; AVX2-NEXT: vpblendvb %ymm3, %ymm2, %ymm0, %ymm0
2022
- ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm1[0],zero,zero,xmm1[0],zero,zero,xmm1[0],zero,zero,xmm1[0],zero,zero,xmm1[0],zero,zero
2023
2022
; AVX2-NEXT: vpaddb (%rsi), %ymm0, %ymm0
2024
2023
; AVX2-NEXT: vpaddb 32(%rsi), %ymm1, %ymm1
2025
2024
; AVX2-NEXT: vmovdqa %ymm1, 32(%rdx)
@@ -2263,15 +2262,14 @@ define void @vec384_i8_widen_to_i48_factor6_broadcast_to_v8i48_factor8(ptr %in.e
2263
2262
;
2264
2263
; AVX2-LABEL: vec384_i8_widen_to_i48_factor6_broadcast_to_v8i48_factor8:
2265
2264
; AVX2: # %bb.0:
2266
- ; AVX2-NEXT: vpbroadcastb (%rdi), %xmm0
2267
- ; AVX2-NEXT: vmovdqa (%rdi), % xmm1
2268
- ; AVX2-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm1[0],zero,zero,zero,zero,zero,xmm1[0],zero,zero,zero,zero,zero,xmm1[0],zero
2269
- ; AVX2-NEXT: vinserti128 $1, %xmm2 , %ymm0, %ymm0
2265
+ ; AVX2-NEXT: vmovdqa (%rdi), %xmm0
2266
+ ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,xmm0[0],zero,zero,zero,zero,zero,xmm0[0],zero,zero,zero,zero,zero
2267
+ ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
2268
+ ; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip) , %ymm0, %ymm0
2270
2269
; AVX2-NEXT: vperm2i128 {{.*#+}} ymm2 = mem[2,3],ymm0[2,3]
2271
2270
; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255]
2272
2271
; AVX2-NEXT: # ymm3 = mem[0,1,0,1]
2273
2272
; AVX2-NEXT: vpblendvb %ymm3, %ymm2, %ymm0, %ymm0
2274
- ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,xmm1[0],zero,zero,zero,zero,zero,xmm1[0],zero,zero,zero,zero,zero
2275
2273
; AVX2-NEXT: vpaddb (%rsi), %ymm0, %ymm0
2276
2274
; AVX2-NEXT: vpaddb 32(%rsi), %ymm1, %ymm1
2277
2275
; AVX2-NEXT: vmovdqa %ymm1, 32(%rdx)
@@ -2515,15 +2513,14 @@ define void @vec384_i8_widen_to_i96_factor12_broadcast_to_v4i96_factor4(ptr %in.
2515
2513
;
2516
2514
; AVX2-LABEL: vec384_i8_widen_to_i96_factor12_broadcast_to_v4i96_factor4:
2517
2515
; AVX2: # %bb.0:
2518
- ; AVX2-NEXT: vpbroadcastb (%rdi), %xmm0
2519
- ; AVX2-NEXT: vmovdqa (%rdi), % xmm1
2520
- ; AVX2-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,zero,zero,zero,zero,zero,zero,xmm1[0],zero,zero,zero,zero,zero,zero,zero
2521
- ; AVX2-NEXT: vinserti128 $1, %xmm2 , %ymm0, %ymm0
2516
+ ; AVX2-NEXT: vmovdqa (%rdi), %xmm0
2517
+ ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,xmm0[0],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
2518
+ ; AVX2-NEXT: vpbroadcastb %xmm0, %ymm0
2519
+ ; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip) , %ymm0, %ymm0
2522
2520
; AVX2-NEXT: vperm2i128 {{.*#+}} ymm2 = mem[2,3],ymm0[2,3]
2523
2521
; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,255,255,255,255,255,255,255,255,255,255,255,0,255,255,255,0,255,255,255,255,255,255,255,255,255,255,255,0,255,255,255]
2524
2522
; AVX2-NEXT: # ymm3 = mem[0,1,0,1]
2525
2523
; AVX2-NEXT: vpblendvb %ymm3, %ymm2, %ymm0, %ymm0
2526
- ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,xmm1[0],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
2527
2524
; AVX2-NEXT: vpaddb (%rsi), %ymm0, %ymm0
2528
2525
; AVX2-NEXT: vpaddb 32(%rsi), %ymm1, %ymm1
2529
2526
; AVX2-NEXT: vmovdqa %ymm1, 32(%rdx)
0 commit comments