65 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=skx --show-mc-encoding | FileCheck %s3 4define i32 @f256(<8 x float> %A, <8 x float> %AA, ptr %B, <4 x double> %C, <4 x double> %CC, i32 %D, <4 x i64> %E, <4 x i64> %EE, ptr %loadptr) {5; CHECK-LABEL: f256:6; CHECK: ## %bb.0:7; CHECK-NEXT: movl (%rdx), %eax ## encoding: [0x8b,0x02]8; CHECK-NEXT: vaddps %ymm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfc,0x58,0xc1]9; CHECK-NEXT: vmovntps %ymm0, (%rdi) ## EVEX TO VEX Compression encoding: [0xc5,0xfc,0x2b,0x07]10; CHECK-NEXT: vpaddq %ymm5, %ymm4, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xdd,0xd4,0xc5]11; CHECK-NEXT: addl (%rdx), %eax ## encoding: [0x03,0x02]12; CHECK-NEXT: vmovntdq %ymm0, (%rdi) ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xe7,0x07]13; CHECK-NEXT: vaddpd %ymm3, %ymm2, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xed,0x58,0xc3]14; CHECK-NEXT: addl (%rdx), %eax ## encoding: [0x03,0x02]15; CHECK-NEXT: vmovntpd %ymm0, (%rdi) ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0x2b,0x07]16; CHECK-NEXT: addl (%rdx), %eax ## encoding: [0x03,0x02]17; CHECK-NEXT: vzeroupper ## encoding: [0xc5,0xf8,0x77]18; CHECK-NEXT: retq ## encoding: [0xc3]19 %v0 = load i32, ptr %loadptr, align 120 %A2 = fadd <8 x float> %A, %AA21 store <8 x float> %A2, ptr %B, align 64, !nontemporal !022 %v1 = load i32, ptr %loadptr, align 123 %E2 = add <4 x i64> %E, %EE24 store <4 x i64> %E2, ptr %B, align 64, !nontemporal !025 %v2 = load i32, ptr %loadptr, align 126 %C2 = fadd <4 x double> %C, %CC27 store <4 x double> %C2, ptr %B, align 64, !nontemporal !028 %v3 = load i32, ptr %loadptr, align 129 %sum1 = add i32 %v0, %v130 %sum2 = add i32 %sum1, %v231 %sum3 = add i32 %sum2, %v332 ret i32 %sum333}34 35define i32 @f128(<4 x float> %A, <4 x float> %AA, ptr %B, <2 x double> %C, <2 x double> %CC, i32 %D, <2 x i64> %E, <2 x i64> %EE, ptr %loadptr) {36; CHECK-LABEL: f128:37; CHECK: ## %bb.0:38; CHECK-NEXT: movl (%rdx), %eax ## encoding: [0x8b,0x02]39; CHECK-NEXT: vaddps %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf8,0x58,0xc1]40; CHECK-NEXT: vmovntps %xmm0, (%rdi) ## EVEX TO VEX Compression encoding: [0xc5,0xf8,0x2b,0x07]41; CHECK-NEXT: vpaddq %xmm5, %xmm4, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xd9,0xd4,0xc5]42; CHECK-NEXT: addl (%rdx), %eax ## encoding: [0x03,0x02]43; CHECK-NEXT: vmovntdq %xmm0, (%rdi) ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xe7,0x07]44; CHECK-NEXT: vaddpd %xmm3, %xmm2, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xe9,0x58,0xc3]45; CHECK-NEXT: addl (%rdx), %eax ## encoding: [0x03,0x02]46; CHECK-NEXT: vmovntpd %xmm0, (%rdi) ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x2b,0x07]47; CHECK-NEXT: addl (%rdx), %eax ## encoding: [0x03,0x02]48; CHECK-NEXT: retq ## encoding: [0xc3]49 %v0 = load i32, ptr %loadptr, align 150 %A2 = fadd <4 x float> %A, %AA51 store <4 x float> %A2, ptr %B, align 64, !nontemporal !052 %v1 = load i32, ptr %loadptr, align 153 %E2 = add <2 x i64> %E, %EE54 store <2 x i64> %E2, ptr %B, align 64, !nontemporal !055 %v2 = load i32, ptr %loadptr, align 156 %C2 = fadd <2 x double> %C, %CC57 store <2 x double> %C2, ptr %B, align 64, !nontemporal !058 %v3 = load i32, ptr %loadptr, align 159 %sum1 = add i32 %v0, %v160 %sum2 = add i32 %sum1, %v261 %sum3 = add i32 %sum2, %v362 ret i32 %sum363}64!0 = !{i32 1}65