142 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64 -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -mattr=+avx512fp16,+avx512vl | FileCheck %s --check-prefix=CHECK_UNSAFE3; RUN: llc < %s -mtriple=x86_64 -mattr=+avx512fp16,+avx512vl | FileCheck %s --check-prefix=CHECK4 5define <32 x half> @test_max_v32f16(ptr %a_ptr, <32 x half> %b) {6; CHECK_UNSAFE-LABEL: test_max_v32f16:7; CHECK_UNSAFE: # %bb.0:8; CHECK_UNSAFE-NEXT: vmaxph (%rdi), %zmm0, %zmm09; CHECK_UNSAFE-NEXT: retq10;11; CHECK-LABEL: test_max_v32f16:12; CHECK: # %bb.0:13; CHECK-NEXT: vmovaps (%rdi), %zmm114; CHECK-NEXT: vmaxph %zmm0, %zmm1, %zmm015; CHECK-NEXT: retq16 %a = load <32 x half>, ptr %a_ptr17 %tmp = fcmp fast ogt <32 x half> %a, %b18 %tmp4 = select <32 x i1> %tmp, <32 x half> %a, <32 x half> %b19 ret <32 x half> %tmp4;20}21 22define <32 x half> @test_min_v32f16(ptr %a_ptr, <32 x half> %b) {23; CHECK_UNSAFE-LABEL: test_min_v32f16:24; CHECK_UNSAFE: # %bb.0:25; CHECK_UNSAFE-NEXT: vminph (%rdi), %zmm0, %zmm026; CHECK_UNSAFE-NEXT: retq27;28; CHECK-LABEL: test_min_v32f16:29; CHECK: # %bb.0:30; CHECK-NEXT: vmovaps (%rdi), %zmm131; CHECK-NEXT: vminph %zmm0, %zmm1, %zmm032; CHECK-NEXT: retq33 %a = load <32 x half>, ptr %a_ptr34 %tmp = fcmp fast olt <32 x half> %a, %b35 %tmp4 = select <32 x i1> %tmp, <32 x half> %a, <32 x half> %b36 ret <32 x half> %tmp4;37}38 39define <16 x half> @test_max_v16f16(ptr %a_ptr, <16 x half> %b) {40; CHECK_UNSAFE-LABEL: test_max_v16f16:41; CHECK_UNSAFE: # %bb.0:42; CHECK_UNSAFE-NEXT: vmaxph (%rdi), %ymm0, %ymm043; CHECK_UNSAFE-NEXT: retq44;45; CHECK-LABEL: test_max_v16f16:46; CHECK: # %bb.0:47; CHECK-NEXT: vmovaps (%rdi), %ymm148; CHECK-NEXT: vmaxph %ymm0, %ymm1, %ymm049; CHECK-NEXT: retq50 %a = load <16 x half>, ptr %a_ptr51 %tmp = fcmp fast ogt <16 x half> %a, %b52 %tmp4 = select <16 x i1> %tmp, <16 x half> %a, <16 x half> %b53 ret <16 x half> %tmp4;54}55 56define <16 x half> @test_min_v16f16(ptr %a_ptr, <16 x half> %b) {57; CHECK_UNSAFE-LABEL: test_min_v16f16:58; CHECK_UNSAFE: # %bb.0:59; CHECK_UNSAFE-NEXT: vminph (%rdi), %ymm0, %ymm060; CHECK_UNSAFE-NEXT: retq61;62; CHECK-LABEL: test_min_v16f16:63; CHECK: # %bb.0:64; CHECK-NEXT: vmovaps (%rdi), %ymm165; CHECK-NEXT: vminph %ymm0, %ymm1, %ymm066; CHECK-NEXT: retq67 %a = load <16 x half>, ptr %a_ptr68 %tmp = fcmp fast olt <16 x half> %a, %b69 %tmp4 = select <16 x i1> %tmp, <16 x half> %a, <16 x half> %b70 ret <16 x half> %tmp4;71}72 73define <8 x half> @test_max_v8f16(ptr %a_ptr, <8 x half> %b) {74; CHECK_UNSAFE-LABEL: test_max_v8f16:75; CHECK_UNSAFE: # %bb.0:76; CHECK_UNSAFE-NEXT: vmaxph (%rdi), %xmm0, %xmm077; CHECK_UNSAFE-NEXT: retq78;79; CHECK-LABEL: test_max_v8f16:80; CHECK: # %bb.0:81; CHECK-NEXT: vmovaps (%rdi), %xmm182; CHECK-NEXT: vmaxph %xmm0, %xmm1, %xmm083; CHECK-NEXT: retq84 %a = load <8 x half>, ptr %a_ptr85 %tmp = fcmp fast ogt <8 x half> %a, %b86 %tmp4 = select <8 x i1> %tmp, <8 x half> %a, <8 x half> %b87 ret <8 x half> %tmp4;88}89 90define <8 x half> @test_min_v8f16(ptr %a_ptr, <8 x half> %b) {91; CHECK_UNSAFE-LABEL: test_min_v8f16:92; CHECK_UNSAFE: # %bb.0:93; CHECK_UNSAFE-NEXT: vminph (%rdi), %xmm0, %xmm094; CHECK_UNSAFE-NEXT: retq95;96; CHECK-LABEL: test_min_v8f16:97; CHECK: # %bb.0:98; CHECK-NEXT: vmovaps (%rdi), %xmm199; CHECK-NEXT: vminph %xmm0, %xmm1, %xmm0100; CHECK-NEXT: retq101 %a = load <8 x half>, ptr %a_ptr102 %tmp = fcmp fast olt <8 x half> %a, %b103 %tmp4 = select <8 x i1> %tmp, <8 x half> %a, <8 x half> %b104 ret <8 x half> %tmp4;105}106 107define half @test_max_f16(half %a, ptr %ptr) {108; CHECK_UNSAFE-LABEL: test_max_f16:109; CHECK_UNSAFE: # %bb.0: # %entry110; CHECK_UNSAFE-NEXT: vmaxsh (%rdi), %xmm0, %xmm0111; CHECK_UNSAFE-NEXT: retq112;113; CHECK-LABEL: test_max_f16:114; CHECK: # %bb.0: # %entry115; CHECK-NEXT: vmovsh {{.*#+}} xmm1 = mem[0],zero,zero,zero,zero,zero,zero,zero116; CHECK-NEXT: vmaxsh %xmm0, %xmm1, %xmm0117; CHECK-NEXT: retq118entry:119 %0 = load half, ptr %ptr120 %1 = fcmp fast ogt half %0, %a121 %2 = select i1 %1, half %0, half %a122 ret half %2123}124 125define half @test_min_f16(half %a, ptr %ptr) {126; CHECK_UNSAFE-LABEL: test_min_f16:127; CHECK_UNSAFE: # %bb.0: # %entry128; CHECK_UNSAFE-NEXT: vminsh (%rdi), %xmm0, %xmm0129; CHECK_UNSAFE-NEXT: retq130;131; CHECK-LABEL: test_min_f16:132; CHECK: # %bb.0: # %entry133; CHECK-NEXT: vmovsh {{.*#+}} xmm1 = mem[0],zero,zero,zero,zero,zero,zero,zero134; CHECK-NEXT: vminsh %xmm0, %xmm1, %xmm0135; CHECK-NEXT: retq136entry:137 %0 = load half, ptr %ptr138 %1 = fcmp fast olt half %0, %a139 %2 = select i1 %1, half %0, half %a140 ret half %2141}142