1; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=core2 | FileCheck %s 2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=btver2 | FileCheck %s --check-prefix=BTVER2 3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+use-recip-est,+avx -x86-recip-refinement-steps=2 | FileCheck %s --check-prefix=REFINE 4 5; If the target's divss/divps instructions are substantially 6; slower than rcpss/rcpps with a Newton-Raphson refinement, 7; we should generate the estimate sequence. 8 9; See PR21385 ( http://llvm.org/bugs/show_bug.cgi?id=21385 ) 10; for details about the accuracy, speed, and implementation 11; differences of x86 reciprocal estimates. 12 13define float @reciprocal_estimate(float %x) #0 { 14 %div = fdiv fast float 1.0, %x 15 ret float %div 16 17; CHECK-LABEL: reciprocal_estimate: 18; CHECK: movss 19; CHECK-NEXT: divss 20; CHECK-NEXT: movaps 21; CHECK-NEXT: retq 22 23; BTVER2-LABEL: reciprocal_estimate: 24; BTVER2: vrcpss 25; BTVER2: vmulss 26; BTVER2: vsubss 27; BTVER2: vmulss 28; BTVER2: vaddss 29; BTVER2-NEXT: retq 30 31; REFINE-LABEL: reciprocal_estimate: 32; REFINE: vrcpss 33; REFINE: vmulss 34; REFINE: vsubss 35; REFINE: vmulss 36; REFINE: vaddss 37; REFINE: vmulss 38; REFINE: vsubss 39; REFINE: vmulss 40; REFINE: vaddss 41; REFINE-NEXT: retq 42} 43 44define <4 x float> @reciprocal_estimate_v4f32(<4 x float> %x) #0 { 45 %div = fdiv fast <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, %x 46 ret <4 x float> %div 47 48; CHECK-LABEL: reciprocal_estimate_v4f32: 49; CHECK: movaps 50; CHECK-NEXT: divps 51; CHECK-NEXT: movaps 52; CHECK-NEXT: retq 53 54; BTVER2-LABEL: reciprocal_estimate_v4f32: 55; BTVER2: vrcpps 56; BTVER2: vmulps 57; BTVER2: vsubps 58; BTVER2: vmulps 59; BTVER2: vaddps 60; BTVER2-NEXT: retq 61 62; REFINE-LABEL: reciprocal_estimate_v4f32: 63; REFINE: vrcpps 64; REFINE: vmulps 65; REFINE: vsubps 66; REFINE: vmulps 67; REFINE: vaddps 68; REFINE: vmulps 69; REFINE: vsubps 70; REFINE: vmulps 71; REFINE: vaddps 72; REFINE-NEXT: retq 73} 74 75define <8 x float> @reciprocal_estimate_v8f32(<8 x float> %x) #0 { 76 %div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x 77 ret <8 x float> %div 78 79; CHECK-LABEL: reciprocal_estimate_v8f32: 80; CHECK: movaps 81; CHECK: movaps 82; CHECK-NEXT: divps 83; CHECK-NEXT: divps 84; CHECK-NEXT: movaps 85; CHECK-NEXT: movaps 86; CHECK-NEXT: retq 87 88; BTVER2-LABEL: reciprocal_estimate_v8f32: 89; BTVER2: vrcpps 90; BTVER2: vmulps 91; BTVER2: vsubps 92; BTVER2: vmulps 93; BTVER2: vaddps 94; BTVER2-NEXT: retq 95 96; REFINE-LABEL: reciprocal_estimate_v8f32: 97; REFINE: vrcpps 98; REFINE: vmulps 99; REFINE: vsubps 100; REFINE: vmulps 101; REFINE: vaddps 102; REFINE: vmulps 103; REFINE: vsubps 104; REFINE: vmulps 105; REFINE: vaddps 106; REFINE-NEXT: retq 107} 108 109attributes #0 = { "unsafe-fp-math"="true" } 110