From 09a127acbc02234b2f2f43018f655d90f2cf343f Mon Sep 17 00:00:00 2001 From: Craig Topper Date: Mon, 6 Nov 2017 05:48:23 +0000 Subject: [PATCH] [X86] Simplify command lines on the fma-instrinsics-x86.ll test and add -show-mc-encoding. Use feature names instead of CPU names. A future commit will add avx512vl command lines to demonstrate missed use of EVEX instructions. llvm-svn: 317451 --- test/CodeGen/X86/fma-intrinsics-x86.ll | 665 ++++++++++++------------- 1 file changed, 331 insertions(+), 334 deletions(-) diff --git a/test/CodeGen/X86/fma-intrinsics-x86.ll b/test/CodeGen/X86/fma-intrinsics-x86.ll index 68f39469a82..36d487c6611 100644 --- a/test/CodeGen/X86/fma-intrinsics-x86.ll +++ b/test/CodeGen/X86/fma-intrinsics-x86.ll @@ -1,29 +1,26 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=corei7-avx -mattr=+fma | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FMA -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=core-avx2 -mattr=+fma,+avx2 | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FMA -; RUN: llc < %s -mtriple=x86_64-pc-windows -mcpu=core-avx2 -mattr=+fma,+avx2 | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FMA-WIN -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=corei7-avx -mattr=+fma4 | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FMA4 -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=bdver2 -mattr=+avx,-fma | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FMA4 -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=bdver2 -mattr=-fma4 | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FMA +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+fma,-fma4 -show-mc-encoding | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FMA +; RUN: llc < %s -mtriple=x86_64-pc-windows -mattr=+fma,-fma4 -show-mc-encoding | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FMA-WIN +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+fma4,-fma -show-mc-encoding | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-FMA4 ; VFMADD define <4 x float> @test_x86_fma_vfmadd_ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmadd_ss: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmadd213ss %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmadd213ss %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xa9,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmadd_ss: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmadd213ss (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x0a] +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x01] +; CHECK-FMA-WIN-NEXT: vfmadd213ss (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xa9,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmadd_ss: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddss %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddss %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x6a,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfmadd.ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) ret <4 x float> %res } @@ -31,21 +28,21 @@ define <4 x float> @test_x86_fma_vfmadd_ss(<4 x float> %a0, <4 x float> %a1, <4 define <4 x float> @test_x86_fma_vfmadd_bac_ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmadd_bac_ss: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmadd213ss %xmm2, %xmm0, %xmm1 -; CHECK-FMA-NEXT: vmovaps %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmadd213ss %xmm2, %xmm0, %xmm1 # encoding: [0xc4,0xe2,0x79,0xa9,0xca] +; CHECK-FMA-NEXT: vmovaps %xmm1, %xmm0 # encoding: [0xc5,0xf8,0x28,0xc1] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmadd_bac_ss: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmadd213ss (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmadd213ss (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xa9,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmadd_bac_ss: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddss %xmm2, %xmm0, %xmm1, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddss %xmm2, %xmm0, %xmm1, %xmm0 # encoding: [0xc4,0xe3,0xf1,0x6a,0xc2,0x00] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfmadd.ss(<4 x float> %a1, <4 x float> %a0, <4 x float> %a2) ret <4 x float> %res } @@ -54,20 +51,20 @@ declare <4 x float> @llvm.x86.fma.vfmadd.ss(<4 x float>, <4 x float>, <4 x float define <2 x double> @test_x86_fma_vfmadd_sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmadd_sd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmadd213sd %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmadd213sd %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0xf1,0xa9,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmadd_sd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmadd213sd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x0a] +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x01] +; CHECK-FMA-WIN-NEXT: vfmadd213sd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xa9,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmadd_sd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddsd %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddsd %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x6b,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfmadd.sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) ret <2 x double> %res } @@ -75,21 +72,21 @@ define <2 x double> @test_x86_fma_vfmadd_sd(<2 x double> %a0, <2 x double> %a1, define <2 x double> @test_x86_fma_vfmadd_bac_sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmadd_bac_sd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmadd213sd %xmm2, %xmm0, %xmm1 -; CHECK-FMA-NEXT: vmovapd %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmadd213sd %xmm2, %xmm0, %xmm1 # encoding: [0xc4,0xe2,0xf9,0xa9,0xca] +; CHECK-FMA-NEXT: vmovapd %xmm1, %xmm0 # encoding: [0xc5,0xf9,0x28,0xc1] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmadd_bac_sd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmadd213sd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmadd213sd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xa9,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmadd_bac_sd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddsd %xmm2, %xmm0, %xmm1, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddsd %xmm2, %xmm0, %xmm1, %xmm0 # encoding: [0xc4,0xe3,0xf1,0x6b,0xc2,0x00] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfmadd.sd(<2 x double> %a1, <2 x double> %a0, <2 x double> %a2) ret <2 x double> %res } @@ -98,20 +95,20 @@ declare <2 x double> @llvm.x86.fma.vfmadd.sd(<2 x double>, <2 x double>, <2 x do define <4 x float> @test_x86_fma_vfmadd_ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmadd_ps: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmadd213ps %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmadd213ps %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xa8,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmadd_ps: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmadd213ps (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmadd213ps (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xa8,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmadd_ps: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddps %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddps %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x68,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfmadd.ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) ret <4 x float> %res } @@ -120,20 +117,20 @@ declare <4 x float> @llvm.x86.fma.vfmadd.ps(<4 x float>, <4 x float>, <4 x float define <2 x double> @test_x86_fma_vfmadd_pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmadd_pd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmadd213pd %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmadd213pd %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0xf1,0xa8,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmadd_pd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmadd213pd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmadd213pd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xa8,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmadd_pd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddpd %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddpd %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x69,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfmadd.pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) ret <2 x double> %res } @@ -142,20 +139,20 @@ declare <2 x double> @llvm.x86.fma.vfmadd.pd(<2 x double>, <2 x double>, <2 x do define <8 x float> @test_x86_fma_vfmadd_ps_256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmadd_ps_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmadd213ps %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmadd213ps %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0x75,0xa8,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmadd_ps_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfmadd213ps (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 # encoding: [0xc5,0xfc,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 # encoding: [0xc5,0xfc,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmadd213ps (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0x75,0xa8,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmadd_ps_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddps %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddps %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x68,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <8 x float> @llvm.x86.fma.vfmadd.ps.256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) ret <8 x float> %res } @@ -164,20 +161,20 @@ declare <8 x float> @llvm.x86.fma.vfmadd.ps.256(<8 x float>, <8 x float>, <8 x f define <4 x double> @test_x86_fma_vfmadd_pd_256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmadd_pd_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmadd213pd %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmadd213pd %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0xf5,0xa8,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmadd_pd_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfmadd213pd (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 # encoding: [0xc5,0xfd,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 # encoding: [0xc5,0xfd,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmadd213pd (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0xf5,0xa8,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmadd_pd_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddpd %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddpd %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x69,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x double> @llvm.x86.fma.vfmadd.pd.256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) ret <4 x double> %res } @@ -187,20 +184,20 @@ declare <4 x double> @llvm.x86.fma.vfmadd.pd.256(<4 x double>, <4 x double>, <4 define <4 x float> @test_x86_fma_vfmsub_ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsub_ss: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsub213ss %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsub213ss %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xab,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsub_ss: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmsub213ss (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x0a] +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x01] +; CHECK-FMA-WIN-NEXT: vfmsub213ss (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xab,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsub_ss: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubss %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubss %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x6e,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfmsub.ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) ret <4 x float> %res } @@ -208,21 +205,21 @@ define <4 x float> @test_x86_fma_vfmsub_ss(<4 x float> %a0, <4 x float> %a1, <4 define <4 x float> @test_x86_fma_vfmsub_bac_ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsub_bac_ss: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsub213ss %xmm2, %xmm0, %xmm1 -; CHECK-FMA-NEXT: vmovaps %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsub213ss %xmm2, %xmm0, %xmm1 # encoding: [0xc4,0xe2,0x79,0xab,0xca] +; CHECK-FMA-NEXT: vmovaps %xmm1, %xmm0 # encoding: [0xc5,0xf8,0x28,0xc1] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsub_bac_ss: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmsub213ss (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmsub213ss (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xab,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsub_bac_ss: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubss %xmm2, %xmm0, %xmm1, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubss %xmm2, %xmm0, %xmm1, %xmm0 # encoding: [0xc4,0xe3,0xf1,0x6e,0xc2,0x00] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfmsub.ss(<4 x float> %a1, <4 x float> %a0, <4 x float> %a2) ret <4 x float> %res } @@ -231,20 +228,20 @@ declare <4 x float> @llvm.x86.fma.vfmsub.ss(<4 x float>, <4 x float>, <4 x float define <2 x double> @test_x86_fma_vfmsub_sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsub_sd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsub213sd %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsub213sd %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0xf1,0xab,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsub_sd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmsub213sd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x0a] +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x01] +; CHECK-FMA-WIN-NEXT: vfmsub213sd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xab,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsub_sd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubsd %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubsd %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x6f,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfmsub.sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) ret <2 x double> %res } @@ -252,21 +249,21 @@ define <2 x double> @test_x86_fma_vfmsub_sd(<2 x double> %a0, <2 x double> %a1, define <2 x double> @test_x86_fma_vfmsub_bac_sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsub_bac_sd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsub213sd %xmm2, %xmm0, %xmm1 -; CHECK-FMA-NEXT: vmovapd %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsub213sd %xmm2, %xmm0, %xmm1 # encoding: [0xc4,0xe2,0xf9,0xab,0xca] +; CHECK-FMA-NEXT: vmovapd %xmm1, %xmm0 # encoding: [0xc5,0xf9,0x28,0xc1] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsub_bac_sd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmsub213sd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmsub213sd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xab,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsub_bac_sd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubsd %xmm2, %xmm0, %xmm1, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubsd %xmm2, %xmm0, %xmm1, %xmm0 # encoding: [0xc4,0xe3,0xf1,0x6f,0xc2,0x00] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfmsub.sd(<2 x double> %a1, <2 x double> %a0, <2 x double> %a2) ret <2 x double> %res } @@ -275,20 +272,20 @@ declare <2 x double> @llvm.x86.fma.vfmsub.sd(<2 x double>, <2 x double>, <2 x do define <4 x float> @test_x86_fma_vfmsub_ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsub_ps: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsub213ps %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsub213ps %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xaa,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsub_ps: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmsub213ps (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmsub213ps (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xaa,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsub_ps: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubps %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubps %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x6c,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfmsub.ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) ret <4 x float> %res } @@ -297,20 +294,20 @@ declare <4 x float> @llvm.x86.fma.vfmsub.ps(<4 x float>, <4 x float>, <4 x float define <2 x double> @test_x86_fma_vfmsub_pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsub_pd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsub213pd %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsub213pd %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0xf1,0xaa,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsub_pd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmsub213pd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmsub213pd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xaa,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsub_pd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubpd %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubpd %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x6d,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfmsub.pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) ret <2 x double> %res } @@ -319,20 +316,20 @@ declare <2 x double> @llvm.x86.fma.vfmsub.pd(<2 x double>, <2 x double>, <2 x do define <8 x float> @test_x86_fma_vfmsub_ps_256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsub_ps_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsub213ps %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsub213ps %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0x75,0xaa,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsub_ps_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfmsub213ps (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 # encoding: [0xc5,0xfc,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 # encoding: [0xc5,0xfc,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmsub213ps (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0x75,0xaa,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsub_ps_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubps %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubps %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x6c,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <8 x float> @llvm.x86.fma.vfmsub.ps.256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) ret <8 x float> %res } @@ -341,20 +338,20 @@ declare <8 x float> @llvm.x86.fma.vfmsub.ps.256(<8 x float>, <8 x float>, <8 x f define <4 x double> @test_x86_fma_vfmsub_pd_256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsub_pd_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsub213pd %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsub213pd %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0xf5,0xaa,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsub_pd_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfmsub213pd (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 # encoding: [0xc5,0xfd,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 # encoding: [0xc5,0xfd,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmsub213pd (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0xf5,0xaa,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsub_pd_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubpd %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubpd %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x6d,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x double> @llvm.x86.fma.vfmsub.pd.256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) ret <4 x double> %res } @@ -364,20 +361,20 @@ declare <4 x double> @llvm.x86.fma.vfmsub.pd.256(<4 x double>, <4 x double>, <4 define <4 x float> @test_x86_fma_vfnmadd_ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmadd_ss: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmadd213ss %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmadd213ss %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xad,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmadd_ss: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmadd213ss (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x0a] +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x01] +; CHECK-FMA-WIN-NEXT: vfnmadd213ss (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xad,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmadd_ss: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmaddss %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmaddss %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x7a,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfnmadd.ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) ret <4 x float> %res } @@ -385,21 +382,21 @@ define <4 x float> @test_x86_fma_vfnmadd_ss(<4 x float> %a0, <4 x float> %a1, <4 define <4 x float> @test_x86_fma_vfnmadd_bac_ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmadd_bac_ss: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmadd213ss %xmm2, %xmm0, %xmm1 -; CHECK-FMA-NEXT: vmovaps %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmadd213ss %xmm2, %xmm0, %xmm1 # encoding: [0xc4,0xe2,0x79,0xad,0xca] +; CHECK-FMA-NEXT: vmovaps %xmm1, %xmm0 # encoding: [0xc5,0xf8,0x28,0xc1] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmadd_bac_ss: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmadd213ss (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmadd213ss (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xad,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmadd_bac_ss: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmaddss %xmm2, %xmm0, %xmm1, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmaddss %xmm2, %xmm0, %xmm1, %xmm0 # encoding: [0xc4,0xe3,0xf1,0x7a,0xc2,0x00] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfnmadd.ss(<4 x float> %a1, <4 x float> %a0, <4 x float> %a2) ret <4 x float> %res } @@ -408,20 +405,20 @@ declare <4 x float> @llvm.x86.fma.vfnmadd.ss(<4 x float>, <4 x float>, <4 x floa define <2 x double> @test_x86_fma_vfnmadd_sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmadd_sd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmadd213sd %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmadd213sd %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0xf1,0xad,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmadd_sd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmadd213sd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x0a] +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x01] +; CHECK-FMA-WIN-NEXT: vfnmadd213sd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xad,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmadd_sd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmaddsd %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmaddsd %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x7b,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfnmadd.sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) ret <2 x double> %res } @@ -429,21 +426,21 @@ define <2 x double> @test_x86_fma_vfnmadd_sd(<2 x double> %a0, <2 x double> %a1, define <2 x double> @test_x86_fma_vfnmadd_bac_sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmadd_bac_sd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmadd213sd %xmm2, %xmm0, %xmm1 -; CHECK-FMA-NEXT: vmovapd %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmadd213sd %xmm2, %xmm0, %xmm1 # encoding: [0xc4,0xe2,0xf9,0xad,0xca] +; CHECK-FMA-NEXT: vmovapd %xmm1, %xmm0 # encoding: [0xc5,0xf9,0x28,0xc1] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmadd_bac_sd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmadd213sd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmadd213sd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xad,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmadd_bac_sd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmaddsd %xmm2, %xmm0, %xmm1, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmaddsd %xmm2, %xmm0, %xmm1, %xmm0 # encoding: [0xc4,0xe3,0xf1,0x7b,0xc2,0x00] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfnmadd.sd(<2 x double> %a1, <2 x double> %a0, <2 x double> %a2) ret <2 x double> %res } @@ -452,20 +449,20 @@ declare <2 x double> @llvm.x86.fma.vfnmadd.sd(<2 x double>, <2 x double>, <2 x d define <4 x float> @test_x86_fma_vfnmadd_ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmadd_ps: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmadd213ps %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmadd213ps %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xac,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmadd_ps: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmadd213ps (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmadd213ps (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xac,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmadd_ps: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmaddps %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmaddps %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x78,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfnmadd.ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) ret <4 x float> %res } @@ -474,20 +471,20 @@ declare <4 x float> @llvm.x86.fma.vfnmadd.ps(<4 x float>, <4 x float>, <4 x floa define <2 x double> @test_x86_fma_vfnmadd_pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmadd_pd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmadd213pd %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmadd213pd %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0xf1,0xac,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmadd_pd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmadd213pd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmadd213pd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xac,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmadd_pd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmaddpd %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmaddpd %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x79,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfnmadd.pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) ret <2 x double> %res } @@ -496,20 +493,20 @@ declare <2 x double> @llvm.x86.fma.vfnmadd.pd(<2 x double>, <2 x double>, <2 x d define <8 x float> @test_x86_fma_vfnmadd_ps_256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmadd_ps_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmadd213ps %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmadd213ps %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0x75,0xac,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmadd_ps_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfnmadd213ps (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 # encoding: [0xc5,0xfc,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 # encoding: [0xc5,0xfc,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmadd213ps (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0x75,0xac,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmadd_ps_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmaddps %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmaddps %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x78,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <8 x float> @llvm.x86.fma.vfnmadd.ps.256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) ret <8 x float> %res } @@ -518,20 +515,20 @@ declare <8 x float> @llvm.x86.fma.vfnmadd.ps.256(<8 x float>, <8 x float>, <8 x define <4 x double> @test_x86_fma_vfnmadd_pd_256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmadd_pd_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmadd213pd %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmadd213pd %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0xf5,0xac,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmadd_pd_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfnmadd213pd (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 # encoding: [0xc5,0xfd,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 # encoding: [0xc5,0xfd,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmadd213pd (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0xf5,0xac,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmadd_pd_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmaddpd %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmaddpd %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x79,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x double> @llvm.x86.fma.vfnmadd.pd.256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) ret <4 x double> %res } @@ -541,20 +538,20 @@ declare <4 x double> @llvm.x86.fma.vfnmadd.pd.256(<4 x double>, <4 x double>, <4 define <4 x float> @test_x86_fma_vfnmsub_ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmsub_ss: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmsub213ss %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmsub213ss %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xaf,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmsub_ss: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmsub213ss (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x0a] +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x01] +; CHECK-FMA-WIN-NEXT: vfnmsub213ss (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xaf,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmsub_ss: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmsubss %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmsubss %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x7e,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfnmsub.ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) ret <4 x float> %res } @@ -562,21 +559,21 @@ define <4 x float> @test_x86_fma_vfnmsub_ss(<4 x float> %a0, <4 x float> %a1, <4 define <4 x float> @test_x86_fma_vfnmsub_bac_ss(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmsub_bac_ss: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmsub213ss %xmm2, %xmm0, %xmm1 -; CHECK-FMA-NEXT: vmovaps %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmsub213ss %xmm2, %xmm0, %xmm1 # encoding: [0xc4,0xe2,0x79,0xaf,0xca] +; CHECK-FMA-NEXT: vmovaps %xmm1, %xmm0 # encoding: [0xc5,0xf8,0x28,0xc1] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmsub_bac_ss: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmsub213ss (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmsub213ss (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xaf,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmsub_bac_ss: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmsubss %xmm2, %xmm0, %xmm1, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmsubss %xmm2, %xmm0, %xmm1, %xmm0 # encoding: [0xc4,0xe3,0xf1,0x7e,0xc2,0x00] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfnmsub.ss(<4 x float> %a1, <4 x float> %a0, <4 x float> %a2) ret <4 x float> %res } @@ -585,20 +582,20 @@ declare <4 x float> @llvm.x86.fma.vfnmsub.ss(<4 x float>, <4 x float>, <4 x floa define <2 x double> @test_x86_fma_vfnmsub_sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmsub_sd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmsub213sd %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmsub213sd %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0xf1,0xaf,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmsub_sd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmsub213sd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x0a] +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x01] +; CHECK-FMA-WIN-NEXT: vfnmsub213sd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xaf,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmsub_sd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmsubsd %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmsubsd %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x7f,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfnmsub.sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) ret <2 x double> %res } @@ -606,21 +603,21 @@ define <2 x double> @test_x86_fma_vfnmsub_sd(<2 x double> %a0, <2 x double> %a1, define <2 x double> @test_x86_fma_vfnmsub_bac_sd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmsub_bac_sd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmsub213sd %xmm2, %xmm0, %xmm1 -; CHECK-FMA-NEXT: vmovapd %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmsub213sd %xmm2, %xmm0, %xmm1 # encoding: [0xc4,0xe2,0xf9,0xaf,0xca] +; CHECK-FMA-NEXT: vmovapd %xmm1, %xmm0 # encoding: [0xc5,0xf9,0x28,0xc1] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmsub_bac_sd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmsub213sd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmsub213sd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xaf,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmsub_bac_sd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmsubsd %xmm2, %xmm0, %xmm1, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmsubsd %xmm2, %xmm0, %xmm1, %xmm0 # encoding: [0xc4,0xe3,0xf1,0x7f,0xc2,0x00] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfnmsub.sd(<2 x double> %a1, <2 x double> %a0, <2 x double> %a2) ret <2 x double> %res } @@ -629,20 +626,20 @@ declare <2 x double> @llvm.x86.fma.vfnmsub.sd(<2 x double>, <2 x double>, <2 x d define <4 x float> @test_x86_fma_vfnmsub_ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmsub_ps: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmsub213ps %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmsub213ps %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xae,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmsub_ps: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmsub213ps (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmsub213ps (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xae,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmsub_ps: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmsubps %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmsubps %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x7c,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfnmsub.ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) ret <4 x float> %res } @@ -651,20 +648,20 @@ declare <4 x float> @llvm.x86.fma.vfnmsub.ps(<4 x float>, <4 x float>, <4 x floa define <2 x double> @test_x86_fma_vfnmsub_pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmsub_pd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmsub213pd %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmsub213pd %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0xf1,0xae,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmsub_pd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfnmsub213pd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmsub213pd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xae,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmsub_pd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmsubpd %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmsubpd %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x7d,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfnmsub.pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) ret <2 x double> %res } @@ -673,20 +670,20 @@ declare <2 x double> @llvm.x86.fma.vfnmsub.pd(<2 x double>, <2 x double>, <2 x d define <8 x float> @test_x86_fma_vfnmsub_ps_256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmsub_ps_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmsub213ps %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmsub213ps %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0x75,0xae,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmsub_ps_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfnmsub213ps (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 # encoding: [0xc5,0xfc,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 # encoding: [0xc5,0xfc,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmsub213ps (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0x75,0xae,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmsub_ps_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmsubps %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmsubps %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x7c,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <8 x float> @llvm.x86.fma.vfnmsub.ps.256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) ret <8 x float> %res } @@ -695,20 +692,20 @@ declare <8 x float> @llvm.x86.fma.vfnmsub.ps.256(<8 x float>, <8 x float>, <8 x define <4 x double> @test_x86_fma_vfnmsub_pd_256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfnmsub_pd_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfnmsub213pd %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfnmsub213pd %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0xf5,0xae,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfnmsub_pd_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfnmsub213pd (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 # encoding: [0xc5,0xfd,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 # encoding: [0xc5,0xfd,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfnmsub213pd (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0xf5,0xae,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfnmsub_pd_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfnmsubpd %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfnmsubpd %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x7d,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x double> @llvm.x86.fma.vfnmsub.pd.256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) ret <4 x double> %res } @@ -718,20 +715,20 @@ declare <4 x double> @llvm.x86.fma.vfnmsub.pd.256(<4 x double>, <4 x double>, <4 define <4 x float> @test_x86_fma_vfmaddsub_ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmaddsub_ps: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmaddsub213ps %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmaddsub213ps %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xa6,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmaddsub_ps: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmaddsub213ps (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmaddsub213ps (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xa6,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmaddsub_ps: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddsubps %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddsubps %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x5c,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfmaddsub.ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) ret <4 x float> %res } @@ -740,20 +737,20 @@ declare <4 x float> @llvm.x86.fma.vfmaddsub.ps(<4 x float>, <4 x float>, <4 x fl define <2 x double> @test_x86_fma_vfmaddsub_pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmaddsub_pd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmaddsub213pd %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmaddsub213pd %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0xf1,0xa6,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmaddsub_pd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmaddsub213pd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmaddsub213pd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xa6,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmaddsub_pd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddsubpd %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddsubpd %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x5d,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfmaddsub.pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) ret <2 x double> %res } @@ -762,20 +759,20 @@ declare <2 x double> @llvm.x86.fma.vfmaddsub.pd(<2 x double>, <2 x double>, <2 x define <8 x float> @test_x86_fma_vfmaddsub_ps_256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmaddsub_ps_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmaddsub213ps %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmaddsub213ps %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0x75,0xa6,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmaddsub_ps_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfmaddsub213ps (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 # encoding: [0xc5,0xfc,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 # encoding: [0xc5,0xfc,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmaddsub213ps (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0x75,0xa6,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmaddsub_ps_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddsubps %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddsubps %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x5c,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <8 x float> @llvm.x86.fma.vfmaddsub.ps.256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) ret <8 x float> %res } @@ -784,20 +781,20 @@ declare <8 x float> @llvm.x86.fma.vfmaddsub.ps.256(<8 x float>, <8 x float>, <8 define <4 x double> @test_x86_fma_vfmaddsub_pd_256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmaddsub_pd_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmaddsub213pd %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmaddsub213pd %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0xf5,0xa6,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmaddsub_pd_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfmaddsub213pd (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 # encoding: [0xc5,0xfd,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 # encoding: [0xc5,0xfd,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmaddsub213pd (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0xf5,0xa6,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmaddsub_pd_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmaddsubpd %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmaddsubpd %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x5d,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x double> @llvm.x86.fma.vfmaddsub.pd.256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) ret <4 x double> %res } @@ -807,20 +804,20 @@ declare <4 x double> @llvm.x86.fma.vfmaddsub.pd.256(<4 x double>, <4 x double>, define <4 x float> @test_x86_fma_vfmsubadd_ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsubadd_ps: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsubadd213ps %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsubadd213ps %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0x71,0xa7,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsubadd_ps: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmsubadd213ps (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %xmm1 # encoding: [0xc5,0xf8,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %xmm0 # encoding: [0xc5,0xf8,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmsubadd213ps (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0x71,0xa7,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsubadd_ps: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubaddps %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubaddps %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x5e,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x float> @llvm.x86.fma.vfmsubadd.ps(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) ret <4 x float> %res } @@ -829,20 +826,20 @@ declare <4 x float> @llvm.x86.fma.vfmsubadd.ps(<4 x float>, <4 x float>, <4 x fl define <2 x double> @test_x86_fma_vfmsubadd_pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsubadd_pd: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsubadd213pd %xmm2, %xmm1, %xmm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsubadd213pd %xmm2, %xmm1, %xmm0 # encoding: [0xc4,0xe2,0xf1,0xa7,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsubadd_pd: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 -; CHECK-FMA-WIN-NEXT: vfmsubadd213pd (%r8), %xmm1, %xmm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %xmm1 # encoding: [0xc5,0xf9,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %xmm0 # encoding: [0xc5,0xf9,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmsubadd213pd (%r8), %xmm1, %xmm0 # encoding: [0xc4,0xc2,0xf1,0xa7,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsubadd_pd: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubaddpd %xmm2, %xmm1, %xmm0, %xmm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubaddpd %xmm2, %xmm1, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0xf9,0x5f,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <2 x double> @llvm.x86.fma.vfmsubadd.pd(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2) ret <2 x double> %res } @@ -851,20 +848,20 @@ declare <2 x double> @llvm.x86.fma.vfmsubadd.pd(<2 x double>, <2 x double>, <2 x define <8 x float> @test_x86_fma_vfmsubadd_ps_256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsubadd_ps_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsubadd213ps %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsubadd213ps %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0x75,0xa7,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsubadd_ps_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfmsubadd213ps (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovaps (%rcx), %ymm1 # encoding: [0xc5,0xfc,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovaps (%rdx), %ymm0 # encoding: [0xc5,0xfc,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmsubadd213ps (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0x75,0xa7,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsubadd_ps_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubaddps %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubaddps %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x5e,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <8 x float> @llvm.x86.fma.vfmsubadd.ps.256(<8 x float> %a0, <8 x float> %a1, <8 x float> %a2) ret <8 x float> %res } @@ -873,20 +870,20 @@ declare <8 x float> @llvm.x86.fma.vfmsubadd.ps.256(<8 x float>, <8 x float>, <8 define <4 x double> @test_x86_fma_vfmsubadd_pd_256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) #0 { ; CHECK-FMA-LABEL: test_x86_fma_vfmsubadd_pd_256: ; CHECK-FMA: # BB#0: -; CHECK-FMA-NEXT: vfmsubadd213pd %ymm2, %ymm1, %ymm0 -; CHECK-FMA-NEXT: retq +; CHECK-FMA-NEXT: vfmsubadd213pd %ymm2, %ymm1, %ymm0 # encoding: [0xc4,0xe2,0xf5,0xa7,0xc2] +; CHECK-FMA-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA-WIN-LABEL: test_x86_fma_vfmsubadd_pd_256: ; CHECK-FMA-WIN: # BB#0: -; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 -; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 -; CHECK-FMA-WIN-NEXT: vfmsubadd213pd (%r8), %ymm1, %ymm0 -; CHECK-FMA-WIN-NEXT: retq +; CHECK-FMA-WIN-NEXT: vmovapd (%rcx), %ymm1 # encoding: [0xc5,0xfd,0x28,0x09] +; CHECK-FMA-WIN-NEXT: vmovapd (%rdx), %ymm0 # encoding: [0xc5,0xfd,0x28,0x02] +; CHECK-FMA-WIN-NEXT: vfmsubadd213pd (%r8), %ymm1, %ymm0 # encoding: [0xc4,0xc2,0xf5,0xa7,0x00] +; CHECK-FMA-WIN-NEXT: retq # encoding: [0xc3] ; ; CHECK-FMA4-LABEL: test_x86_fma_vfmsubadd_pd_256: ; CHECK-FMA4: # BB#0: -; CHECK-FMA4-NEXT: vfmsubaddpd %ymm2, %ymm1, %ymm0, %ymm0 -; CHECK-FMA4-NEXT: retq +; CHECK-FMA4-NEXT: vfmsubaddpd %ymm2, %ymm1, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0xfd,0x5f,0xc2,0x10] +; CHECK-FMA4-NEXT: retq # encoding: [0xc3] %res = call <4 x double> @llvm.x86.fma.vfmsubadd.pd.256(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) ret <4 x double> %res }