[llvm] [X86] combineConcatVectorOps - add handling to concat ISD::FROUND/FFLOOR intrinsics together (PR #170176)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Dec 1 09:47:36 PST 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-x86
Author: Simon Pilgrim (RKSimon)
<details>
<summary>Changes</summary>
These were missed in #<!-- -->170160
---
Patch is 23.55 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/170176.diff
3 Files Affected:
- (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+2)
- (modified) llvm/test/CodeGen/X86/combine-ffloor.ll (+52-37)
- (modified) llvm/test/CodeGen/X86/combine-fround.ll (+94-100)
``````````diff
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 539b238d5043f..ed2539d163d11 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -59463,7 +59463,9 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
case ISD::FTRUNC:
case ISD::FRINT:
case ISD::FNEARBYINT:
+ case ISD::FROUND:
case ISD::FROUNDEVEN:
+ case ISD::FFLOOR:
if (!IsSplat && (VT.is256BitVector() ||
(VT.is512BitVector() && Subtarget.useAVX512Regs()))) {
return DAG.getNode(Opcode, DL, VT, ConcatSubOperand(VT, Ops, 0));
diff --git a/llvm/test/CodeGen/X86/combine-ffloor.ll b/llvm/test/CodeGen/X86/combine-ffloor.ll
index 35c28d2c85399..5cde95ec7aa4f 100644
--- a/llvm/test/CodeGen/X86/combine-ffloor.ll
+++ b/llvm/test/CodeGen/X86/combine-ffloor.ll
@@ -13,9 +13,9 @@ define <4 x double> @concat_floor_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1
;
; AVX-LABEL: concat_floor_v4f64_v2f64:
; AVX: # %bb.0:
-; AVX-NEXT: vroundpd $9, %xmm0, %xmm0
-; AVX-NEXT: vroundpd $9, %xmm1, %xmm1
+; AVX-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT: vroundpd $9, %ymm0, %ymm0
; AVX-NEXT: retq
%v0 = call <2 x double> @llvm.floor.v2f64(<2 x double> %a0)
%v1 = call <2 x double> @llvm.floor.v2f64(<2 x double> %a1)
@@ -32,9 +32,9 @@ define <8 x float> @concat_floor_v8f32_v4f32(<4 x float> %a0, <4 x float> %a1) {
;
; AVX-LABEL: concat_floor_v8f32_v4f32:
; AVX: # %bb.0:
-; AVX-NEXT: vroundps $9, %xmm0, %xmm0
-; AVX-NEXT: vroundps $9, %xmm1, %xmm1
+; AVX-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT: vroundps $9, %ymm0, %ymm0
; AVX-NEXT: retq
%v0 = call <4 x float> @llvm.floor.v4f32(<4 x float> %a0)
%v1 = call <4 x float> @llvm.floor.v4f32(<4 x float> %a1)
@@ -51,25 +51,34 @@ define <8 x double> @concat_floor_v8f64_v2f64(<2 x double> %a0, <2 x double> %a1
; SSE-NEXT: roundpd $9, %xmm3, %xmm3
; SSE-NEXT: retq
;
-; AVX1OR2-LABEL: concat_floor_v8f64_v2f64:
-; AVX1OR2: # %bb.0:
-; AVX1OR2-NEXT: vroundpd $9, %xmm0, %xmm0
-; AVX1OR2-NEXT: vroundpd $9, %xmm1, %xmm1
-; AVX1OR2-NEXT: vroundpd $9, %xmm2, %xmm2
-; AVX1OR2-NEXT: vroundpd $9, %xmm3, %xmm3
-; AVX1OR2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT: retq
+; AVX1-LABEL: concat_floor_v8f64_v2f64:
+; AVX1: # %bb.0:
+; AVX1-NEXT: # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT: vroundpd $9, %ymm0, %ymm0
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT: vroundpd $9, %ymm1, %ymm1
+; AVX1-NEXT: retq
+;
+; AVX2-LABEL: concat_floor_v8f64_v2f64:
+; AVX2: # %bb.0:
+; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT: vroundpd $9, %ymm0, %ymm0
+; AVX2-NEXT: # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT: vroundpd $9, %ymm1, %ymm1
+; AVX2-NEXT: retq
;
; AVX512-LABEL: concat_floor_v8f64_v2f64:
; AVX512: # %bb.0:
-; AVX512-NEXT: vroundpd $9, %xmm0, %xmm0
-; AVX512-NEXT: vroundpd $9, %xmm1, %xmm1
-; AVX512-NEXT: vroundpd $9, %xmm2, %xmm2
-; AVX512-NEXT: vroundpd $9, %xmm3, %xmm3
+; AVX512-NEXT: # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX512-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
; AVX512-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
; AVX512-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX512-NEXT: vinsertf64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-NEXT: vrndscalepd $9, %zmm0, %zmm0
; AVX512-NEXT: retq
%v0 = call <2 x double> @llvm.floor.v2f64(<2 x double> %a0)
%v1 = call <2 x double> @llvm.floor.v2f64(<2 x double> %a1)
@@ -90,25 +99,34 @@ define <16 x float> @concat_floor_v16f32_v4f32(<4 x float> %a0, <4 x float> %a1,
; SSE-NEXT: roundps $9, %xmm3, %xmm3
; SSE-NEXT: retq
;
-; AVX1OR2-LABEL: concat_floor_v16f32_v4f32:
-; AVX1OR2: # %bb.0:
-; AVX1OR2-NEXT: vroundps $9, %xmm0, %xmm0
-; AVX1OR2-NEXT: vroundps $9, %xmm1, %xmm1
-; AVX1OR2-NEXT: vroundps $9, %xmm2, %xmm2
-; AVX1OR2-NEXT: vroundps $9, %xmm3, %xmm3
-; AVX1OR2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT: retq
+; AVX1-LABEL: concat_floor_v16f32_v4f32:
+; AVX1: # %bb.0:
+; AVX1-NEXT: # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT: vroundps $9, %ymm0, %ymm0
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT: vroundps $9, %ymm1, %ymm1
+; AVX1-NEXT: retq
+;
+; AVX2-LABEL: concat_floor_v16f32_v4f32:
+; AVX2: # %bb.0:
+; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT: vroundps $9, %ymm0, %ymm0
+; AVX2-NEXT: # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT: vroundps $9, %ymm1, %ymm1
+; AVX2-NEXT: retq
;
; AVX512-LABEL: concat_floor_v16f32_v4f32:
; AVX512: # %bb.0:
-; AVX512-NEXT: vroundps $9, %xmm0, %xmm0
-; AVX512-NEXT: vroundps $9, %xmm1, %xmm1
-; AVX512-NEXT: vroundps $9, %xmm2, %xmm2
-; AVX512-NEXT: vroundps $9, %xmm3, %xmm3
+; AVX512-NEXT: # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX512-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
; AVX512-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
; AVX512-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX512-NEXT: vinsertf64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512-NEXT: vrndscaleps $9, %zmm0, %zmm0
; AVX512-NEXT: retq
%v0 = call <4 x float> @llvm.floor.v4f32(<4 x float> %a0)
%v1 = call <4 x float> @llvm.floor.v4f32(<4 x float> %a1)
@@ -137,9 +155,9 @@ define <8 x double> @concat_floor_v8f64_v4f64(<4 x double> %a0, <4 x double> %a1
;
; AVX512-LABEL: concat_floor_v8f64_v4f64:
; AVX512: # %bb.0:
-; AVX512-NEXT: vroundpd $9, %ymm0, %ymm0
-; AVX512-NEXT: vroundpd $9, %ymm1, %ymm1
+; AVX512-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
; AVX512-NEXT: vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT: vrndscalepd $9, %zmm0, %zmm0
; AVX512-NEXT: retq
%v0 = call <4 x double> @llvm.floor.v4f64(<4 x double> %a0)
%v1 = call <4 x double> @llvm.floor.v4f64(<4 x double> %a1)
@@ -164,15 +182,12 @@ define <16 x float> @concat_floor_v16f32_v8f32(<8 x float> %a0, <8 x float> %a1)
;
; AVX512-LABEL: concat_floor_v16f32_v8f32:
; AVX512: # %bb.0:
-; AVX512-NEXT: vroundps $9, %ymm0, %ymm0
-; AVX512-NEXT: vroundps $9, %ymm1, %ymm1
+; AVX512-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
; AVX512-NEXT: vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT: vrndscaleps $9, %zmm0, %zmm0
; AVX512-NEXT: retq
%v0 = call <8 x float> @llvm.floor.v8f32(<8 x float> %a0)
%v1 = call <8 x float> @llvm.floor.v8f32(<8 x float> %a1)
%res = shufflevector <8 x float> %v0, <8 x float> %v1, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
ret <16 x float> %res
}
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; AVX1: {{.*}}
-; AVX2: {{.*}}
diff --git a/llvm/test/CodeGen/X86/combine-fround.ll b/llvm/test/CodeGen/X86/combine-fround.ll
index 44865544ecb00..42dbaf234dbc7 100644
--- a/llvm/test/CodeGen/X86/combine-fround.ll
+++ b/llvm/test/CodeGen/X86/combine-fround.ll
@@ -1,7 +1,7 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v2 | FileCheck %s --check-prefixes=SSE
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=sandybridge | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX1
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v3 | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX2
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=sandybridge | FileCheck %s --check-prefixes=AVX,AVX1
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v3 | FileCheck %s --check-prefixes=AVX,AVX2
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v4 | FileCheck %s --check-prefixes=AVX,AVX512
define <4 x double> @concat_round_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1) {
@@ -20,26 +20,32 @@ define <4 x double> @concat_round_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1
; SSE-NEXT: roundpd $11, %xmm2, %xmm1
; SSE-NEXT: retq
;
-; AVX1OR2-LABEL: concat_round_v4f64_v2f64:
-; AVX1OR2: # %bb.0:
-; AVX1OR2-NEXT: vmovddup {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0]
-; AVX1OR2-NEXT: # xmm2 = mem[0,0]
-; AVX1OR2-NEXT: vandpd %xmm2, %xmm0, %xmm3
-; AVX1OR2-NEXT: vmovddup {{.*#+}} xmm4 = [4.9999999999999994E-1,4.9999999999999994E-1]
-; AVX1OR2-NEXT: # xmm4 = mem[0,0]
-; AVX1OR2-NEXT: vorpd %xmm4, %xmm3, %xmm3
-; AVX1OR2-NEXT: vaddpd %xmm3, %xmm0, %xmm0
-; AVX1OR2-NEXT: vandpd %xmm2, %xmm1, %xmm2
-; AVX1OR2-NEXT: vorpd %xmm4, %xmm2, %xmm2
-; AVX1OR2-NEXT: vaddpd %xmm2, %xmm1, %xmm1
-; AVX1OR2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT: vroundpd $11, %ymm0, %ymm0
-; AVX1OR2-NEXT: retq
+; AVX1-LABEL: concat_round_v4f64_v2f64:
+; AVX1: # %bb.0:
+; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT: vandpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX1-NEXT: vorpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX1-NEXT: vaddpd %ymm1, %ymm0, %ymm0
+; AVX1-NEXT: vroundpd $11, %ymm0, %ymm0
+; AVX1-NEXT: retq
+;
+; AVX2-LABEL: concat_round_v4f64_v2f64:
+; AVX2: # %bb.0:
+; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX2-NEXT: vandpd %ymm1, %ymm0, %ymm1
+; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm2 = [4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1]
+; AVX2-NEXT: vorpd %ymm2, %ymm1, %ymm1
+; AVX2-NEXT: vaddpd %ymm1, %ymm0, %ymm0
+; AVX2-NEXT: vroundpd $11, %ymm0, %ymm0
+; AVX2-NEXT: retq
;
; AVX512-LABEL: concat_round_v4f64_v2f64:
; AVX512: # %bb.0:
; AVX512-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
-; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm2 = [4602678819172646911,4602678819172646911,4602678819172646911,4602678819172646911]
+; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm2 = [4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1]
; AVX512-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512-NEXT: vpternlogq {{.*#+}} ymm2 = ymm2 | (ymm0 & m64bcst)
; AVX512-NEXT: vaddpd %ymm2, %ymm0, %ymm0
@@ -67,24 +73,32 @@ define <8 x float> @concat_round_v8f32_v4f32(<4 x float> %a0, <4 x float> %a1) {
; SSE-NEXT: roundps $11, %xmm2, %xmm1
; SSE-NEXT: retq
;
-; AVX1OR2-LABEL: concat_round_v8f32_v4f32:
-; AVX1OR2: # %bb.0:
-; AVX1OR2-NEXT: vbroadcastss {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
-; AVX1OR2-NEXT: vandps %xmm2, %xmm0, %xmm3
-; AVX1OR2-NEXT: vbroadcastss {{.*#+}} xmm4 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
-; AVX1OR2-NEXT: vorps %xmm4, %xmm3, %xmm3
-; AVX1OR2-NEXT: vaddps %xmm3, %xmm0, %xmm0
-; AVX1OR2-NEXT: vandps %xmm2, %xmm1, %xmm2
-; AVX1OR2-NEXT: vorps %xmm4, %xmm2, %xmm2
-; AVX1OR2-NEXT: vaddps %xmm2, %xmm1, %xmm1
-; AVX1OR2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT: vroundps $11, %ymm0, %ymm0
-; AVX1OR2-NEXT: retq
+; AVX1-LABEL: concat_round_v8f32_v4f32:
+; AVX1: # %bb.0:
+; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
+; AVX1-NEXT: vorps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX1-NEXT: vaddps %ymm1, %ymm0, %ymm0
+; AVX1-NEXT: vroundps $11, %ymm0, %ymm0
+; AVX1-NEXT: retq
+;
+; AVX2-LABEL: concat_round_v8f32_v4f32:
+; AVX2: # %bb.0:
+; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT: vbroadcastss {{.*#+}} ymm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX2-NEXT: vandps %ymm1, %ymm0, %ymm1
+; AVX2-NEXT: vbroadcastss {{.*#+}} ymm2 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; AVX2-NEXT: vorps %ymm2, %ymm1, %ymm1
+; AVX2-NEXT: vaddps %ymm1, %ymm0, %ymm0
+; AVX2-NEXT: vroundps $11, %ymm0, %ymm0
+; AVX2-NEXT: retq
;
; AVX512-LABEL: concat_round_v8f32_v4f32:
; AVX512: # %bb.0:
; AVX512-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
-; AVX512-NEXT: vpbroadcastd {{.*#+}} ymm2 = [1056964607,1056964607,1056964607,1056964607,1056964607,1056964607,1056964607,1056964607]
+; AVX512-NEXT: vpbroadcastd {{.*#+}} ymm2 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
; AVX512-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512-NEXT: vpternlogd {{.*#+}} ymm2 = ymm2 | (ymm0 & m32bcst)
; AVX512-NEXT: vaddps %ymm2, %ymm0, %ymm0
@@ -124,49 +138,37 @@ define <8 x double> @concat_round_v8f64_v2f64(<2 x double> %a0, <2 x double> %a1
;
; AVX1-LABEL: concat_round_v8f64_v2f64:
; AVX1: # %bb.0:
-; AVX1-NEXT: vmovddup {{.*#+}} xmm4 = [-0.0E+0,-0.0E+0]
-; AVX1-NEXT: # xmm4 = mem[0,0]
-; AVX1-NEXT: vandpd %xmm4, %xmm0, %xmm5
-; AVX1-NEXT: vmovddup {{.*#+}} xmm6 = [4.9999999999999994E-1,4.9999999999999994E-1]
-; AVX1-NEXT: # xmm6 = mem[0,0]
-; AVX1-NEXT: vorpd %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vaddpd %xmm5, %xmm0, %xmm0
-; AVX1-NEXT: vandpd %xmm4, %xmm1, %xmm5
-; AVX1-NEXT: vorpd %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vaddpd %xmm5, %xmm1, %xmm1
-; AVX1-NEXT: vandpd %xmm4, %xmm2, %xmm5
-; AVX1-NEXT: vorpd %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vaddpd %xmm5, %xmm2, %xmm2
-; AVX1-NEXT: vandpd %xmm4, %xmm3, %xmm4
-; AVX1-NEXT: vorpd %xmm6, %xmm4, %xmm4
-; AVX1-NEXT: vaddpd %xmm4, %xmm3, %xmm3
+; AVX1-NEXT: # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT: vmovapd {{.*#+}} ymm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX1-NEXT: vandpd %ymm1, %ymm0, %ymm4
+; AVX1-NEXT: vmovapd {{.*#+}} ymm5 = [4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1]
+; AVX1-NEXT: vorpd %ymm5, %ymm4, %ymm4
+; AVX1-NEXT: vaddpd %ymm4, %ymm0, %ymm0
; AVX1-NEXT: vroundpd $11, %ymm0, %ymm0
-; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT: vandpd %ymm1, %ymm2, %ymm1
+; AVX1-NEXT: vorpd %ymm5, %ymm1, %ymm1
+; AVX1-NEXT: vaddpd %ymm1, %ymm2, %ymm1
; AVX1-NEXT: vroundpd $11, %ymm1, %ymm1
; AVX1-NEXT: retq
;
; AVX2-LABEL: concat_round_v8f64_v2f64:
; AVX2: # %bb.0:
-; AVX2-NEXT: vmovddup {{.*#+}} xmm4 = [-0.0E+0,-0.0E+0]
-; AVX2-NEXT: # xmm4 = mem[0,0]
-; AVX2-NEXT: vandpd %xmm4, %xmm0, %xmm5
-; AVX2-NEXT: vmovddup {{.*#+}} xmm6 = [4.9999999999999994E-1,4.9999999999999994E-1]
-; AVX2-NEXT: # xmm6 = mem[0,0]
-; AVX2-NEXT: vorpd %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vaddpd %xmm5, %xmm0, %xmm0
-; AVX2-NEXT: vandpd %xmm4, %xmm1, %xmm5
-; AVX2-NEXT: vorpd %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vaddpd %xmm5, %xmm1, %xmm1
-; AVX2-NEXT: vandpd %xmm4, %xmm2, %xmm5
-; AVX2-NEXT: vorpd %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vaddpd %xmm5, %xmm2, %xmm2
-; AVX2-NEXT: vandpd %xmm4, %xmm3, %xmm4
-; AVX2-NEXT: vorpd %xmm6, %xmm4, %xmm4
+; AVX2-NEXT: # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX2-NEXT: vandpd %ymm1, %ymm0, %ymm4
+; AVX2-NEXT: vbroadcastsd {{.*#+}} ymm5 = [4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1]
+; AVX2-NEXT: vorpd %ymm5, %ymm4, %ymm4
+; AVX2-NEXT: vaddpd %ymm4, %ymm0, %ymm0
; AVX2-NEXT: vroundpd $11, %ymm0, %ymm0
-; AVX2-NEXT: vaddpd %xmm4, %xmm3, %xmm1
-; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
+; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX2-NEXT: vandpd %ymm1, %ymm2, %ymm1
+; AVX2-NEXT: vorpd %ymm5, %ymm1, %ymm1
+; AVX2-NEXT: vaddpd %ymm1, %ymm2, %ymm1
; AVX2-NEXT: vroundpd $11, %ymm1, %ymm1
; AVX2-NEXT: retq
;
@@ -176,7 +178,7 @@ define <8 x double> @concat_round_v8f64_v2f64(<2 x double> %a0, <2 x double> %a1
; AVX512-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
; AVX512-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2
; AVX512-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX512-NEXT: vpbroadcastq {{.*#+}} zmm1 = [4602678819172646911,4602678819172646911,4602678819172646911,4602678819172646911,4602678819172646911,4602678819172646911,4602678819172646911,4602678819172646911]
+; AVX512-NEXT: vpbroadcastq {{.*#+}} zmm1 = [4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1,4.9999999999999994E-1]
; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0
; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 | (zmm0 & m64bcst)
; AVX512-NEXT: vaddpd %zmm1, %zmm0, %zmm0
@@ -220,45 +222,37 @@ define <16 x float> @concat_round_v16f32_v4f32(<4 x float> %a0, <4 x float> %a1,
;
; AVX1-LABEL: concat_round_v16f32_v4f32:
; AVX1: # %bb.0:
-; AVX1-NEXT: vbroadcastss {{.*#+}} xmm4 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
-; AVX1-NEXT: vandps %xmm4, %xmm0, %xmm5
-; AVX1-NEXT: vbroadcastss {{.*#+}} xmm6 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
-; AVX1-NEXT: vorps %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vaddps %xmm5, %xmm0, %xmm0
-; AVX1-NEXT: vandps %xmm4, %xmm1, %xmm5
-; AVX1-NEXT: vorps %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vaddps %xmm5, %xmm1, %xmm1
-; AVX1-NEXT: vandps %xmm4, %xmm2, %xmm5
-; AVX1-NEXT: vorps %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vaddps %xmm5, %xmm2, %xmm2
-; AVX1-NEXT: vandps %xmm4, %xmm3, %xmm4
-; AVX1-NEXT: vorps %xmm6, %xmm4, %xmm4
-; AVX1-NEXT: vaddps %xmm4, %xmm3, %xmm3
+; AVX1-NEXT: # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX1-NEXT: # kill: def $xmm0 killed $xmm0 def $ymm0
; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT: vmovaps {{.*#+}} ymm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; AVX1-NEXT: vandps %ymm1, %ymm0, %ymm4
+; AVX1-NEXT: vmovaps {{.*#+}} ymm5 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
+; AVX1-NEXT: vorps %ymm5, %ymm4, %ymm4
+; AVX1-NEXT: vaddps %ymm4, %ymm0, %ymm0
; AVX1-NEXT: vroundps $11, %ymm0, %ymm0
-; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT: vandps %ymm1, %ymm2, %ymm1
+; AVX1-NEXT: vorps %ymm5, %ymm1, %ymm1
+; AVX1-NEXT: vaddps %ymm1, %ymm2, %ymm1
; AVX1-NEXT: vroundps $11, %ymm1, %ymm1
; AVX1-NEXT: retq
;
; AVX2-LABEL: concat_round_v16f32_v4f32:
; AVX2: # %bb.0:
-; AVX2-NEXT: vbroadcastss {{.*#+}} xmm4 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
-; AVX2-NEXT: vandps %xmm4, %xmm0, %xmm5
-; AVX2-NEXT: vbroadcastss {{.*#+}} xmm6 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
-; AVX2-NEXT: vorps %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vaddps %xmm5, %xmm0, %xmm0
-; AVX2-NEXT: vandps %xmm4, %xmm1, %xmm5
-; AVX2-NEXT: vorps %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vaddps %xmm5, %xmm1, %xmm1
-; AVX2-NEXT: vandps %xmm4, %xmm2, %xmm5
-; AVX2-NEXT: vorps %xmm6, %xmm5, %xmm5
-; AVX2-NEXT: vaddps %xmm5, %xmm2, %xmm2
-; AVX2-NEXT: vandps %xmm4, %xmm3, %xmm4
-; AVX2-NEXT: vorps...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/170176
More information about the llvm-commits
mailing list