From 04347d834cf9c330378074573d55c5817cbb183d Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Mon, 1 Dec 2025 15:46:52 +0000
Subject: [PATCH] [X86] combineConcatVectorOps - add handling to concat fp
 rounding intrinsics together

---
 llvm/lib/Target/X86/X86ISelLowering.cpp     | 16 ++++
 llvm/test/CodeGen/X86/combine-fceil.ll      | 90 ++++++++++++---------
 llvm/test/CodeGen/X86/combine-fnearbyint.ll | 90 ++++++++++++---------
 llvm/test/CodeGen/X86/combine-frint.ll      | 90 ++++++++++++---------
 llvm/test/CodeGen/X86/combine-froundeven.ll | 90 ++++++++++++---------
 llvm/test/CodeGen/X86/combine-ftrunc.ll     | 90 ++++++++++++---------
 llvm/test/CodeGen/X86/combine-rndscale.ll   | 90 ++++++++++++---------
 7 files changed, 340 insertions(+), 216 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 74a02711bd98a..539b238d5043f 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -59459,6 +59459,11 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
       }
       break;
     case ISD::FSQRT:
+    case ISD::FCEIL:
+    case ISD::FTRUNC:
+    case ISD::FRINT:
+    case ISD::FNEARBYINT:
+    case ISD::FROUNDEVEN:
       if (!IsSplat && (VT.is256BitVector() ||
                        (VT.is512BitVector() && Subtarget.useAVX512Regs()))) {
         return DAG.getNode(Opcode, DL, VT, ConcatSubOperand(VT, Ops, 0));
@@ -59470,6 +59475,17 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
         return DAG.getNode(Opcode, DL, VT, ConcatSubOperand(VT, Ops, 0));
       }
       break;
+    case X86ISD::VRNDSCALE:
+      if (!IsSplat &&
+          (VT.is256BitVector() ||
+           (VT.is512BitVector() && Subtarget.useAVX512Regs())) &&
+          llvm::all_of(Ops, [Op0](SDValue Op) {
+            return Op0.getOperand(1) == Op.getOperand(1);
+          })) {
+        return DAG.getNode(Opcode, DL, VT, ConcatSubOperand(VT, Ops, 0),
+                           Op0.getOperand(1));
+      }
+      break;
     case X86ISD::HADD:
     case X86ISD::HSUB:
     case X86ISD::FHADD:
diff --git a/llvm/test/CodeGen/X86/combine-fceil.ll b/llvm/test/CodeGen/X86/combine-fceil.ll
index 78f1476a49152..a3f55e8f64b80 100644
--- a/llvm/test/CodeGen/X86/combine-fceil.ll
+++ b/llvm/test/CodeGen/X86/combine-fceil.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
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v3 | FileCheck %s --check-prefixes=AVX,AVX1OR2
+; 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=x86-64-v4 | FileCheck %s --check-prefixes=AVX,AVX512
 
 define <4 x double> @concat_ceil_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1) {
@@ -13,9 +13,9 @@ define <4 x double> @concat_ceil_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1)
 ;
 ; AVX-LABEL: concat_ceil_v4f64_v2f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundpd $10, %xmm0, %xmm0
-; AVX-NEXT:    vroundpd $10, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundpd $10, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <2 x double> @llvm.ceil.v2f64(<2 x double> %a0)
   %v1 = call <2 x double> @llvm.ceil.v2f64(<2 x double> %a1)
@@ -32,9 +32,9 @@ define <8 x float> @concat_ceil_v8f32_v4f32(<4 x float> %a0, <4 x float> %a1) {
 ;
 ; AVX-LABEL: concat_ceil_v8f32_v4f32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundps $10, %xmm0, %xmm0
-; AVX-NEXT:    vroundps $10, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundps $10, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <4 x float> @llvm.ceil.v4f32(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.ceil.v4f32(<4 x float> %a1)
@@ -51,25 +51,34 @@ define <8 x double> @concat_ceil_v8f64_v2f64(<2 x double> %a0, <2 x double> %a1,
 ; SSE-NEXT:    roundpd $10, %xmm3, %xmm3
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: concat_ceil_v8f64_v2f64:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundpd $10, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundpd $10, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundpd $10, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundpd $10, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_ceil_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 $10, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundpd $10, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_ceil_v8f64_v2f64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundpd $10, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundpd $10, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_ceil_v8f64_v2f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $10, %xmm0, %xmm0
-; AVX512-NEXT:    vroundpd $10, %xmm1, %xmm1
-; AVX512-NEXT:    vroundpd $10, %xmm2, %xmm2
-; AVX512-NEXT:    vroundpd $10, %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 $10, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <2 x double> @llvm.ceil.v2f64(<2 x double> %a0)
   %v1 = call <2 x double> @llvm.ceil.v2f64(<2 x double> %a1)
@@ -90,25 +99,34 @@ define <16 x float> @concat_ceil_v16f32_v4f32(<4 x float> %a0, <4 x float> %a1,
 ; SSE-NEXT:    roundps $10, %xmm3, %xmm3
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: concat_ceil_v16f32_v4f32:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundps $10, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundps $10, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundps $10, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundps $10, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_ceil_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 $10, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundps $10, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_ceil_v16f32_v4f32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundps $10, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundps $10, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_ceil_v16f32_v4f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $10, %xmm0, %xmm0
-; AVX512-NEXT:    vroundps $10, %xmm1, %xmm1
-; AVX512-NEXT:    vroundps $10, %xmm2, %xmm2
-; AVX512-NEXT:    vroundps $10, %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 $10, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x float> @llvm.ceil.v4f32(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.ceil.v4f32(<4 x float> %a1)
@@ -137,9 +155,9 @@ define <8 x double> @concat_ceil_v8f64_v4f64(<4 x double> %a0, <4 x double> %a1)
 ;
 ; AVX512-LABEL: concat_ceil_v8f64_v4f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $10, %ymm0, %ymm0
-; AVX512-NEXT:    vroundpd $10, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscalepd $10, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x double> @llvm.ceil.v4f64(<4 x double> %a0)
   %v1 = call <4 x double> @llvm.ceil.v4f64(<4 x double> %a1)
@@ -164,9 +182,9 @@ define <16 x float> @concat_ceil_v16f32_v8f32(<8 x float> %a0, <8 x float> %a1)
 ;
 ; AVX512-LABEL: concat_ceil_v16f32_v8f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $10, %ymm0, %ymm0
-; AVX512-NEXT:    vroundps $10, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscaleps $10, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <8 x float> @llvm.ceil.v8f32(<8 x float> %a0)
   %v1 = call <8 x float> @llvm.ceil.v8f32(<8 x float> %a1)
diff --git a/llvm/test/CodeGen/X86/combine-fnearbyint.ll b/llvm/test/CodeGen/X86/combine-fnearbyint.ll
index 14d1017aec630..fde136af7c4c2 100644
--- a/llvm/test/CodeGen/X86/combine-fnearbyint.ll
+++ b/llvm/test/CodeGen/X86/combine-fnearbyint.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
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v3 | FileCheck %s --check-prefixes=AVX,AVX1OR2
+; 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=x86-64-v4 | FileCheck %s --check-prefixes=AVX,AVX512
 
 define <4 x double> @concat_nearbyint_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1) {
@@ -13,9 +13,9 @@ define <4 x double> @concat_nearbyint_v4f64_v2f64(<2 x double> %a0, <2 x double>
 ;
 ; AVX-LABEL: concat_nearbyint_v4f64_v2f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundpd $12, %xmm0, %xmm0
-; AVX-NEXT:    vroundpd $12, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundpd $12, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <2 x double> @llvm.nearbyint.v2f64(<2 x double> %a0)
   %v1 = call <2 x double> @llvm.nearbyint.v2f64(<2 x double> %a1)
@@ -32,9 +32,9 @@ define <8 x float> @concat_nearbyint_v8f32_v4f32(<4 x float> %a0, <4 x float> %a
 ;
 ; AVX-LABEL: concat_nearbyint_v8f32_v4f32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundps $12, %xmm0, %xmm0
-; AVX-NEXT:    vroundps $12, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundps $12, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <4 x float> @llvm.nearbyint.v4f32(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.nearbyint.v4f32(<4 x float> %a1)
@@ -51,25 +51,34 @@ define <8 x double> @concat_nearbyint_v8f64_v2f64(<2 x double> %a0, <2 x double>
 ; SSE-NEXT:    roundpd $12, %xmm3, %xmm3
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: concat_nearbyint_v8f64_v2f64:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundpd $12, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundpd $12, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundpd $12, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundpd $12, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_nearbyint_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 $12, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundpd $12, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_nearbyint_v8f64_v2f64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundpd $12, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundpd $12, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_nearbyint_v8f64_v2f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $12, %xmm0, %xmm0
-; AVX512-NEXT:    vroundpd $12, %xmm1, %xmm1
-; AVX512-NEXT:    vroundpd $12, %xmm2, %xmm2
-; AVX512-NEXT:    vroundpd $12, %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 $12, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <2 x double> @llvm.nearbyint.v2f64(<2 x double> %a0)
   %v1 = call <2 x double> @llvm.nearbyint.v2f64(<2 x double> %a1)
@@ -90,25 +99,34 @@ define <16 x float> @concat_nearbyint_v16f32_v4f32(<4 x float> %a0, <4 x float>
 ; SSE-NEXT:    roundps $12, %xmm3, %xmm3
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: concat_nearbyint_v16f32_v4f32:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundps $12, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundps $12, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundps $12, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundps $12, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_nearbyint_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 $12, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundps $12, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_nearbyint_v16f32_v4f32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundps $12, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundps $12, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_nearbyint_v16f32_v4f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $12, %xmm0, %xmm0
-; AVX512-NEXT:    vroundps $12, %xmm1, %xmm1
-; AVX512-NEXT:    vroundps $12, %xmm2, %xmm2
-; AVX512-NEXT:    vroundps $12, %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 $12, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x float> @llvm.nearbyint.v4f32(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.nearbyint.v4f32(<4 x float> %a1)
@@ -137,9 +155,9 @@ define <8 x double> @concat_nearbyint_v8f64_v4f64(<4 x double> %a0, <4 x double>
 ;
 ; AVX512-LABEL: concat_nearbyint_v8f64_v4f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $12, %ymm0, %ymm0
-; AVX512-NEXT:    vroundpd $12, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscalepd $12, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x double> @llvm.nearbyint.v4f64(<4 x double> %a0)
   %v1 = call <4 x double> @llvm.nearbyint.v4f64(<4 x double> %a1)
@@ -164,9 +182,9 @@ define <16 x float> @concat_nearbyint_v16f32_v8f32(<8 x float> %a0, <8 x float>
 ;
 ; AVX512-LABEL: concat_nearbyint_v16f32_v8f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $12, %ymm0, %ymm0
-; AVX512-NEXT:    vroundps $12, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscaleps $12, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <8 x float> @llvm.nearbyint.v8f32(<8 x float> %a0)
   %v1 = call <8 x float> @llvm.nearbyint.v8f32(<8 x float> %a1)
diff --git a/llvm/test/CodeGen/X86/combine-frint.ll b/llvm/test/CodeGen/X86/combine-frint.ll
index 901ce2c1f0d82..1c52529e8386c 100644
--- a/llvm/test/CodeGen/X86/combine-frint.ll
+++ b/llvm/test/CodeGen/X86/combine-frint.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
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v3 | FileCheck %s --check-prefixes=AVX,AVX1OR2
+; 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=x86-64-v4 | FileCheck %s --check-prefixes=AVX,AVX512
 
 define <4 x double> @concat_rint_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1) {
@@ -13,9 +13,9 @@ define <4 x double> @concat_rint_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1)
 ;
 ; AVX-LABEL: concat_rint_v4f64_v2f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundpd $4, %xmm0, %xmm0
-; AVX-NEXT:    vroundpd $4, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundpd $4, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <2 x double> @llvm.rint.v2f64(<2 x double> %a0)
   %v1 = call <2 x double> @llvm.rint.v2f64(<2 x double> %a1)
@@ -32,9 +32,9 @@ define <8 x float> @concat_rint_v8f32_v4f32(<4 x float> %a0, <4 x float> %a1) {
 ;
 ; AVX-LABEL: concat_rint_v8f32_v4f32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundps $4, %xmm0, %xmm0
-; AVX-NEXT:    vroundps $4, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundps $4, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <4 x float> @llvm.rint.v4f32(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.rint.v4f32(<4 x float> %a1)
@@ -51,25 +51,34 @@ define <8 x double> @concat_rint_v8f64_v2f64(<2 x double> %a0, <2 x double> %a1,
 ; SSE-NEXT:    roundpd $4, %xmm3, %xmm3
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: concat_rint_v8f64_v2f64:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundpd $4, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundpd $4, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundpd $4, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundpd $4, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_rint_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 $4, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundpd $4, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_rint_v8f64_v2f64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundpd $4, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundpd $4, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_rint_v8f64_v2f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $4, %xmm0, %xmm0
-; AVX512-NEXT:    vroundpd $4, %xmm1, %xmm1
-; AVX512-NEXT:    vroundpd $4, %xmm2, %xmm2
-; AVX512-NEXT:    vroundpd $4, %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 $4, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <2 x double> @llvm.rint.v2f64(<2 x double> %a0)
   %v1 = call <2 x double> @llvm.rint.v2f64(<2 x double> %a1)
@@ -90,25 +99,34 @@ define <16 x float> @concat_rint_v16f32_v4f32(<4 x float> %a0, <4 x float> %a1,
 ; SSE-NEXT:    roundps $4, %xmm3, %xmm3
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: concat_rint_v16f32_v4f32:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundps $4, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundps $4, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundps $4, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundps $4, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_rint_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 $4, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundps $4, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_rint_v16f32_v4f32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundps $4, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundps $4, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_rint_v16f32_v4f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $4, %xmm0, %xmm0
-; AVX512-NEXT:    vroundps $4, %xmm1, %xmm1
-; AVX512-NEXT:    vroundps $4, %xmm2, %xmm2
-; AVX512-NEXT:    vroundps $4, %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 $4, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x float> @llvm.rint.v4f32(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.rint.v4f32(<4 x float> %a1)
@@ -137,9 +155,9 @@ define <8 x double> @concat_rint_v8f64_v4f64(<4 x double> %a0, <4 x double> %a1)
 ;
 ; AVX512-LABEL: concat_rint_v8f64_v4f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $4, %ymm0, %ymm0
-; AVX512-NEXT:    vroundpd $4, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscalepd $4, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x double> @llvm.rint.v4f64(<4 x double> %a0)
   %v1 = call <4 x double> @llvm.rint.v4f64(<4 x double> %a1)
@@ -164,9 +182,9 @@ define <16 x float> @concat_rint_v16f32_v8f32(<8 x float> %a0, <8 x float> %a1)
 ;
 ; AVX512-LABEL: concat_rint_v16f32_v8f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $4, %ymm0, %ymm0
-; AVX512-NEXT:    vroundps $4, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscaleps $4, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <8 x float> @llvm.rint.v8f32(<8 x float> %a0)
   %v1 = call <8 x float> @llvm.rint.v8f32(<8 x float> %a1)
diff --git a/llvm/test/CodeGen/X86/combine-froundeven.ll b/llvm/test/CodeGen/X86/combine-froundeven.ll
index 484e3a9680450..4bf1e86d887ae 100644
--- a/llvm/test/CodeGen/X86/combine-froundeven.ll
+++ b/llvm/test/CodeGen/X86/combine-froundeven.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
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v3 | FileCheck %s --check-prefixes=AVX,AVX1OR2
+; 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=x86-64-v4 | FileCheck %s --check-prefixes=AVX,AVX512
 
 define <4 x double> @concat_roundeven_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1) {
@@ -13,9 +13,9 @@ define <4 x double> @concat_roundeven_v4f64_v2f64(<2 x double> %a0, <2 x double>
 ;
 ; AVX-LABEL: concat_roundeven_v4f64_v2f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundpd $8, %xmm0, %xmm0
-; AVX-NEXT:    vroundpd $8, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundpd $8, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <2 x double> @llvm.roundeven.v2f64(<2 x double> %a0)
   %v1 = call <2 x double> @llvm.roundeven.v2f64(<2 x double> %a1)
@@ -32,9 +32,9 @@ define <8 x float> @concat_roundeven_v8f32_v4f32(<4 x float> %a0, <4 x float> %a
 ;
 ; AVX-LABEL: concat_roundeven_v8f32_v4f32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundps $8, %xmm0, %xmm0
-; AVX-NEXT:    vroundps $8, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundps $8, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <4 x float> @llvm.roundeven.v4f32(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.roundeven.v4f32(<4 x float> %a1)
@@ -51,25 +51,34 @@ define <8 x double> @concat_roundeven_v8f64_v2f64(<2 x double> %a0, <2 x double>
 ; SSE-NEXT:    roundpd $8, %xmm3, %xmm3
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: concat_roundeven_v8f64_v2f64:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundpd $8, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundpd $8, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundpd $8, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundpd $8, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_roundeven_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 $8, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundpd $8, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_roundeven_v8f64_v2f64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundpd $8, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundpd $8, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_roundeven_v8f64_v2f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $8, %xmm0, %xmm0
-; AVX512-NEXT:    vroundpd $8, %xmm1, %xmm1
-; AVX512-NEXT:    vroundpd $8, %xmm2, %xmm2
-; AVX512-NEXT:    vroundpd $8, %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 $8, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <2 x double> @llvm.roundeven.v2f64(<2 x double> %a0)
   %v1 = call <2 x double> @llvm.roundeven.v2f64(<2 x double> %a1)
@@ -90,25 +99,34 @@ define <16 x float> @concat_roundeven_v16f32_v4f32(<4 x float> %a0, <4 x float>
 ; SSE-NEXT:    roundps $8, %xmm3, %xmm3
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: concat_roundeven_v16f32_v4f32:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundps $8, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundps $8, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundps $8, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundps $8, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_roundeven_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 $8, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundps $8, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_roundeven_v16f32_v4f32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundps $8, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundps $8, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_roundeven_v16f32_v4f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $8, %xmm0, %xmm0
-; AVX512-NEXT:    vroundps $8, %xmm1, %xmm1
-; AVX512-NEXT:    vroundps $8, %xmm2, %xmm2
-; AVX512-NEXT:    vroundps $8, %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 $8, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x float> @llvm.roundeven.v4f32(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.roundeven.v4f32(<4 x float> %a1)
@@ -137,9 +155,9 @@ define <8 x double> @concat_roundeven_v8f64_v4f64(<4 x double> %a0, <4 x double>
 ;
 ; AVX512-LABEL: concat_roundeven_v8f64_v4f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $8, %ymm0, %ymm0
-; AVX512-NEXT:    vroundpd $8, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscalepd $8, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x double> @llvm.roundeven.v4f64(<4 x double> %a0)
   %v1 = call <4 x double> @llvm.roundeven.v4f64(<4 x double> %a1)
@@ -164,9 +182,9 @@ define <16 x float> @concat_roundeven_v16f32_v8f32(<8 x float> %a0, <8 x float>
 ;
 ; AVX512-LABEL: concat_roundeven_v16f32_v8f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $8, %ymm0, %ymm0
-; AVX512-NEXT:    vroundps $8, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscaleps $8, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <8 x float> @llvm.roundeven.v8f32(<8 x float> %a0)
   %v1 = call <8 x float> @llvm.roundeven.v8f32(<8 x float> %a1)
diff --git a/llvm/test/CodeGen/X86/combine-ftrunc.ll b/llvm/test/CodeGen/X86/combine-ftrunc.ll
index a6c703a1cbeae..3dde226db73df 100644
--- a/llvm/test/CodeGen/X86/combine-ftrunc.ll
+++ b/llvm/test/CodeGen/X86/combine-ftrunc.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
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v3 | FileCheck %s --check-prefixes=AVX,AVX1OR2
+; 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=x86-64-v4 | FileCheck %s --check-prefixes=AVX,AVX512
 
 define <4 x double> @concat_trunc_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1) {
@@ -13,9 +13,9 @@ define <4 x double> @concat_trunc_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1
 ;
 ; AVX-LABEL: concat_trunc_v4f64_v2f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundpd $11, %xmm0, %xmm0
-; AVX-NEXT:    vroundpd $11, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundpd $11, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <2 x double> @llvm.trunc.v2f64(<2 x double> %a0)
   %v1 = call <2 x double> @llvm.trunc.v2f64(<2 x double> %a1)
@@ -32,9 +32,9 @@ define <8 x float> @concat_trunc_v8f32_v4f32(<4 x float> %a0, <4 x float> %a1) {
 ;
 ; AVX-LABEL: concat_trunc_v8f32_v4f32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundps $11, %xmm0, %xmm0
-; AVX-NEXT:    vroundps $11, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundps $11, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <4 x float> @llvm.trunc.v4f32(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.trunc.v4f32(<4 x float> %a1)
@@ -51,25 +51,34 @@ define <8 x double> @concat_trunc_v8f64_v2f64(<2 x double> %a0, <2 x double> %a1
 ; SSE-NEXT:    roundpd $11, %xmm3, %xmm3
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: concat_trunc_v8f64_v2f64:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundpd $11, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundpd $11, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundpd $11, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundpd $11, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_trunc_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 $11, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundpd $11, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_trunc_v8f64_v2f64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundpd $11, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundpd $11, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_trunc_v8f64_v2f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $11, %xmm0, %xmm0
-; AVX512-NEXT:    vroundpd $11, %xmm1, %xmm1
-; AVX512-NEXT:    vroundpd $11, %xmm2, %xmm2
-; AVX512-NEXT:    vroundpd $11, %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 $11, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <2 x double> @llvm.trunc.v2f64(<2 x double> %a0)
   %v1 = call <2 x double> @llvm.trunc.v2f64(<2 x double> %a1)
@@ -90,25 +99,34 @@ define <16 x float> @concat_trunc_v16f32_v4f32(<4 x float> %a0, <4 x float> %a1,
 ; SSE-NEXT:    roundps $11, %xmm3, %xmm3
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: concat_trunc_v16f32_v4f32:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundps $11, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundps $11, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundps $11, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundps $11, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_trunc_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 $11, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundps $11, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_trunc_v16f32_v4f32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundps $11, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundps $11, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_trunc_v16f32_v4f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $11, %xmm0, %xmm0
-; AVX512-NEXT:    vroundps $11, %xmm1, %xmm1
-; AVX512-NEXT:    vroundps $11, %xmm2, %xmm2
-; AVX512-NEXT:    vroundps $11, %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 $11, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x float> @llvm.trunc.v4f32(<4 x float> %a0)
   %v1 = call <4 x float> @llvm.trunc.v4f32(<4 x float> %a1)
@@ -137,9 +155,9 @@ define <8 x double> @concat_trunc_v8f64_v4f64(<4 x double> %a0, <4 x double> %a1
 ;
 ; AVX512-LABEL: concat_trunc_v8f64_v4f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $11, %ymm0, %ymm0
-; AVX512-NEXT:    vroundpd $11, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscalepd $11, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x double> @llvm.trunc.v4f64(<4 x double> %a0)
   %v1 = call <4 x double> @llvm.trunc.v4f64(<4 x double> %a1)
@@ -164,9 +182,9 @@ define <16 x float> @concat_trunc_v16f32_v8f32(<8 x float> %a0, <8 x float> %a1)
 ;
 ; AVX512-LABEL: concat_trunc_v16f32_v8f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $11, %ymm0, %ymm0
-; AVX512-NEXT:    vroundps $11, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscaleps $11, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <8 x float> @llvm.trunc.v8f32(<8 x float> %a0)
   %v1 = call <8 x float> @llvm.trunc.v8f32(<8 x float> %a1)
diff --git a/llvm/test/CodeGen/X86/combine-rndscale.ll b/llvm/test/CodeGen/X86/combine-rndscale.ll
index 25117e864b512..b557dd8106d8e 100644
--- a/llvm/test/CodeGen/X86/combine-rndscale.ll
+++ b/llvm/test/CodeGen/X86/combine-rndscale.ll
@@ -1,14 +1,14 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=sandybridge | FileCheck %s --check-prefixes=AVX,AVX1OR2
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64-v3 | FileCheck %s --check-prefixes=AVX,AVX1OR2
+; 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=x86-64-v4 | FileCheck %s --check-prefixes=AVX,AVX512
 
 define <4 x double> @concat_roundpd_v4f64_v2f64(<2 x double> %a0, <2 x double> %a1) {
 ; AVX-LABEL: concat_roundpd_v4f64_v2f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundpd $4, %xmm0, %xmm0
-; AVX-NEXT:    vroundpd $4, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundpd $4, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <2 x double> @llvm.x86.sse41.round.pd(<2 x double> %a0, i32 4)
   %v1 = call <2 x double> @llvm.x86.sse41.round.pd(<2 x double> %a1, i32 4)
@@ -19,9 +19,9 @@ define <4 x double> @concat_roundpd_v4f64_v2f64(<2 x double> %a0, <2 x double> %
 define <8 x float> @concat_roundps_v8f32_v4f32(<4 x float> %a0, <4 x float> %a1) {
 ; AVX-LABEL: concat_roundps_v8f32_v4f32:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vroundps $4, %xmm0, %xmm0
-; AVX-NEXT:    vroundps $4, %xmm1, %xmm1
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX-NEXT:    vroundps $4, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %v0 = call <4 x float> @llvm.x86.sse41.round.ps(<4 x float> %a0, i32 4)
   %v1 = call <4 x float> @llvm.x86.sse41.round.ps(<4 x float> %a1, i32 4)
@@ -30,25 +30,34 @@ define <8 x float> @concat_roundps_v8f32_v4f32(<4 x float> %a0, <4 x float> %a1)
 }
 
 define <8 x double> @concat_roundpd_v8f64_v2f64(<2 x double> %a0, <2 x double> %a1, <2 x double> %a2, <2 x double> %a3) {
-; AVX1OR2-LABEL: concat_roundpd_v8f64_v2f64:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundpd $4, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundpd $4, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundpd $4, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundpd $4, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_roundpd_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 $4, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundpd $4, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_roundpd_v8f64_v2f64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundpd $4, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundpd $4, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_roundpd_v8f64_v2f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $4, %xmm0, %xmm0
-; AVX512-NEXT:    vroundpd $4, %xmm1, %xmm1
-; AVX512-NEXT:    vroundpd $4, %xmm2, %xmm2
-; AVX512-NEXT:    vroundpd $4, %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 $4, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <2 x double> @llvm.x86.sse41.round.pd(<2 x double> %a0, i32 4)
   %v1 = call <2 x double> @llvm.x86.sse41.round.pd(<2 x double> %a1, i32 4)
@@ -61,25 +70,34 @@ define <8 x double> @concat_roundpd_v8f64_v2f64(<2 x double> %a0, <2 x double> %
 }
 
 define <16 x float> @concat_roundps_v16f32_v4f32(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2, <4 x float> %a3) {
-; AVX1OR2-LABEL: concat_roundps_v16f32_v4f32:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vroundps $4, %xmm0, %xmm0
-; AVX1OR2-NEXT:    vroundps $4, %xmm1, %xmm1
-; AVX1OR2-NEXT:    vroundps $4, %xmm2, %xmm2
-; AVX1OR2-NEXT:    vroundps $4, %xmm3, %xmm3
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX1OR2-NEXT:    retq
+; AVX1-LABEL: concat_roundps_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 $4, %ymm0, %ymm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX1-NEXT:    vroundps $4, %ymm1, %ymm1
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: concat_roundps_v16f32_v4f32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vroundps $4, %ymm0, %ymm0
+; AVX2-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
+; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
+; AVX2-NEXT:    vroundps $4, %ymm1, %ymm1
+; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: concat_roundps_v16f32_v4f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $4, %xmm0, %xmm0
-; AVX512-NEXT:    vroundps $4, %xmm1, %xmm1
-; AVX512-NEXT:    vroundps $4, %xmm2, %xmm2
-; AVX512-NEXT:    vroundps $4, %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 $4, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x float> @llvm.x86.sse41.round.ps(<4 x float> %a0, i32 4)
   %v1 = call <4 x float> @llvm.x86.sse41.round.ps(<4 x float> %a1, i32 4)
@@ -100,9 +118,9 @@ define <8 x double> @concat_roundpd_v8f64_v4f64(<4 x double> %a0, <4 x double> %
 ;
 ; AVX512-LABEL: concat_roundpd_v8f64_v4f64:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundpd $4, %ymm0, %ymm0
-; AVX512-NEXT:    vroundpd $4, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscalepd $4, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <4 x double> @llvm.x86.avx.round.pd.256(<4 x double> %a0, i32 4)
   %v1 = call <4 x double> @llvm.x86.avx.round.pd.256(<4 x double> %a1, i32 4)
@@ -119,9 +137,9 @@ define <16 x float> @concat_roundps_v16f32_v8f32(<8 x float> %a0, <8 x float> %a
 ;
 ; AVX512-LABEL: concat_roundps_v16f32_v8f32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vroundps $4, %ymm0, %ymm0
-; AVX512-NEXT:    vroundps $4, %ymm1, %ymm1
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512-NEXT:    vrndscaleps $4, %zmm0, %zmm0
 ; AVX512-NEXT:    retq
   %v0 = call <8 x float> @llvm.x86.avx.round.ps.256(<8 x float> %a0, i32 4)
   %v1 = call <8 x float> @llvm.x86.avx.round.ps.256(<8 x float> %a1, i32 4)