From 396a07242341c06b69608187f1336b196551aa01 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Tue, 3 Oct 2017 09:54:03 +0000
Subject: [PATCH] [X86][SSE] Add support for shuffle combining from
 PACKSS/PACKUS

Mentioned in D38472

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@314777 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86ISelLowering.cpp |  4 ++++
 test/CodeGen/X86/vector-mul.ll     | 20 ++++----------------
 2 files changed, 8 insertions(+), 16 deletions(-)

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 3dd4d74ca40..4ba5c2c8c8d 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -31918,6 +31918,10 @@ static SDValue combineVectorPack(SDNode *N, SelectionDAG &DAG,
     return getConstVector(Bits, Undefs, VT.getSimpleVT(), DAG, SDLoc(N));
   }
 
+  // Attempt to combine as shuffle.
+  SDValue Op(N, 0);
+  combineX86ShufflesRecursively({Op}, 0, Op, {0}, {}, /*Depth*/ 1,
+                                /*HasVarMask*/ false, DAG, DCI, Subtarget);
   return SDValue();
 }
 
diff --git a/test/CodeGen/X86/vector-mul.ll b/test/CodeGen/X86/vector-mul.ll
index 19f87493d5c..95e64e9049d 100644
--- a/test/CodeGen/X86/vector-mul.ll
+++ b/test/CodeGen/X86/vector-mul.ll
@@ -358,13 +358,10 @@ define <16 x i8> @mul_v16i8_17(<16 x i8> %a0) nounwind {
 ; X64-XOP-NEXT:    vpmovsxbw %xmm0, %xmm1
 ; X64-XOP-NEXT:    vmovdqa {{.*#+}} xmm2 = [17,17,17,17,17,17,17,17]
 ; X64-XOP-NEXT:    vpmullw %xmm2, %xmm1, %xmm1
-; X64-XOP-NEXT:    vmovdqa {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
-; X64-XOP-NEXT:    vpand %xmm3, %xmm1, %xmm1
 ; X64-XOP-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 ; X64-XOP-NEXT:    vpmovsxbw %xmm0, %xmm0
 ; X64-XOP-NEXT:    vpmullw %xmm2, %xmm0, %xmm0
-; X64-XOP-NEXT:    vpand %xmm3, %xmm0, %xmm0
-; X64-XOP-NEXT:    vpackuswb %xmm0, %xmm1, %xmm0
+; X64-XOP-NEXT:    vpperm {{.*#+}} xmm0 = xmm1[0,2,4,6,8,10,12,14],xmm0[0,2,4,6,8,10,12,14]
 ; X64-XOP-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: mul_v16i8_17:
@@ -493,13 +490,10 @@ define <16 x i8> @mul_v16i8_2_3_9_17_33_65_129_2_3_9_17_33_65_129_2_3(<16 x i8>
 ; X64-XOP:       # BB#0:
 ; X64-XOP-NEXT:    vpmovsxbw %xmm0, %xmm1
 ; X64-XOP-NEXT:    vpmullw {{.*}}(%rip), %xmm1, %xmm1
-; X64-XOP-NEXT:    vmovdqa {{.*#+}} xmm2 = [255,255,255,255,255,255,255,255]
-; X64-XOP-NEXT:    vpand %xmm2, %xmm1, %xmm1
 ; X64-XOP-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 ; X64-XOP-NEXT:    vpmovsxbw %xmm0, %xmm0
 ; X64-XOP-NEXT:    vpmullw {{.*}}(%rip), %xmm0, %xmm0
-; X64-XOP-NEXT:    vpand %xmm2, %xmm0, %xmm0
-; X64-XOP-NEXT:    vpackuswb %xmm0, %xmm1, %xmm0
+; X64-XOP-NEXT:    vpperm {{.*#+}} xmm0 = xmm1[0,2,4,6,8,10,12,14],xmm0[0,2,4,6,8,10,12,14]
 ; X64-XOP-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: mul_v16i8_2_3_9_17_33_65_129_2_3_9_17_33_65_129_2_3:
@@ -637,13 +631,10 @@ define <16 x i8> @mul_v16i8_31(<16 x i8> %a0) nounwind {
 ; X64-XOP-NEXT:    vpmovsxbw %xmm0, %xmm1
 ; X64-XOP-NEXT:    vmovdqa {{.*#+}} xmm2 = [31,31,31,31,31,31,31,31]
 ; X64-XOP-NEXT:    vpmullw %xmm2, %xmm1, %xmm1
-; X64-XOP-NEXT:    vmovdqa {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
-; X64-XOP-NEXT:    vpand %xmm3, %xmm1, %xmm1
 ; X64-XOP-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 ; X64-XOP-NEXT:    vpmovsxbw %xmm0, %xmm0
 ; X64-XOP-NEXT:    vpmullw %xmm2, %xmm0, %xmm0
-; X64-XOP-NEXT:    vpand %xmm3, %xmm0, %xmm0
-; X64-XOP-NEXT:    vpackuswb %xmm0, %xmm1, %xmm0
+; X64-XOP-NEXT:    vpperm {{.*#+}} xmm0 = xmm1[0,2,4,6,8,10,12,14],xmm0[0,2,4,6,8,10,12,14]
 ; X64-XOP-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: mul_v16i8_31:
@@ -1005,13 +996,10 @@ define <16 x i8> @mul_v16i8_0_1_3_7_15_31_63_127_0_1_3_7_15_31_63_127(<16 x i8>
 ; X64-XOP-NEXT:    vpmovsxbw %xmm0, %xmm1
 ; X64-XOP-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1,3,7,15,31,63,127]
 ; X64-XOP-NEXT:    vpmullw %xmm2, %xmm1, %xmm1
-; X64-XOP-NEXT:    vmovdqa {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
-; X64-XOP-NEXT:    vpand %xmm3, %xmm1, %xmm1
 ; X64-XOP-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 ; X64-XOP-NEXT:    vpmovsxbw %xmm0, %xmm0
 ; X64-XOP-NEXT:    vpmullw %xmm2, %xmm0, %xmm0
-; X64-XOP-NEXT:    vpand %xmm3, %xmm0, %xmm0
-; X64-XOP-NEXT:    vpackuswb %xmm0, %xmm1, %xmm0
+; X64-XOP-NEXT:    vpperm {{.*#+}} xmm0 = xmm1[0,2,4,6,8,10,12,14],xmm0[0,2,4,6,8,10,12,14]
 ; X64-XOP-NEXT:    retq
 ;
 ; X64-AVX2-LABEL: mul_v16i8_0_1_3_7_15_31_63_127_0_1_3_7_15_31_63_127:
-- 
2.40.0