From: Craig Topper <craig.topper@intel.com>
Date: Mon, 28 Aug 2017 15:28:33 +0000 (+0000)
Subject: [DAGCombiner] Teach visitEXTRACT_SUBVECTOR to turn extracts of BUILD_VECTOR into... 
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=337c2dfa0bd216d0c211e5f4ab44442e97c20e77;p=llvm

[DAGCombiner] Teach visitEXTRACT_SUBVECTOR to turn extracts of BUILD_VECTOR into smaller BUILD_VECTORs

Only do this before operations are legalized of BUILD_VECTOR is Legal for the target.

Differential Revision: https://reviews.llvm.org/D37186

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311892 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index df38ef423fa..306fd6f4984 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -15157,6 +15157,29 @@ SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
   // Skip bitcasting
   V = peekThroughBitcast(V);
 
+  // If the input is a build vector. Try to make a smaller build vector.
+  if (V->getOpcode() == ISD::BUILD_VECTOR) {
+    if (auto *Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
+      EVT InVT = V->getValueType(0);
+      unsigned NumElems = NVT.getSizeInBits() / InVT.getScalarSizeInBits();
+      if (NumElems > 0) {
+        EVT ExtractVT = EVT::getVectorVT(*DAG.getContext(),
+                                         InVT.getVectorElementType(), NumElems);
+        if (!LegalOperations ||
+            TLI.isOperationLegal(ISD::BUILD_VECTOR, ExtractVT)) {
+          unsigned IdxVal = Idx->getZExtValue() * NVT.getScalarSizeInBits() /
+                            InVT.getScalarSizeInBits();
+
+          // Extract the pieces from the original build_vector.
+          SDValue BuildVec = DAG.getBuildVector(ExtractVT, SDLoc(N),
+                                            makeArrayRef(V->op_begin() + IdxVal,
+                                                         NumElems));
+          return DAG.getBitcast(NVT, BuildVec);
+        }
+      }
+    }
+  }
+
   if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
     // Handle only simple case where vector being inserted and vector
     // being extracted are of same size.
diff --git a/test/CodeGen/X86/fold-vector-sext-zext.ll b/test/CodeGen/X86/fold-vector-sext-zext.ll
index 575bd5897e4..39e728816b0 100644
--- a/test/CodeGen/X86/fold-vector-sext-zext.ll
+++ b/test/CodeGen/X86/fold-vector-sext-zext.ll
@@ -83,8 +83,7 @@ define <4 x i32> @test_sext_4i8_4i32_undef() {
 define <4 x i64> @test_sext_4i8_4i64() {
 ; X32-LABEL: test_sext_4i8_4i64:
 ; X32:       # BB#0:
-; X32-NEXT:    vmovaps {{.*#+}} xmm0 = [0,0,4294967295,4294967295]
-; X32-NEXT:    vinsertf128 $1, {{\.LCPI.*}}, %ymm0, %ymm0
+; X32-NEXT:    vmovaps {{.*#+}} ymm0 = [0,0,4294967295,4294967295,2,0,4294967293,4294967295]
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: test_sext_4i8_4i64:
@@ -102,8 +101,7 @@ define <4 x i64> @test_sext_4i8_4i64() {
 define <4 x i64> @test_sext_4i8_4i64_undef() {
 ; X32-LABEL: test_sext_4i8_4i64_undef:
 ; X32:       # BB#0:
-; X32-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; X32-NEXT:    vinsertf128 $1, {{\.LCPI.*}}, %ymm0, %ymm0
+; X32-NEXT:    vmovaps {{.*#+}} ymm0 = <u,u,4294967295,4294967295,u,u,4294967293,4294967295>
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: test_sext_4i8_4i64_undef:
@@ -245,8 +243,7 @@ define <4 x i32> @test_zext_4i8_4i32() {
 define <4 x i64> @test_zext_4i8_4i64() {
 ; X32-LABEL: test_zext_4i8_4i64:
 ; X32:       # BB#0:
-; X32-NEXT:    vmovaps {{.*#+}} xmm0 = [0,0,255,0]
-; X32-NEXT:    vinsertf128 $1, {{\.LCPI.*}}, %ymm0, %ymm0
+; X32-NEXT:    vmovaps {{.*#+}} ymm0 = [0,0,255,0,2,0,253,0]
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: test_zext_4i8_4i64:
@@ -300,10 +297,7 @@ define <4 x i32> @test_zext_4i8_4i32_undef() {
 define <4 x i64> @test_zext_4i8_4i64_undef() {
 ; X32-LABEL: test_zext_4i8_4i64_undef:
 ; X32:       # BB#0:
-; X32-NEXT:    vmovaps {{.*#+}} xmm0 = <u,u,255,0>
-; X32-NEXT:    movl $2, %eax
-; X32-NEXT:    vmovd %eax, %xmm1
-; X32-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; X32-NEXT:    vmovaps {{.*#+}} ymm0 = <u,u,255,0,2,0,u,u>
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: test_zext_4i8_4i64_undef:
diff --git a/test/CodeGen/X86/widen_extract-1.ll b/test/CodeGen/X86/widen_extract-1.ll
index 3737ea96773..332dc940b17 100644
--- a/test/CodeGen/X86/widen_extract-1.ll
+++ b/test/CodeGen/X86/widen_extract-1.ll
@@ -7,8 +7,8 @@
 define void @convert(<2 x double>* %dst.addr, <3 x double> %src)  {
 ; X32-LABEL: convert:
 ; X32:       # BB#0: # %entry
-; X32-NEXT:    movups {{[0-9]+}}(%esp), %xmm0
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NEXT:    movups {{[0-9]+}}(%esp), %xmm0
 ; X32-NEXT:    movaps %xmm0, (%eax)
 ; X32-NEXT:    retl
 ;