From d9e2ef1edc3dfa8e6917148dabe58d0cba4a848a Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Thu, 28 Mar 2019 11:34:21 +0000
Subject: [PATCH] [DAGCombiner] Fold truncate(build_vector(x,y)) ->
 build_vector(truncate(x),truncate(y))

If scalar truncates are free, attempt to pre-truncate build_vectors source operands.

Only attempt to do this before legalization as we often end up with truncations/extensions during build_vector lowering.

Differential Revision: https://reviews.llvm.org/D59654

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@357161 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/CodeGen/SelectionDAG/DAGCombiner.cpp  | 16 +++++++++++++++-
 test/CodeGen/X86/bool-vector.ll           | 20 ++++++++++----------
 test/CodeGen/X86/known-signbits-vector.ll | 15 ++++-----------
 3 files changed, 29 insertions(+), 22 deletions(-)

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index ded4f04c57a..1c44a5d5ce3 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -9895,10 +9895,11 @@ SDValue DAGCombiner::visitZERO_EXTEND_VECTOR_INREG(SDNode *N) {
 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
   SDValue N0 = N->getOperand(0);
   EVT VT = N->getValueType(0);
+  EVT SrcVT = N0.getValueType();
   bool isLE = DAG.getDataLayout().isLittleEndian();
 
   // noop truncate
-  if (N0.getValueType() == N->getValueType(0))
+  if (SrcVT == VT)
     return N0;
 
   // fold (truncate (truncate x)) -> (truncate x)
@@ -9999,6 +10000,19 @@ SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
     }
   }
 
+  // Attempt to pre-truncate BUILD_VECTOR sources.
+  if (N0.getOpcode() == ISD::BUILD_VECTOR && !LegalOperations &&
+      TLI.isTruncateFree(SrcVT.getScalarType(), VT.getScalarType())) {
+    SDLoc DL(N);
+    EVT SVT = VT.getScalarType();
+    SmallVector<SDValue, 8> TruncOps;
+    for (const SDValue &Op : N0->op_values()) {
+      SDValue TruncOp = DAG.getNode(ISD::TRUNCATE, DL, SVT, Op);
+      TruncOps.push_back(TruncOp);
+    }
+    return DAG.getBuildVector(VT, DL, TruncOps);
+  }
+
   // Fold a series of buildvector, bitcast, and truncate if possible.
   // For example fold
   //   (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
diff --git a/test/CodeGen/X86/bool-vector.ll b/test/CodeGen/X86/bool-vector.ll
index 1afa3bc5eb7..7850d06bd7d 100644
--- a/test/CodeGen/X86/bool-vector.ll
+++ b/test/CodeGen/X86/bool-vector.ll
@@ -77,17 +77,17 @@ define i32 @PR15215_good(<4 x i32> %input) {
 ; X32-NEXT:    pushl %esi
 ; X32-NEXT:    .cfi_def_cfa_offset 8
 ; X32-NEXT:    .cfi_offset %esi, -8
-; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    andl $1, %eax
-; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
-; X32-NEXT:    andl $1, %ecx
-; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %edx
-; X32-NEXT:    andl $1, %edx
-; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %esi
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %esi
 ; X32-NEXT:    andl $1, %esi
-; X32-NEXT:    leal (%eax,%ecx,2), %eax
-; X32-NEXT:    leal (%eax,%edx,4), %eax
-; X32-NEXT:    leal (%eax,%esi,8), %eax
+; X32-NEXT:    andl $1, %edx
+; X32-NEXT:    andl $1, %ecx
+; X32-NEXT:    andl $1, %eax
+; X32-NEXT:    leal (%esi,%edx,2), %edx
+; X32-NEXT:    leal (%edx,%ecx,4), %ecx
+; X32-NEXT:    leal (%ecx,%eax,8), %eax
 ; X32-NEXT:    popl %esi
 ; X32-NEXT:    .cfi_def_cfa_offset 4
 ; X32-NEXT:    retl
diff --git a/test/CodeGen/X86/known-signbits-vector.ll b/test/CodeGen/X86/known-signbits-vector.ll
index 0e6c067c159..1a6217f62f7 100644
--- a/test/CodeGen/X86/known-signbits-vector.ll
+++ b/test/CodeGen/X86/known-signbits-vector.ll
@@ -36,17 +36,10 @@ define <4 x float> @signbits_sext_v4i64_sitofp_v4f32(i8 signext %a0, i16 signext
 ;
 ; X64-LABEL: signbits_sext_v4i64_sitofp_v4f32:
 ; X64:       # %bb.0:
-; X64-NEXT:    movslq %edi, %rax
-; X64-NEXT:    movslq %esi, %rsi
-; X64-NEXT:    movslq %edx, %rdx
-; X64-NEXT:    movslq %ecx, %rcx
-; X64-NEXT:    vmovq %rcx, %xmm0
-; X64-NEXT:    vmovq %rdx, %xmm1
-; X64-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; X64-NEXT:    vmovq %rsi, %xmm1
-; X64-NEXT:    vmovq %rax, %xmm2
-; X64-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
-; X64-NEXT:    vshufps {{.*#+}} xmm0 = xmm1[0,2],xmm0[0,2]
+; X64-NEXT:    vmovd %edi, %xmm0
+; X64-NEXT:    vpinsrd $1, %esi, %xmm0, %xmm0
+; X64-NEXT:    vpinsrd $2, %edx, %xmm0, %xmm0
+; X64-NEXT:    vpinsrd $3, %ecx, %xmm0, %xmm0
 ; X64-NEXT:    vcvtdq2ps %xmm0, %xmm0
 ; X64-NEXT:    retq
   %1 = sext i8 %a0 to i64
-- 
2.50.1