From dd6ca639d57fce79fffb74caa35c72c91a6f83b0 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Tue, 6 Dec 2016 19:09:37 +0000
Subject: [PATCH] [DAGCombine] Add (sext_in_reg (zext x)) -> (sext x) combine

Handle the case where a sign extension has ended up being split into separate stages (typically to get around vector legal ops) and a zext + sext_in_reg gets inserted.

Differential Revision: https://reviews.llvm.org/D27461

llvm-svn: 288842
---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp |  9 +++++++++
 llvm/test/CodeGen/X86/combine-sext-in-reg.ll  | 12 ------------
 2 files changed, 9 insertions(+), 12 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 590b10b..0fa0b91 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -7137,6 +7137,15 @@ SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
       return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
   }
 
+  // fold (sext_in_reg (zext x)) -> (sext x)
+  // iff we are extending the source sign bit.
+  if (N0.getOpcode() == ISD::ZERO_EXTEND) {
+    SDValue N00 = N0.getOperand(0);
+    if (N00.getScalarValueSizeInBits() == EVTBits &&
+        (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
+      return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
+  }
+
   // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
   if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
     return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT.getScalarType());
diff --git a/llvm/test/CodeGen/X86/combine-sext-in-reg.ll b/llvm/test/CodeGen/X86/combine-sext-in-reg.ll
index 9407d1b..3e60f3b 100644
--- a/llvm/test/CodeGen/X86/combine-sext-in-reg.ll
+++ b/llvm/test/CodeGen/X86/combine-sext-in-reg.ll
@@ -30,24 +30,12 @@ define <4 x i64> @sextinreg_zext_sext_v16i8_4i64(<16 x i8> %a0) {
 ; SSE-NEXT:    pmovsxbq %xmm0, %xmm2
 ; SSE-NEXT:    psrld $16, %xmm0
 ; SSE-NEXT:    pmovsxbq %xmm0, %xmm1
-; SSE-NEXT:    psllq $32, %xmm2
-; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
-; SSE-NEXT:    psrad $31, %xmm2
-; SSE-NEXT:    pblendw {{.*#+}} xmm2 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
-; SSE-NEXT:    psllq $32, %xmm1
-; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
-; SSE-NEXT:    psrad $31, %xmm1
-; SSE-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
 ; SSE-NEXT:    movdqa %xmm2, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: sextinreg_zext_sext_v16i8_4i64:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vpmovsxbq %xmm0, %ymm0
-; AVX-NEXT:    vpsllq $32, %ymm0, %ymm0
-; AVX-NEXT:    vpsrad $31, %ymm0, %ymm1
-; AVX-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[1,1,3,3,5,5,7,7]
-; AVX-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
 ; AVX-NEXT:    retq
   %1 = shufflevector <16 x i8> %a0, <16 x i8> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %2 = sext <4 x i8> %1 to <4 x i32>
-- 
2.7.4