From a8dd6f4f309b55c43a5b263c3aa178fa5bca4b30 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Sat, 30 Sep 2017 17:57:34 +0000
Subject: [PATCH] [X86][SSE] Fold (VSRAI (VSHLI X, C1), C1) --> X iff
 NumSignBits(X) > C1

Remove sign extend in register style pattern if the sign is already extended enough

llvm-svn: 314599
---
 llvm/lib/Target/X86/X86ISelLowering.cpp |  9 +++++++++
 llvm/test/CodeGen/X86/vector-trunc.ll   | 28 ++++------------------------
 2 files changed, 13 insertions(+), 24 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index b927178..2d7cf5c 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -31845,6 +31845,15 @@ static SDValue combineVectorShiftImm(SDNode *N, SelectionDAG &DAG,
       N0.getOpcode() == X86ISD::VSRAI)
     return DAG.getNode(X86ISD::VSRLI, SDLoc(N), VT, N0.getOperand(0), N1);
 
+  // fold (VSRAI (VSHLI X, C1), C1) --> X iff NumSignBits(X) > C1
+  if (Opcode == X86ISD::VSRAI && N0.getOpcode() == X86ISD::VSHLI &&
+      N1 == N0.getOperand(1)) {
+    SDValue N00 = N0.getOperand(0);
+    unsigned NumSignBits = DAG.ComputeNumSignBits(N00);
+    if (ShiftVal.ult(NumSignBits))
+      return N00;
+  }
+
   // We can decode 'whole byte' logical bit shifts as shuffles.
   if (LogicalShift && (ShiftVal.getZExtValue() % 8) == 0) {
     SDValue Op(N, 0);
diff --git a/llvm/test/CodeGen/X86/vector-trunc.ll b/llvm/test/CodeGen/X86/vector-trunc.ll
index 5548ee1..bb8fd19 100644
--- a/llvm/test/CodeGen/X86/vector-trunc.ll
+++ b/llvm/test/CodeGen/X86/vector-trunc.ll
@@ -389,11 +389,7 @@ entry:
 define <8 x i16> @trunc8i32_8i16_ashr(<8 x i32> %a) {
 ; SSE2-LABEL: trunc8i32_8i16_ashr:
 ; SSE2:       # BB#0: # %entry
-; SSE2-NEXT:    psrad $16, %xmm0
-; SSE2-NEXT:    psrad $16, %xmm1
-; SSE2-NEXT:    pslld $16, %xmm1
 ; SSE2-NEXT:    psrad $16, %xmm1
-; SSE2-NEXT:    pslld $16, %xmm0
 ; SSE2-NEXT:    psrad $16, %xmm0
 ; SSE2-NEXT:    packssdw %xmm1, %xmm0
 ; SSE2-NEXT:    retq
@@ -727,40 +723,24 @@ entry:
 define void @trunc16i32_16i16_ashr(<16 x i32> %a) {
 ; SSE2-LABEL: trunc16i32_16i16_ashr:
 ; SSE2:       # BB#0: # %entry
-; SSE2-NEXT:    psrad $16, %xmm2
 ; SSE2-NEXT:    psrad $16, %xmm3
-; SSE2-NEXT:    psrad $16, %xmm0
-; SSE2-NEXT:    psrad $16, %xmm1
-; SSE2-NEXT:    pslld $16, %xmm1
+; SSE2-NEXT:    psrad $16, %xmm2
+; SSE2-NEXT:    packssdw %xmm3, %xmm2
 ; SSE2-NEXT:    psrad $16, %xmm1
-; SSE2-NEXT:    pslld $16, %xmm0
 ; SSE2-NEXT:    psrad $16, %xmm0
 ; SSE2-NEXT:    packssdw %xmm1, %xmm0
-; SSE2-NEXT:    pslld $16, %xmm3
-; SSE2-NEXT:    psrad $16, %xmm3
-; SSE2-NEXT:    pslld $16, %xmm2
-; SSE2-NEXT:    psrad $16, %xmm2
-; SSE2-NEXT:    packssdw %xmm3, %xmm2
 ; SSE2-NEXT:    movdqu %xmm2, (%rax)
 ; SSE2-NEXT:    movdqu %xmm0, (%rax)
 ; SSE2-NEXT:    retq
 ;
 ; SSSE3-LABEL: trunc16i32_16i16_ashr:
 ; SSSE3:       # BB#0: # %entry
-; SSSE3-NEXT:    psrad $16, %xmm2
 ; SSSE3-NEXT:    psrad $16, %xmm3
-; SSSE3-NEXT:    psrad $16, %xmm0
-; SSSE3-NEXT:    psrad $16, %xmm1
-; SSSE3-NEXT:    pslld $16, %xmm1
+; SSSE3-NEXT:    psrad $16, %xmm2
+; SSSE3-NEXT:    packssdw %xmm3, %xmm2
 ; SSSE3-NEXT:    psrad $16, %xmm1
-; SSSE3-NEXT:    pslld $16, %xmm0
 ; SSSE3-NEXT:    psrad $16, %xmm0
 ; SSSE3-NEXT:    packssdw %xmm1, %xmm0
-; SSSE3-NEXT:    pslld $16, %xmm3
-; SSSE3-NEXT:    psrad $16, %xmm3
-; SSSE3-NEXT:    pslld $16, %xmm2
-; SSSE3-NEXT:    psrad $16, %xmm2
-; SSSE3-NEXT:    packssdw %xmm3, %xmm2
 ; SSSE3-NEXT:    movdqu %xmm2, (%rax)
 ; SSSE3-NEXT:    movdqu %xmm0, (%rax)
 ; SSSE3-NEXT:    retq
-- 
2.7.4