From 5e8c3f328f5ea227c3770ca550c55245d5ae4d91 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Tue, 24 Oct 2017 17:04:57 +0000
Subject: [PATCH] [X86][AVX] ComputeNumSignBitsForTargetNode - add support for
 X86ISD::VTRUNC

llvm-svn: 316462
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 10 ++++++++++
 llvm/test/CodeGen/X86/avg-mask.ll       |  8 ++------
 llvm/test/CodeGen/X86/vselect-packss.ll | 12 +++---------
 3 files changed, 15 insertions(+), 15 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index e3501e4..c0215d2 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -27239,6 +27239,16 @@ unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
     return Tmp;
   }
 
+  case X86ISD::VTRUNC: {
+    SDValue Src = Op.getOperand(0);
+    unsigned NumSrcBits = Src.getScalarValueSizeInBits();
+    assert(VTBits < NumSrcBits && "Illegal truncation input type");
+    unsigned Tmp = DAG.ComputeNumSignBits(Src, Depth + 1);
+    if (Tmp > (NumSrcBits - VTBits))
+      return Tmp - (NumSrcBits - VTBits);
+    return 1;
+  }
+
   case X86ISD::PACKSS: {
     // PACKSS is just a truncation if the sign bits extend to the packed size.
     // TODO: Add DemandedElts support.
diff --git a/llvm/test/CodeGen/X86/avg-mask.ll b/llvm/test/CodeGen/X86/avg-mask.ll
index 6672a83..f594494 100644
--- a/llvm/test/CodeGen/X86/avg-mask.ll
+++ b/llvm/test/CodeGen/X86/avg-mask.ll
@@ -254,9 +254,7 @@ define <8 x i16> @avg_v8i16_mask(<8 x i16> %a, <8 x i16> %b, <8 x i16> %src, i8
 ; AVX512F-NEXT:    kmovw %edi, %k1
 ; AVX512F-NEXT:    vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
 ; AVX512F-NEXT:    vpmovqw %zmm1, %xmm1
-; AVX512F-NEXT:    vpand %xmm1, %xmm0, %xmm0
-; AVX512F-NEXT:    vpandn %xmm2, %xmm1, %xmm1
-; AVX512F-NEXT:    vpor %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT:    vpblendvb %xmm1, %xmm0, %xmm2, %xmm0
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
@@ -311,9 +309,7 @@ define <16 x i16> @avg_v16i16_mask(<16 x i16> %a, <16 x i16> %b, <16 x i16> %src
 ; AVX512F-NEXT:    kmovw %edi, %k1
 ; AVX512F-NEXT:    vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
 ; AVX512F-NEXT:    vpmovdw %zmm1, %ymm1
-; AVX512F-NEXT:    vpand %ymm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vpandn %ymm2, %ymm1, %ymm1
-; AVX512F-NEXT:    vpor %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT:    vpblendvb %ymm1, %ymm0, %ymm2, %ymm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512BWVL-LABEL: avg_v16i16_mask:
diff --git a/llvm/test/CodeGen/X86/vselect-packss.ll b/llvm/test/CodeGen/X86/vselect-packss.ll
index a93e84d..17bc671 100644
--- a/llvm/test/CodeGen/X86/vselect-packss.ll
+++ b/llvm/test/CodeGen/X86/vselect-packss.ll
@@ -54,9 +54,7 @@ define <16 x i8> @vselect_packss_v16i16(<16 x i16> %a0, <16 x i16> %a1, <16 x i8
 ; AVX512-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
 ; AVX512-NEXT:    vpmovsxwd %ymm0, %zmm0
 ; AVX512-NEXT:    vpmovdb %zmm0, %xmm0
-; AVX512-NEXT:    vpand %xmm2, %xmm0, %xmm1
-; AVX512-NEXT:    vpandn %xmm3, %xmm0, %xmm0
-; AVX512-NEXT:    vpor %xmm0, %xmm1, %xmm0
+; AVX512-NEXT:    vpblendvb %xmm0, %xmm2, %xmm3, %xmm0
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
   %1 = icmp eq <16 x i16> %a0, %a1
@@ -134,9 +132,7 @@ define <16 x i8> @vselect_packss_v16i32(<16 x i32> %a0, <16 x i32> %a1, <16 x i8
 ; AVX512-NEXT:    vpcmpeqd %zmm1, %zmm0, %k1
 ; AVX512-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512-NEXT:    vpmovdb %zmm0, %xmm0
-; AVX512-NEXT:    vpand %xmm2, %xmm0, %xmm1
-; AVX512-NEXT:    vpandn %xmm3, %xmm0, %xmm0
-; AVX512-NEXT:    vpor %xmm0, %xmm1, %xmm0
+; AVX512-NEXT:    vpblendvb %xmm0, %xmm2, %xmm3, %xmm0
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
   %1 = icmp eq <16 x i32> %a0, %a1
@@ -392,9 +388,7 @@ define <16 x i8> @vselect_packss_v16i64(<16 x i64> %a0, <16 x i64> %a1, <16 x i8
 ; AVX512-NEXT:    vpmovqd %zmm0, %ymm0
 ; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm8, %zmm0
 ; AVX512-NEXT:    vpmovdb %zmm0, %xmm0
-; AVX512-NEXT:    vpand %xmm4, %xmm0, %xmm1
-; AVX512-NEXT:    vpandn %xmm5, %xmm0, %xmm0
-; AVX512-NEXT:    vpor %xmm0, %xmm1, %xmm0
+; AVX512-NEXT:    vpblendvb %xmm0, %xmm4, %xmm5, %xmm0
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
   %1 = icmp eq <16 x i64> %a0, %a1
-- 
2.7.4