[llvm] r316462 - [X86][AVX] ComputeNumSignBitsForTargetNode - add support for X86ISD::VTRUNC

[Simon Pilgrim via llvm-commits]

Author: rksimon
Date: Tue Oct 24 10:04:57 2017
New Revision: 316462

URL: http://llvm.org/viewvc/llvm-project?rev=316462&view=rev
Log:
[X86][AVX] ComputeNumSignBitsForTargetNode - add support for X86ISD::VTRUNC

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
    llvm/trunk/test/CodeGen/X86/avg-mask.ll
    llvm/trunk/test/CodeGen/X86/vselect-packss.ll

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=316462&r1=316461&r2=316462&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Tue Oct 24 10:04:57 2017
@@ -27239,6 +27239,16 @@ unsigned X86TargetLowering::ComputeNumSi
     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.

Modified: llvm/trunk/test/CodeGen/X86/avg-mask.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/avg-mask.ll?rev=316462&r1=316461&r2=316462&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/avg-mask.ll (original)
+++ llvm/trunk/test/CodeGen/X86/avg-mask.ll Tue Oct 24 10:04:57 2017
@@ -254,9 +254,7 @@ define <8 x i16> @avg_v8i16_mask(<8 x i1
 ; 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
 ; 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:

Modified: llvm/trunk/test/CodeGen/X86/vselect-packss.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/vselect-packss.ll?rev=316462&r1=316461&r2=316462&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/vselect-packss.ll (original)
+++ llvm/trunk/test/CodeGen/X86/vselect-packss.ll Tue Oct 24 10:04:57 2017
@@ -54,9 +54,7 @@ define <16 x i8> @vselect_packss_v16i16(
 ; 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(
 ; 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(
 ; 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