[llvm] r322454 - [X86] Add X86ISD::VTRUNC to computeKnownBitsForTargetNode.

[Craig Topper via llvm-commits]

Author: ctopper
Date: Sun Jan 14 00:11:33 2018
New Revision: 322454

URL: http://llvm.org/viewvc/llvm-project?rev=322454&view=rev
Log:
[X86] Add X86ISD::VTRUNC to computeKnownBitsForTargetNode.

We have to take special care to avoid the cases where the result of the truncate would be padded with zero elements.

Ideally we'd just use ISD::TRUNCATE for these cases instead.

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
    llvm/trunk/test/CodeGen/X86/avx512-trunc.ll

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=322454&r1=322453&r2=322454&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Sun Jan 14 00:11:33 2018
@@ -27837,6 +27837,18 @@ void X86TargetLowering::computeKnownBits
     Known.Zero.setBitsFrom(InBitWidth);
     break;
   }
+  case X86ISD::VTRUNC: {
+    // TODO: Add DemandedElts support.
+    SDValue N0 = Op.getOperand(0);
+    // We can only handle cases with the same number of elements. Otherwise
+    // the truncate fills with zero elements.
+    // TODO: Maybe we could just discard any 1s we found instead of skipping?
+    if (VT.getVectorNumElements() != N0.getValueType().getVectorNumElements())
+      break;
+    DAG.computeKnownBits(N0, Known, Depth+1);
+    Known = Known.trunc(BitWidth);
+    break;
+  }
   case X86ISD::CMOV: {
     DAG.computeKnownBits(Op.getOperand(1), Known, Depth+1);
     // If we don't know any bits, early out.

Modified: llvm/trunk/test/CodeGen/X86/avx512-trunc.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/avx512-trunc.ll?rev=322454&r1=322453&r2=322454&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/avx512-trunc.ll (original)
+++ llvm/trunk/test/CodeGen/X86/avx512-trunc.ll Sun Jan 14 00:11:33 2018
@@ -748,7 +748,7 @@ define <16 x i8> @usat_trunc_db_256(<8 x
 ; KNL-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255]
 ; KNL-NEXT:    vpminud %ymm1, %ymm0, %ymm0
 ; KNL-NEXT:    vpmovdw %zmm0, %ymm0
-; KNL-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u]
+; KNL-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
 ; KNL-NEXT:    vzeroupper
 ; KNL-NEXT:    retq
 ;
@@ -756,7 +756,7 @@ define <16 x i8> @usat_trunc_db_256(<8 x
 ; SKX:       ## %bb.0:
 ; SKX-NEXT:    vpminud {{.*}}(%rip){1to8}, %ymm0, %ymm0
 ; SKX-NEXT:    vpmovdw %ymm0, %xmm0
-; SKX-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u]
+; SKX-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
 ; SKX-NEXT:    vzeroupper
 ; SKX-NEXT:    retq
   %tmp1 = icmp ult <8 x i32> %x, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>