[llvm] r275262 - [X86][SSE] Check for lane crossing shuffles before trying to combine to PSHUFB

[Simon Pilgrim via llvm-commits]

Author: rksimon
Date: Wed Jul 13 07:48:41 2016
New Revision: 275262

URL: http://llvm.org/viewvc/llvm-project?rev=275262&view=rev
Log:
[X86][SSE] Check for lane crossing shuffles before trying to combine to PSHUFB

Removes a return-on-fail that was making it tricky to add other variable mask shuffles.

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=275262&r1=275261&r2=275262&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Wed Jul 13 07:48:41 2016
@@ -25082,6 +25082,14 @@ static bool combineX86ShuffleChain(SDVal
   if (MaskEltSizeInBits > 64)
     return false;
 
+  // Determine the effective mask value type.
+  bool FloatDomain =
+      (VT.isFloatingPoint() || (VT.is256BitVector() && !Subtarget.hasAVX2())) &&
+      (32 <= MaskEltSizeInBits);
+  MVT MaskVT = FloatDomain ? MVT::getFloatingPointVT(MaskEltSizeInBits)
+                           : MVT::getIntegerVT(MaskEltSizeInBits);
+  MaskVT = MVT::getVectorVT(MaskVT, NumMaskElts);
+
   // Attempt to match the mask against known shuffle patterns.
   MVT ShuffleVT;
   unsigned Shuffle, PermuteImm;
@@ -25130,11 +25138,7 @@ static bool combineX86ShuffleChain(SDVal
        (Subtarget.hasAVX() && VT.is256BitVector()))) {
     // Convert VT to a type compatible with X86ISD::BLENDI.
     // TODO - add 16i16 support (requires lane duplication).
-    bool FloatDomain = VT.isFloatingPoint();
-    MVT ShuffleVT = FloatDomain ? MVT::getFloatingPointVT(MaskEltSizeInBits)
-                                : MVT::getIntegerVT(MaskEltSizeInBits);
-    ShuffleVT = MVT::getVectorVT(ShuffleVT, NumMaskElts);
-
+    MVT ShuffleVT = MaskVT;
     if (Subtarget.hasAVX2()) {
       if (ShuffleVT == MVT::v4i64)
         ShuffleVT = MVT::v8i32;
@@ -25213,6 +25217,7 @@ static bool combineX86ShuffleChain(SDVal
   // instructions, but in practice PSHUFB tends to be *very* fast so we're
   // more aggressive.
   if ((Depth >= 3 || HasVariableMask) &&
+      !is128BitLaneCrossingShuffleMask(MaskVT, Mask) &&
       ((VT.is128BitVector() && Subtarget.hasSSSE3()) ||
        (VT.is256BitVector() && Subtarget.hasAVX2()) ||
        (VT.is512BitVector() && Subtarget.hasBWI()))) {
@@ -25230,9 +25235,7 @@ static bool combineX86ShuffleChain(SDVal
         continue;
       }
       M = Ratio * M + i % Ratio;
-      // Check that we are not crossing lanes.
-      if ((M / 16) != (i / 16))
-        return false;
+      assert ((M / 16) == (i / 16) && "Lane crossing detected");
       PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
     }
     MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);