[PATCH] D42899: [X86] Teach X86DAGToDAGISel::shrinkAndImmediate to preserve upper 32 zeroes of a 64 bit mask.

[Craig Topper via Phabricator via llvm-commits]

craig.topper created this revision.
craig.topper added reviewers: RKSimon, spatel.

If the upper 32 bits of a 64 bit mask are all zeros, we have special isel patterns to use a 32-bit and instead of a 64-bit and by relying on the impliciting zeroing of 32 bit ops.

This patch teachs shrinkAndImmediate not to break that optimization.


Repository:
  rL LLVM

https://reviews.llvm.org/D42899

Files:
  lib/Target/X86/X86ISelDAGToDAG.cpp
  test/CodeGen/X86/and-encoding.ll
  test/CodeGen/X86/shift-pair.ll


Index: test/CodeGen/X86/shift-pair.ll
===================================================================
--- test/CodeGen/X86/shift-pair.ll
+++ test/CodeGen/X86/shift-pair.ll
@@ -5,7 +5,7 @@
 ; CHECK-LABEL: test:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    shrq $54, %rdi
-; CHECK-NEXT:    andq $-4, %rdi
+; CHECK-NEXT:    andl $-4, %edi
 ; CHECK-NEXT:    movq %rdi, %rax
 ; CHECK-NEXT:    retq
     %B = lshr i64 %A, 56
Index: test/CodeGen/X86/and-encoding.ll
===================================================================
--- test/CodeGen/X86/and-encoding.ll
+++ test/CodeGen/X86/and-encoding.ll
@@ -61,7 +61,7 @@
 ; CHECK-LABEL: lopped64_32to8:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    shrq $36, %rdi # encoding: [0x48,0xc1,0xef,0x24]
-; CHECK-NEXT:    andq $-16, %rdi # encoding: [0x48,0x83,0xe7,0xf0]
+; CHECK-NEXT:    andl $-16, %edi # encoding: [0x83,0xe7,0xf0]
 ; CHECK-NEXT:    movq %rdi, %rax # encoding: [0x48,0x89,0xf8]
 ; CHECK-NEXT:    retq # encoding: [0xc3]
   %shr = lshr i64 %x, 36
Index: lib/Target/X86/X86ISelDAGToDAG.cpp
===================================================================
--- lib/Target/X86/X86ISelDAGToDAG.cpp
+++ lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -2489,19 +2489,31 @@
   // Bail out if the mask constant is already negative. It can't shrink more.
   APInt MaskVal = And1C->getAPIntValue();
   unsigned MaskLZ = MaskVal.countLeadingZeros();
-  if (!MaskLZ)
+  if (!MaskLZ || (VT == MVT::i64 && MaskLZ == 32))
     return false;
 
+  // Don't extend into the upper 32 bits of a 64 bit mask.
+  if (VT == MVT::i64 && MaskLZ >= 32) {
+    MaskLZ -= 32;
+    MaskVal = MaskVal.trunc(32);
+  }
+
   SDValue And0 = And->getOperand(0);
-  APInt HighZeros = APInt::getHighBitsSet(VT.getSizeInBits(), MaskLZ);
+  APInt HighZeros = APInt::getHighBitsSet(MaskVal.getBitWidth(), MaskLZ);
   APInt NegMaskVal = MaskVal | HighZeros;
 
   // If a negative constant would not allow a smaller encoding, there's no need
   // to continue. Only change the constant when we know it's a win.
   unsigned MinWidth = NegMaskVal.getMinSignedBits();
   if (MinWidth > 32 || (MinWidth > 8 && MaskVal.getMinSignedBits() <= 32))
     return false;
 
+  // Extend masks if we truncated above.
+  if (VT == MVT::i64 && MaskVal.getBitWidth() < 64) {
+    NegMaskVal = NegMaskVal.zext(64);
+    HighZeros = HighZeros.zext(64);
+  }
+
   // The variable operand must be all zeros in the top bits to allow using the
   // new, negative constant as the mask.
   if (!CurDAG->MaskedValueIsZero(And0, HighZeros))


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D42899.132784.patch
Type: text/x-patch
Size: 2526 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180205/60d47a04/attachment.bin>