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

[craig.topper]

Summary

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.

[RKSimon]

Don't know this code that well - @spatel any comments?

lib/Target/X86/X86ISelDAGToDAG.cpp

2489

Update comment

[spatel]

In D42899#997736, @RKSimon wrote:

Don't know this code that well - @spatel any comments?

This is an improvement to D42088.

LGTM with the updated comment. I'd just copy the summary text as shown here as the code comment.

lib/Target/X86/X86ISelDAGToDAG.cpp

bool X86DAGToDAGISel::shrinkAndImmediate(SDNode *And) {
   MVT VT = And->getSimpleValueType(0);
   if (VT != MVT::i32 && VT != MVT::i64)
     return false;
 
   auto *And1C = dyn_cast<ConstantSDNode>(And->getOperand(1));
   if (!And1C)
     return false;
 
   // 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))
     return false;
 
   // Check if the mask is -1. In that case, this is an unnecessary instruction
   // that escaped earlier analysis.
   if (NegMaskVal.isAllOnesValue()) {

test/CodeGen/X86/and-encoding.ll

 }
 
 ; The immediate (0x0ffffff0) can be made into an i8 by making it negative.
 
 define i64 @lopped64_32to8(i64 %x) {
 ; 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
   %and = and i64 %shr, 268435440
   ret i64 %and
 }
 
 ; The immediate (0x0ffffffffffffff0) can be made into an i8 by making it negative.

test/CodeGen/X86/shift-pair.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-- | FileCheck %s
 
 define i64 @test(i64 %A) {
 ; 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
     %C = shl i64 %B, 2
     ret i64 %C
 }