[SystemZ] Fix choice of known-zero mask in insertion optimization

The backend converts 64-bit ORs into subreg moves if the upper 32 bits
of one operand and the low 32 bits of the other are known to be zero.
It then tries to peel away redundant ANDs from the upper 32 bits.

Since AND masks are canonicalized to exclude known-zero bits,
the test ORs the mask and the known-zero bits together before
checking for redundancy.  The problem was that it was using the
wrong node when checking for known-zero bits, so could drop ANDs
that were still needed.

llvm-svn: 196267
This commit is contained in:
Richard Sandiford 2013-12-03 11:01:54 +00:00
parent a48a12cf81
commit ccc2a7c1a0
2 changed files with 17 additions and 4 deletions

View File

@ -1933,10 +1933,10 @@ SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const {
// high 32 bits and just masks out low bits. We can skip it if so.
if (HighOp.getOpcode() == ISD::AND &&
HighOp.getOperand(1).getOpcode() == ISD::Constant) {
ConstantSDNode *MaskNode = cast<ConstantSDNode>(HighOp.getOperand(1));
uint64_t Mask = MaskNode->getZExtValue() | Masks[High];
if ((Mask >> 32) == 0xffffffff)
HighOp = HighOp.getOperand(0);
SDValue HighOp0 = HighOp.getOperand(0);
uint64_t Mask = cast<ConstantSDNode>(HighOp.getOperand(1))->getZExtValue();
if (DAG.MaskedValueIsZero(HighOp0, APInt(64, ~(Mask | 0xffffffff))))
HighOp = HighOp0;
}
// Take advantage of the fact that all GR32 operations only change the

View File

@ -165,3 +165,16 @@ define i64 @f13(i64 %a, i32 %b) {
%or = or i64 %shift, %low
ret i64 %or
}
; We previously wrongly removed the upper AND as dead.
define i64 @f14(i64 %a, i64 %b) {
; CHECK-LABEL: f14:
; CHECK: risbg {{%r[0-5]}}, %r2, 6, 134, 0
; CHECK: br %r14
%and1 = and i64 %a, 144115188075855872
%and2 = and i64 %b, 15
%or = or i64 %and1, %and2
%res = icmp eq i64 %or, 0
%ext = sext i1 %res to i64
ret i64 %ext
}