summaryrefslogtreecommitdiffstats
path: root/llvm/lib/Target/ARM/ARMISelLowering.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/lib/Target/ARM/ARMISelLowering.cpp')
-rw-r--r--llvm/lib/Target/ARM/ARMISelLowering.cpp37
1 files changed, 37 insertions, 0 deletions
diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 8fb7ef4847f..a87770125fd 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -4073,6 +4073,43 @@ SDValue ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
std::swap(LHS, RHS);
}
+ // Thumb1 has very limited immediate modes, so turning an "and" into a
+ // shift can save multiple instructions.
+ //
+ // If we have (x & C1), and C1 is an appropriate mask, we can transform it
+ // into "((x << n) >> n)". But that isn't necessarily profitable on its
+ // own. If it's the operand to an unsigned comparison with an immediate,
+ // we can eliminate one of the shifts: we transform
+ // "((x << n) >> n) == C2" to "(x << n) == (C2 << n)".
+ //
+ // We avoid transforming cases which aren't profitable due to encoding
+ // details:
+ //
+ // 1. C2 fits into the immediate field of a cmp, and the transformed version
+ // would not; in that case, we're essentially trading one immediate load for
+ // another.
+ // 2. C1 is 255 or 65535, so we can use uxtb or uxth.
+ // 3. C2 is zero; we have other code for this special case.
+ //
+ // FIXME: Figure out profitability for Thumb2; we usually can't save an
+ // instruction, since the AND is always one instruction anyway, but we could
+ // use narrow instructions in some cases.
+ if (Subtarget->isThumb1Only() && LHS->getOpcode() == ISD::AND &&
+ LHS->hasOneUse() && isa<ConstantSDNode>(LHS.getOperand(1)) &&
+ LHS.getValueType() == MVT::i32 && isa<ConstantSDNode>(RHS) &&
+ !isSignedIntSetCC(CC)) {
+ unsigned Mask = cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue();
+ auto *RHSC = cast<ConstantSDNode>(RHS.getNode());
+ uint64_t RHSV = RHSC->getZExtValue();
+ if (isMask_32(Mask) && (RHSV & ~Mask) == 0 && Mask != 255 && Mask != 65535) {
+ unsigned ShiftBits = countLeadingZeros(Mask);
+ if (RHSV && (RHSV > 255 || (RHSV << ShiftBits) <= 255)) {
+ SDValue ShiftAmt = DAG.getConstant(ShiftBits, dl, MVT::i32);
+ LHS = DAG.getNode(ISD::SHL, dl, MVT::i32, LHS.getOperand(0), ShiftAmt);
+ RHS = DAG.getConstant(RHSV << ShiftBits, dl, MVT::i32);
+ }
+ }
+ }
ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
// If the RHS is a constant zero then the V (overflow) flag will never be
OpenPOWER on IntegriCloud