Add support for AssertSext and AssertZext, folding other extensions with
them. This allows for elminination of redundant extends in the entry
blocks of functions on PowerPC.
Add support for i32 x i32 -> i64 multiplies, by recognizing when the inputs
to ISD::MUL in ExpandOp are actually just extended i32 values and not real
i64 values. this allows us to codegen
int mulhs(int a, int b) { return ((long long)a * b) >> 32; }
as:
_mulhs:
mulhw r3, r4, r3
blr
instead of:
_mulhs:
mulhwu r2, r4, r3
srawi r5, r3, 31
mullw r5, r4, r5
add r2, r2, r5
srawi r4, r4, 31
mullw r3, r4, r3
add r3, r2, r3
blr
with a similar improvement on x86.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23147 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index dab757e..c929806 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -631,7 +631,8 @@
unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
MVT::ValueType ExtDstTy = N1.getValueType();
unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
-
+
+ if (Cond == ISD::SETEQ || Cond == ISD::SETNE) {
// If the extended part has any inconsistent bits, it cannot ever
// compare equal. In other words, they have to be all ones or all
// zeros.
@@ -644,6 +645,7 @@
return getSetCC(VT, getZeroExtendInReg(N1.getOperand(0), ExtSrcTy),
getConstant(C2 & (~0ULL >> 64-ExtSrcTyBits), ExtDstTy),
Cond);
+ }
}
uint64_t MinVal, MaxVal;
@@ -1192,6 +1194,8 @@
assert(EVT <= VT && "Not rounding down!");
break;
}
+ case ISD::AssertSext:
+ case ISD::AssertZext:
case ISD::SIGN_EXTEND_INREG: {
MVT::ValueType EVT = cast<VTSDNode>(N2)->getVT();
assert(VT == N1.getValueType() && "Not an inreg extend!");
@@ -1394,6 +1398,14 @@
// we know the result of the AND will be the AND mask itself.
return N2;
}
+ } else if (N1.getOpcode() == ISD::AssertZext) {
+ // If we are masking out the part of our input that was already masked
+ // out, just return the input directly.
+ unsigned ExtendBits =
+ MVT::getSizeInBits(cast<VTSDNode>(N1.getOperand(1))->getVT());
+ uint64_t ExtendMask = (1ULL << ExtendBits) - 1;
+ if (ExtendMask == C2)
+ return N1.getOperand(0);
}
break;
case ISD::OR:
@@ -1579,14 +1591,15 @@
if (EVT == VT) return N1; // Not actually extending
// If we are sign extending an extension, use the original source.
- if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG)
+ if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG ||
+ N1.getOpcode() == ISD::AssertSext)
if (cast<VTSDNode>(N1.getOperand(1))->getVT() <= EVT)
return N1;
// If we are sign extending a sextload, return just the load.
if (N1.getOpcode() == ISD::SEXTLOAD)
if (cast<VTSDNode>(N1.getOperand(3))->getVT() <= EVT)
- return N1;
+ return N1;
// If we are extending the result of a setcc, and we already know the
// contents of the top bits, eliminate the extension.
@@ -1594,7 +1607,7 @@
TLI.getSetCCResultContents() ==
TargetLowering::ZeroOrNegativeOneSetCCResult)
return N1;
-
+
// If we are sign extending the result of an (and X, C) operation, and we
// know the extended bits are zeros already, don't do the extend.
if (N1.getOpcode() == ISD::AND)
@@ -2195,6 +2208,8 @@
case ISD::VALUETYPE: return "ValueType";
case ISD::EntryToken: return "EntryToken";
case ISD::TokenFactor: return "TokenFactor";
+ case ISD::AssertSext: return "AssertSext";
+ case ISD::AssertZext: return "AssertZext";
case ISD::Constant: return "Constant";
case ISD::TargetConstant: return "TargetConstant";
case ISD::ConstantFP: return "ConstantFP";