[AArch64] Report icmp as free if it can be folded into ands #143286
Conversation
|
@llvm/pr-subscribers-llvm-analysis @llvm/pr-subscribers-backend-aarch64 Author: AZero13 (AZero13) ChangesSince changing the backend to fold x >= 1 / x < 1 -> x > 0 / x <= 0 and x <= -1 / x > -1 -> x > 0 / x <= 0, this should be reflected in the cost. Full diff: https://github.com/llvm/llvm-project/pull/143286.diff 1 Files Affected:
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 68aec80f07e1d..72b60eaa45b78 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -4346,14 +4346,23 @@ InstructionCost AArch64TTIImpl::getCmpSelInstrCost(
}
// Treat the icmp in icmp(and, 0) as free, as we can make use of ands.
- // FIXME: This can apply to more conditions and add/sub if it can be shown to
- // be profitable.
if (ValTy->isIntegerTy() && ISD == ISD::SETCC && I &&
- ICmpInst::isEquality(VecPred) &&
+ !CmpInst::isUnsigned(VecPred) &&
TLI->isTypeLegal(TLI->getValueType(DL, ValTy)) &&
- match(I->getOperand(1), m_Zero()) &&
- match(I->getOperand(0), m_And(m_Value(), m_Value())))
- return 0;
+ match(I->getOperand(0), m_And(m_Value(), m_Value()))) {
+ if (match(I->getOperand(1), m_Zero()))
+ return 0;
+
+ // x >= 1 / x < 1 -> x > 0 / x <= 0
+ if (match(I->getOperand(1), m_One()) &&
+ (VecPred == CmpInst::ICMP_SLT || VecPred == CmpInst::ICMP_SGE))
+ return 0;
+
+ // x <= -1 / x > -1 -> x > 0 / x <= 0
+ if (match(I->getOperand(1), m_AllOnes()) &&
+ (VecPred == CmpInst::ICMP_SLE || VecPred == CmpInst::ICMP_SGT))
+ return 0;
+ }
// The base case handles scalable vectors fine for now, since it treats the
// cost as 1 * legalization cost.
|
llvmbot
added
the
llvm:analysis
|
@davemgreen Thoughts? |
davemgreen
left a comment
davemgreen
left a comment
There was a problem hiding this comment.
Hi - I think this sounds OK, some of the cases will be canonicalized by the mid-end to compare against 0. Out of interest do you have some motivating example you are trying to optimize, or is this a general improvement as you were looking at the backend patterns?
Can you add some extra tests for each of the cases?
|
The cost model needs to be accurate as possible |
Since changing the backend to fold x >= 1 / x < 1 -> x > 0 / x <= 0 and x <= -1 / x > -1 -> x > 0 / x <= 0, this should be reflected in the cost.
SamTebbs33
left a comment
SamTebbs33
left a comment
There was a problem hiding this comment.
LGTM, thanks. Do let @davemgreen have another look though.
davemgreen
left a comment
davemgreen
left a comment
There was a problem hiding this comment.
Sure, LGTM. Are you happy for this to be merged?
Yes |
4 participants
Since changing the backend to fold x >= 1 / x < 1 -> x > 0 / x <= 0 and x <= -1 / x > -1 -> x > 0 / x <= 0, this should be reflected in the cost.