We recently added IR intrinsics for min/max/abs, and there are several areas that need to be updated to support and optimize these. Please comment here if you are or are planning to work on any part of this, so we do not duplicate effort. Partial list: * ConstantFolding * ValueTracking * InstSimplify * InstCombine * PatternMatch * DemandedBits * Vectorization * Cost Modeling * LVI, CVP, SCCP * SCEV * IndVars
I'm working on ConstantFolding and InstSimplify currently.
I'm working on computeKnownBits for abs
Default + X86 costs have been committed.
SLP - basic abs/min/max vectorization works fine I'm going to look at adding min/max intrinsic reduction support
This was the last InstSimplify that was on my list to adapt: https://reviews.llvm.org/rG1470ce4a76fc Constant folding should be complete too.
I think we have most of it covered now. I'm not as familiar with the last 3 on the list (LVI/CVP/SCCP, SCEV, IndVars). What needs to be done before we start producing intrinsics in place of cmp+select in instcombine transforms?
The main thing that still needs to be done is recognizing min/max intrinsics in SCEV.
First try to get abs() canonicalized: https://reviews.llvm.org/D87188 As I mentioned there, I don't think we have our *size* cost models updated (but let me know if I'm misreading the code). We updated the throughput models here: https://github.com/llvm/llvm-project/commit/c5ef1f1edd4d539df530b07daf7a0c8b94d0efc8 https://github.com/llvm/llvm-project/commit/0a0f28254ab7ae5cc9e1868a0ef1ba03d945434d That's what is used by the vectorizers, but that's probably not what is used by unrolling, inlining, and simplifycfg.
The basic cost model handling is in place as of: https://reviews.llvm.org/D90554 / https://reviews.llvm.org/rGe32bd3512043 Still trying to get abs canonicalization in with: https://reviews.llvm.org/D87188
(In reply to Sanjay Patel from comment #9) > The basic cost model handling is in place as of: > https://reviews.llvm.org/D90554 / https://reviews.llvm.org/rGe32bd3512043 > Still trying to get abs canonicalization in with: > https://reviews.llvm.org/D87188 Done.
Awesome! Only part left now is to canonicalize min/max... I took a quick look at that, here's how the InstCombine test diff looks like: https://gist.github.com/nikic/f69061736ea8e6b6c159b28b09991599 We're still missing quite a few folds, e.g. pushing nots across min/max and forming unsigned saturating math.
Over in bug 48816, we discovered that we need narrowing combines for abs to avoid regressions. I'll work on the related folds for min/max.
(In reply to Sanjay Patel from comment #12) > I'll work on the related folds for min/max. Those were: https://reviews.llvm.org/rG09a136bcc694 https://reviews.llvm.org/rG0ce2920f1707
This is likely the best view of current state: https://reviews.llvm.org/D98152 There are still missing parts of analysis/instcombines as noted. I think SLP is done (and the sooner we can switch to intrinsics, the better because supporting cmp+select is still at risk for hard-to-spot bugs as shown in https://reviews.llvm.org/D99753 ). AFAIK, LoopVectorizer is partially done. I'll add some PhaseOrdering tests to increase visibility.
Added CVP handling for abs/min/max
What about saturating intrinsics? The same pattern matching discrepancy seems to apply, no?