As discussed in post-commit review in https://reviews.llvm.org/D73501
if the goal of this is to help vectorizer, then we should actually
be teaching vectorizer to do this, because right now this rewrite
is still budget-limited, which isn't what we'd want.
Additionally, while the rest of the patch series was universally profitable,
this particular patch is reportedly (https://reviews.llvm.org/D73501#1905171)
exposing cost-modeling issues on ARM.
So let's just back this particular patch out. Once there's an undo transform,
this could be considered for reintegration.
This reverts commit 44edc6fd2c.
Summary:
The widenIVUse avoids generating trunc by evaluating the use as AddRec, this
will not work when:
1) SCEV traces back to an instruction inside the loop that SCEV can not
expand, eg. add %indvar, (load %addr)
2) SCEV finds a loop variant, eg. add %indvar, %loopvariant
While SCEV fails to avoid trunc, we can still try to use instruction
combining approach to prove trunc is not required. This can be further
extended with other instruction combining checks, but for now we handle the
following case (sub can be "add" and "mul", "nsw + sext" can be "nus + zext")
```
Src:
%c = sub nsw %b, %indvar
%d = sext %c to i64
Dst:
%indvar.ext1 = sext %indvar to i64
%m = sext %b to i64
%d = sub nsw i64 %m, %indvar.ext1
```
Therefore, as long as the result of add/sub/mul is extended to wide type with
right extension and overflow wrap combination, no
trunc is required regardless of how %b is generated. This pattern is common
when calculating address in 64 bit architecture.
Note that this patch reuse almost all the code from D49151 by @az:
https://reviews.llvm.org/D49151
It extends it by providing proof of why trunc is unnecessary in more general case,
it should also resolve some of the concerns from the following discussion with @reames.
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20180910/585945.html
Reviewers: sanjoy, efriedma, sebpop, reames, az, javed.absar, amehsan
Reviewed By: az, amehsan
Subscribers: hiraditya, llvm-commits, amehsan, reames, az
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73059
Summary:
Replacing uses of IV outside of the loop is likely generally useful,
but `rewriteLoopExitValues()` is cautious, and if it isn't told to always
perform the replacement, and there are hard uses of IV in loop,
it doesn't replace.
In [[ https://bugs.llvm.org/show_bug.cgi?id=44668 | PR44668 ]],
that prevents `-indvars` from replacing uses of induction variable
after the loop, which might be one of the optimization failures
preventing that code from being vectorized.
Instead, now that the cost model is fixed, i believe we should be
a little bit more optimistic, and also perform replacement
if we believe it is within our budget.
Fixes [[ https://bugs.llvm.org/show_bug.cgi?id=44668 | PR44668 ]].
Reviewers: reames, mkazantsev, asbirlea, fhahn, skatkov
Reviewed By: mkazantsev
Subscribers: nikic, hiraditya, zzheng, javed.absar, dmgreen, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73501
Summary:
Previosly we simply always said that `SCEVMinMaxExpr` is too costly to expand.
But this isn't really true, it expands into just a comparison+swap pair.
And again much like with add/mul, there will be one less such pair
than the number of operands. And we need to count the cost of operands themselves.
This does change a number of testcases, and as far as i can tell,
all of these changes are improvements, in the sense that
we fixed up more latches to do the [in]equality comparison.
This concludes cost-modelling changes, no other SCEV expressions exist as of now.
This is a part of addressing [[ https://bugs.llvm.org/show_bug.cgi?id=44668 | PR44668 ]].
Reviewers: reames, mkazantsev, wmi, sanjoy
Reviewed By: mkazantsev
Subscribers: hiraditya, javed.absar, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73744
Summary:
While this resolves the regression from D73722 in `llvm/test/Transforms/IndVarSimplify/exit_value_test2.ll`,
this now regresses `llvm/test/Transforms/IndVarSimplify/elim-extend.ll` `@nestedIV` test,
we no longer can perform that expansion within default budget of `4`, but require budget of `6`.
That regression is being addressed by D73777.
The basic idea here is simple.
```
Op0, Op1, Op2 ...
| | |
\--+--/ |
| |
\---+---/
```
I.e. given N operands, we will have N-1 operations,
so we have to add cost of an add (mul) for **every** Op processed,
**except** the first one, plus we need to recurse into *every* Op.
I'm guessing there's already canonicalization that ensures we won't
have `1` operand in `scMulExpr`, and no `0` in `scAddExpr`/`scMulExpr`.
Reviewers: reames, mkazantsev, wmi, sanjoy
Reviewed By: mkazantsev
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73728
Summary:
If we don't believe this UDiv is actually a LShr in disguise, things are much worse.
First, we try to see if this UDiv actually originates from user code,
by looking for `S + 1`, and if found considering this UDiv to be free.
But otherwise, we always considered this UDiv to be high-cost.
However that is no longer the case with TTI-driven cost model:
our default budget is 4, which matches the default cost of UDiv,
so now we allow a single UDiv to not be counted as high-cost.
While that is the case, it is evident this is actually a regression
due to the fact that cost-modelling is incomplete - we did not account
for the `add`, `mul` costs yet. That is being addressed in D73728.
Cost-modelling for UDiv also seems pretty straight-forward:
subtract cost of the UDiv itself, and recurse into both the LHS and RHS.
Reviewers: reames, mkazantsev, wmi, sanjoy
Reviewed By: mkazantsev
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73722
In builds with assertions enabled (!NDEBUG), IndVarSimplify does an
additional query to ScalarEvolution which may change future SCEV queries
since it fills the internal cache differently. The result is actually
only used with the -verify-indvars command line option. We fix the issue
by only calling SE->getBackedgeTakenCount(L) if -verify-indvars is
enabled such that only -verify-indvars shows the behavior, but not debug
builds themselves. Also add a remark to the description of
-verify-indvars about this behavior.
Fixes llvm.org/PR44815
Differential Revision: https://reviews.llvm.org/D74810
Teach SCEV about the @loop.decrement.reg intrinsic, which has exactly the same
semantics as a sub expression. This allows us to query hardware-loops, which
contain this @loop.decrement.reg intrinsic, so that we can calculate iteration
counts, exit values, etc. of hardwareloops.
This "int_loop_decrement_reg" intrinsic is defined as "IntrNoDuplicate". Thus,
while hardware-loops and tripcounts now become analysable by SCEV, this
prevents the usual loop transformations from applying transformations on
hardware-loops, which is what we want at this point, for which I have added
test cases for loopunrolling and IndVarSimplify and LFTR.
Differential Revision: https://reviews.llvm.org/D71563
The basic idea of the transform is to convert variant loop exit conditions into invariant exit conditions by changing the iteration on which the exit is taken when we know that the trip count is unobservable. See the original patch which introduced the code for a more complete explanation.
The individual parts of this have been reviewed, the result has been fuzzed, and then further analyzed by hand, but despite all of that, I will not be suprised to see breakage here. If you see problems, please don't hesitate to revert - though please do provide a test case. The most likely class of issues are latent SCEV bugs and without a reduced test case, I'll be essentially stuck on reducing them.
(Note: A bunch of tests were opted out of the new transform to preserve coverage. That landed in a previous commit to simplify revert cycles if they turn out to be needed.)
I'm about to enable the new loop predication transform by default. It has the effect of completely destroying many read only loops - which happen to be a super common idiom in our test cases. So as to preserve test coverage of other transforms, disable the new transform where it would cause sharp test coverage regressions.
(This is semantically part of the enabling commit. It's committed separate to ease revert if the actual flag flip gets reverted.)
We can end up with two loop exits whose exit counts are equivalent, but whose textual representation is different and non-obvious. For the sub-case where we have a series of exits which dominate one another (common), eliminate any exits which would iterate *after* a previous exit on the exiting iteration.
As noted in the TODO being removed, I'd always thought this was a good idea, but I've now seen this in a real workload as well.
Interestingly, in review, Nikita pointed out there's let another oppurtunity to leverage SCEV's reasoning. If we kept track of the min of dominanting exits so far, we could discharge exits with EC >= MDE. This is less powerful than the existing transform (since later exits aren't considered), but potentially more powerful for any case where SCEV can prove a >= b, but neither a == b or a > b. I don't have an example to illustrate that oppurtunity, but won't be suprised if we find one and return to handle that case as well.
Differential Revision: https://reviews.llvm.org/D69009
llvm-svn: 375379
The problem is that we can have two loop exits, 'a' and 'b', where 'a' and 'b' would exit at the same iteration, 'a' precedes 'b' along some path, and 'b' is predicated while 'a' is not. In this case (see the previously submitted test case), we causing the loop to exit through 'b' whereas it should have exited through 'a'.
This only applies to loop exits where the exit counts are not provably inequal, but that isn't as much of a restriction as it appears. If we could order the exit counts, we'd have already removed one of the two exits. In theory, we might be able to prove inequality w/o ordering, but I didn't really explore that piece. Instead, I went for the obvious restriction and ensured we didn't predicate exits following non-predicateable exits.
Credit goes to Evgeny Brevnov for figuring out the problematic case. Fuzzing probably also found it (failures seen), but due to some silly infrastructure problems I hadn't gotten to the results before Evgeny hand reduced it from a benchmark (he manually enabled the transform). Once this is fixed, I'll try to filter through the fuzzer failures to see if there's anything additional lurking.
Differential Revision https://reviews.llvm.org/D68956
llvm-svn: 375038
Credit goes to Evgeny Brevnov for figuring out the problematic case.
Fuzzing probably also found it (lots of failures), but due to some silly infrastructure problems I hadn't gotten to the results before Evgeny hand reduced it from a benchmark.
llvm-svn: 374812
This patch implements a variation of a well known techniques for JIT compilers - we have an implementation in tree as LoopPredication - but with an interesting twist. This version does not assume the ability to execute a path which wasn't taken in the original program (such as a guard or widenable.condition intrinsic). The benefit is that this works for arbitrary IR from any frontend (including C/C++/Fortran). The tradeoff is that it's restricted to read only loops without implicit exits.
This builds on SCEV, and can thus eliminate the loop varying portion of the any early exit where all exits are understandable by SCEV. A key advantage is that fixing deficiency exposed in SCEV - already found one while writing test cases - will also benefit all of full redundancy elimination (and most other loop transforms).
I haven't seen anything in the literature which quite matches this. Given that, I'm not entirely sure that keeping the name "loop predication" is helpful. Anyone have suggestions for a better name? This is analogous to partial redundancy elimination - since we remove the condition flowing around the backedge - and has some parallels to our existing transforms which try to make conditions invariant in loops.
Factoring wise, I chose to put this in IndVarSimplify since it's a generally applicable to all workloads. I could split this off into it's own pass, but we'd then probably want to add that new pass every place we use IndVars. One solid argument for splitting it off into it's own pass is that this transform is "too good". It breaks a huge number of existing IndVars test cases as they tend to be simple read only loops. At the moment, I've opted it off by default, but if we add this to IndVars and enable, we'll have to update around 20 test files to add side effects or disable this transform.
Near term plan is to fuzz this extensively while off by default, reflect and discuss on the factoring issue mentioned just above, and then enable by default. I also need to give some though to supporting widenable conditions in this framing.
Differential Revision: https://reviews.llvm.org/D67408
llvm-svn: 373351
Summary:
Previously, if the threshold was 2, we were willing to speculatively
execute 2 cheap instructions in both basic blocks (thus we were willing
to speculatively execute cost = 4), but weren't willing to speculate
when one BB had 3 instructions and other one had no instructions,
even thought that would have total cost of 3.
This looks inconsistent to me.
I don't think `cmov`-like instructions will start executing
until both of it's inputs are available: https://godbolt.org/z/zgHePf
So i don't see why the existing behavior is the correct one.
Also, let's add it's own `cl::opt` for this threshold,
with default=4, so it is not stricter than the previous threshold:
will allow to fold when there are 2 BB's each with cost=2.
And since the logic has changed, it will also allow to fold when
one BB has cost=3 and other cost=1, or there is only one BB with cost=4.
This is an alternative solution to D65148:
This fix is mainly motivated by `signbit-like-value-extension.ll` test.
That pattern comes up in JPEG decoding, see e.g.
`Figure F.12 – Extending the sign bit of a decoded value in V`
of `ITU T.81` (JPEG specification).
That branch is not predictable, and it is within the innermost loop,
so the fact that that pattern ends up being stuck with a branch
instead of `select` (i.e. `CMOV` for x86) is unlikely to be beneficial.
This has great results on the final assembly (vanilla test-suite + RawSpeed): (metric pass - D67240)
| metric | old | new | delta | % |
| x86-mi-counting.NumMachineFunctions | 37720 | 37721 | 1 | 0.00% |
| x86-mi-counting.NumMachineBasicBlocks | 773545 | 771181 | -2364 | -0.31% |
| x86-mi-counting.NumMachineInstructions | 7488843 | 7486442 | -2401 | -0.03% |
| x86-mi-counting.NumUncondBR | 135770 | 135543 | -227 | -0.17% |
| x86-mi-counting.NumCondBR | 423753 | 422187 | -1566 | -0.37% |
| x86-mi-counting.NumCMOV | 24815 | 25731 | 916 | 3.69% |
| x86-mi-counting.NumVecBlend | 17 | 17 | 0 | 0.00% |
We significantly decrease basic block count, notably decrease instruction count,
significantly decrease branch count and very significantly increase `cmov` count.
Performance-wise, unsurprisingly, this has great effect on
target RawSpeed benchmark. I'm seeing 5 **major** improvements:
```
Benchmark Time CPU Time Old Time New CPU Old CPU New
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Samsung/NX3000/_3184416.SRW/threads:8/process_time/real_time_pvalue 0.0000 0.0000 U Test, Repetitions: 49 vs 49
Samsung/NX3000/_3184416.SRW/threads:8/process_time/real_time_mean -0.3064 -0.3064 226.9913 157.4452 226.9800 157.4384
Samsung/NX3000/_3184416.SRW/threads:8/process_time/real_time_median -0.3057 -0.3057 226.8407 157.4926 226.8282 157.4828
Samsung/NX3000/_3184416.SRW/threads:8/process_time/real_time_stddev -0.4985 -0.4954 0.3051 0.1530 0.3040 0.1534
Kodak/DCS760C/86L57188.DCR/threads:8/process_time/real_time_pvalue 0.0000 0.0000 U Test, Repetitions: 49 vs 49
Kodak/DCS760C/86L57188.DCR/threads:8/process_time/real_time_mean -0.1747 -0.1747 80.4787 66.4227 80.4771 66.4146
Kodak/DCS760C/86L57188.DCR/threads:8/process_time/real_time_median -0.1742 -0.1743 80.4686 66.4542 80.4690 66.4436
Kodak/DCS760C/86L57188.DCR/threads:8/process_time/real_time_stddev +0.6089 +0.5797 0.0670 0.1078 0.0673 0.1062
Sony/DSLR-A230/DSC08026.ARW/threads:8/process_time/real_time_pvalue 0.0000 0.0000 U Test, Repetitions: 49 vs 49
Sony/DSLR-A230/DSC08026.ARW/threads:8/process_time/real_time_mean -0.1598 -0.1598 171.6996 144.2575 171.6915 144.2538
Sony/DSLR-A230/DSC08026.ARW/threads:8/process_time/real_time_median -0.1598 -0.1597 171.7109 144.2755 171.7018 144.2766
Sony/DSLR-A230/DSC08026.ARW/threads:8/process_time/real_time_stddev +0.4024 +0.3850 0.0847 0.1187 0.0848 0.1175
Canon/EOS 77D/IMG_4049.CR2/threads:8/process_time/real_time_pvalue 0.0000 0.0000 U Test, Repetitions: 49 vs 49
Canon/EOS 77D/IMG_4049.CR2/threads:8/process_time/real_time_mean -0.0550 -0.0551 280.3046 264.8800 280.3017 264.8559
Canon/EOS 77D/IMG_4049.CR2/threads:8/process_time/real_time_median -0.0554 -0.0554 280.2628 264.7360 280.2574 264.7297
Canon/EOS 77D/IMG_4049.CR2/threads:8/process_time/real_time_stddev +0.7005 +0.7041 0.2779 0.4725 0.2775 0.4729
Canon/EOS 5DS/2K4A9929.CR2/threads:8/process_time/real_time_pvalue 0.0000 0.0000 U Test, Repetitions: 49 vs 49
Canon/EOS 5DS/2K4A9929.CR2/threads:8/process_time/real_time_mean -0.0354 -0.0355 316.7396 305.5208 316.7342 305.4890
Canon/EOS 5DS/2K4A9929.CR2/threads:8/process_time/real_time_median -0.0354 -0.0356 316.6969 305.4798 316.6917 305.4324
Canon/EOS 5DS/2K4A9929.CR2/threads:8/process_time/real_time_stddev +0.0493 +0.0330 0.3562 0.3737 0.3563 0.3681
```
That being said, it's always best-effort, so there will likely
be cases where this worsens things.
Reviewers: efriedma, craig.topper, dmgreen, jmolloy, fhahn, Carrot, hfinkel, chandlerc
Reviewed By: jmolloy
Subscribers: xbolva00, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67318
llvm-svn: 372009
We were computing the loop exit value, but not ensuring the addrec belonged to the loop whose exit value we were computing. I couldn't actually trip this; the test case shows the basic setup which *might* trip this, but none of the variations I've tried actually do.
llvm-svn: 369730
We already supported rewriting loop exit values for multiple exit loops, but if any of the loop exits were not computable, we gave up on all loop exit values. This patch generalizes the existing code to handle individual computable loop exits where possible.
As discussed in the review, this is a starting point for figuring out a better API. The code is a bit ugly, but getting it in lets us test as we go.
Differential Revision: https://reviews.llvm.org/D65544
llvm-svn: 368898
This is a prepatory patch for future work on support exit value rewriting in loops with a mixture of computable and non-computable exit counts. The intention is to be "mostly NFC" - i.e. not enable any interesting new transforms - but in practice, there are some small output changes.
The test differences are caused by cases wherewhere getSCEVAtScope can simplify a single entry phi without needing any knowledge of the loop.
llvm-svn: 367485
The original code failed to account for the fact that one exit can have a pointer exit count without all of them having pointer exit counts. This could cause two separate bugs:
1) We might exit the loop early, and leave optimizations undone. This is what triggered the assertion failure in the reported test case.
2) We might optimize one exit, then exit without indicating a change. This could result in an analysis invalidaton bug if no other transform is done by the rest of indvars.
Note that the pointer exit counts are a really fragile concept. They show up only when we have a pointer IV w/o a datalayout to provide their size. It's really questionable to me whether the complexity implied is worth it.
llvm-svn: 366829
Continue in the spirit of D63618, and use exit count reasoning to prove away loop exits which can not be taken since the backedge taken count of the loop as a whole is provably less than the minimal BE count required to take this particular loop exit.
As demonstrated in the newly added tests, this triggers in a number of cases where IndVars was previously unable to discharge obviously redundant exit tests. And some not so obvious ones.
Differential Revision: https://reviews.llvm.org/D63733
llvm-svn: 365920
As noted in the test change, this is not trivially NFC, but all of the changes in output are cases where the SCEVExpander form is more canonical/optimal than the hand generation.
llvm-svn: 365075
The motivation for this is two fold:
1) Make the output (and thus tests) a bit more readable to a human trying to understand the result of the transform
2) Reduce spurious diffs in a potential future change to restructure all of this logic to use SCEVExpander (which hoists by default)
llvm-svn: 365066
Fixes https://bugs.llvm.org/show_bug.cgi?id=41998. Usually when we
have a truncated exit count we'll truncate the IV when comparing
against the limit, in which case exit count overflow in post-inc
form doesn't matter. However, for pointer IVs we don't do that, so
we have to be careful about incrementing the IV in the wide type.
I'm fixing this by removing the IVCount variable (which was
ExitCount or ExitCount+1) and replacing it with a UsePostInc flag,
and then moving the actual limit adjustment to the individual cases
(which are: pointer IV where we add to the wide type, integer IV
where we add to the narrow type, and constant integer IV where we
add to the wide type).
Differential Revision: https://reviews.llvm.org/D63686
llvm-svn: 364709
In rL364135, I taught IndVars to fold exiting branches in loops with a zero backedge taken count (i.e. loops that only run one iteration). This extends that to eliminate the dead comparison left around.
llvm-svn: 364155
This turned out to be surprisingly effective. I was originally doing this just for completeness sake, but it seems like there are a lot of cases where SCEV's exit count reasoning is stronger than it's isKnownPredicate reasoning.
Once this is in, I'm thinking about trying to build on the same infrastructure to eliminate provably untaken checks. There may be something generally interesting here.
Differential Revision: https://reviews.llvm.org/D63618
llvm-svn: 364135
Thought of this case while working on something else. We appear to get it right in all of the variations I tried, but that's by accident. So, add a test which would catch the potential bug.
llvm-svn: 363953
Teach IndVarSimply's LinearFunctionTestReplace transform to handle multiple exit loops. LFTR does two key things 1) it rewrites (all) exit tests in terms of a common IV potentially eliminating one in the process and 2) it moves any offset/indexing/f(i) style logic out of the loop.
This turns out to actually be pretty easy to implement. SCEV already has all the information we need to know what the backedge taken count is for each individual exit. (We use that when computing the BE taken count for the loop as a whole.) We basically just need to iterate through the exiting blocks and apply the existing logic with the exit specific BE taken count. (The previously landed NFC makes this super obvious.)
I chose to go ahead and apply this to all loop exits instead of only latch exits as originally proposed. After reviewing other passes, the only case I could find where LFTR form was harmful was LoopPredication. I've fixed the latch case, and guards aren't LFTRed anyways. We'll have some more work to do on the way towards widenable_conditions, but that's easily deferred.
I do want to note that I added one bit after the review. When running tests, I saw a new failure (no idea why didn't see previously) which pointed out LFTR can rewrite a constant condition back to a loop varying one. This was theoretically possible with a single exit, but the zero case covered it in practice. With multiple exits, we saw this happening in practice for the eliminate-comparison.ll test case because we'd compute a ExitCount for one of the exits which was guaranteed to never actually be reached. Since LFTR ran after simplifyAndExtend, we'd immediately turn around and undo the simplication work we'd just done. The solution seemed obvious, so I didn't bother with another round of review.
Differential Revision: https://reviews.llvm.org/D62625
llvm-svn: 363883
This patch really contains two pieces:
Teach SCEV how to fold a phi in the header of a loop to the value on the backedge when a) the backedge is known to execute at least once, and b) the value is safe to use globally within the scope dominated by the original phi.
Teach IndVarSimplify's rewriteLoopExitValues to allow loop invariant expressions which already exist (and thus don't need new computation inserted) even in loops where we can't optimize away other uses.
Differential Revision: https://reviews.llvm.org/D63224
llvm-svn: 363619
Recommit r363289 with a bug fix for crash identified in pr42279. Issue was that a loop exit test does not have to be an icmp, leading to a null dereference crash when new logic was exercised for that case. Test case previously committed in r363601.
Original commit comment follows:
This contains fixes for two cases where we might invalidate inbounds and leave it stale in the IR (a miscompile). Case 1 is when switching to an IV with no dynamically live uses, and case 2 is when doing pre-to-post conversion on the same pointer type IV.
The basic scheme used is to prove that using the given IV (pre or post increment forms) would have to already trigger UB on the path to the test we're modifying. As such, our potential UB triggering use does not change the semantics of the original program.
As was pointed out in the review thread by Nikita, this is defending against a separate issue from the hasConcreteDef case. This is about poison, that's about undef. Unfortunately, the two are different, see Nikita's comment for a fuller explanation, he explains it well.
(Note: I'm going to address Nikita's last style comment in a separate commit just to minimize chance of subtle bugs being introduced due to typos.)
Differential Revision: https://reviews.llvm.org/D62939
llvm-svn: 363613
If we can detect that saturating math that depends on an IV cannot
overflow, replace it with simple math. This is similar to the CVP
optimization from D62703, just based on a different underlying
analysis (SCEV vs LVI) that catches different cases.
Differential Revision: https://reviews.llvm.org/D62792
llvm-svn: 363489
Roman Lebedev