As it's causing some bot failures (and per request from kbarton).
This reverts commit r358543/ab70da07286e618016e78247e4a24fcb84077fda.
llvm-svn: 358546
Summary:
Between building the pair map and querying it there are a few places that
erase and create Values. It's rare but the address of these newly created
Values is occasionally the same as a just-erased Value that we already
have in the pair map. These coincidences should be accounted for to avoid
non-determinism.
Thanks to Roman Tereshin for the test case.
Reviewers: rtereshin, bogner
Reviewed By: rtereshin
Subscribers: mgrang, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59401
llvm-svn: 356803
This patch fixes the issue noticed in D54532.
The problem is that cst_pred_ty-based matchers like m_Zero() currently do not match
scalar undefs (as expected), but *do* match vector undefs. This may lead to optimization
inconsistencies in rare cases.
There is only one existing test for which output changes, reverting the change from D53205.
The reason here is that vector fsub undef, %x is no longer matched as an m_FNeg(). While I
think that the new output is technically worse than the previous one, it is consistent with
scalar, and I don't think it's really important either way (generally that undef should have
been folded away prior to reassociation.)
I've also added another test case for this issue based on InstructionSimplify. It took some
effort to find that one, as in most cases undef folds are either checked first -- and in the
cases where they aren't it usually happens to not make a difference in the end. This is the
only case I was able to come up with. Prior to this patch the test case simplified to undef
in the scalar case, but zeroinitializer in the vector case.
Patch by: @nikic (Nikita Popov)
Differential Revision: https://reviews.llvm.org/D54631
llvm-svn: 347318
This work is to avoid regressions when we seperate FNeg from the FSub IR instruction.
Differential Revision: https://reviews.llvm.org/D53205
llvm-svn: 345146
We need to update this code before introducing an 'fneg' instruction in IR,
so we might as well kill off the integer neg/not queries too.
This is no-functional-change-intended for scalar code and most vector code.
For vectors, we can see that the 'match' API allows for undef elements in
constants, so we optimize those cases better.
Ideally, there would be a test for each code diff, but I don't see evidence
of that for the existing code, so I didn't try very hard to come up with new
vector tests for each code change.
Differential Revision: https://reviews.llvm.org/D53533
llvm-svn: 345042
I was trying to provide test coverage for D53533
with rL344964, but these don't do it...and I don't
think they add any value, so deleting.
llvm-svn: 344969
Also, regenerate checks for these files. We should do better
on the vector tests by using the PatternMatch API instead of
BinaryOperator::isNot/isNeg.
llvm-svn: 344964
If we have a pair of binops feeding another pair of binops, rearrange the operands so
the matching pair are together because that allows easy factorization folds to happen
in instcombine:
((X << S) & Y) & (Z << S) --> ((X << S) & (Z << S)) & Y (reassociation)
--> ((X & Z) << S) & Y (factorize shift from 'and' ops optimization)
This is part of solving PR37098:
https://bugs.llvm.org/show_bug.cgi?id=37098
Note that there's an instcombine version of this patch attached there, but we're trying
to make instcombine have less responsibility to improve compile-time efficiency.
For reasons I still don't completely understand, reassociate does this kind of transform
sometimes, but misses everything in my motivating cases.
This patch on its own is gluing an independent cleanup chunk to the end of the existing
RewriteExprTree() loop. We can build on it and do something stronger to better order the
full expression tree like D40049. That might be an alternative to the proposal to add a
separate reassociation pass like D41574.
Differential Revision: https://reviews.llvm.org/D45842
llvm-svn: 341288
As the test diffs show, the current users of getBinOpIdentity()
are InstCombine and Reassociate. SLP vectorizer is a candidate
for using this functionality too (D28907).
The InstCombine shuffle improvements are part of the planned
enhancements noted in D48830.
InstCombine actually has several other uses of getBinOpIdentity()
via SimplifyUsingDistributiveLaws(), but we don't call that for
any FP ops. Fixing that might be another part of removing the
custom reassociation in InstCombine that is only done for fadd+fmul.
llvm-svn: 336215
Phi nodes can reside in live blocks but one of their incoming
arguments can come from a dead block. Dead blocks and reassociate
don't play nice together. In fact, reassociate performs an RPO
as a first step to avoid processing dead blocks.
The reason why Reassociate might not fixpoint when examining
dead blocks is that the following:
%xor0 = xor i16 %xor1, undef
%xor1 = xor i16 %xor0, undef
is perfectly valid LLVM IR (if it appears in a dead block),
so the worklist algorithm keeps pushing the two instructions for
reexamination. Note that this is not Reassociate fault, at least
not entirely. It's llvm that has a weird definition of dominance.
Fixes PR37390.
llvm-svn: 332100
In order to set breakpoints on labels and list source code around
labels, we need collect debug information for labels, i.e., label
name, the function label belong, line number in the file, and the
address label located. In order to keep these information in LLVM
IR and to allow backend to generate debug information correctly.
We create a new kind of metadata for labels, DILabel. The format
of DILabel is
!DILabel(scope: !1, name: "foo", file: !2, line: 3)
We hope to keep debug information as much as possible even the
code is optimized. So, we create a new kind of intrinsic for label
metadata to avoid the metadata is eliminated with basic block.
The intrinsic will keep existing if we keep it from optimized out.
The format of the intrinsic is
llvm.dbg.label(metadata !1)
It has only one argument, that is the DILabel metadata. The
intrinsic will follow the label immediately. Backend could get the
label metadata through the intrinsic's parameter.
We also create DIBuilder API for labels to be used by Frontend.
Frontend could use createLabel() to allocate DILabel objects, and use
insertLabel() to insert llvm.dbg.label intrinsic in LLVM IR.
Differential Revision: https://reviews.llvm.org/D45024
Patch by Hsiangkai Wang.
llvm-svn: 331841
Summary:
Some of our internal testing detected a major compile time regression which I've
tracked down to:
r278938 - Revert "Reassociate: Reprocess RedoInsts after each inst".
It appears that processing long chains of reassociatable instructions causes
non-linear (potentially exponential) growth in the number of times an
instruction is revisited. For example, the included test revisits instructions
220 times in a 20-instruction test.
It appears that r278938 reversed the order instructions were visited and that
this is preventing scheduled revisits from being cancelled as a result of
visiting the instructions naturally during normal processing. However, simply
reversing the order also harmed the generated code. Upon closer inspection, it
was discovered that revisits occurred in the opposite order to the first pass
(Thanks to escha for spotting that).
This patch makes the revisit order consistent with the first pass which allows
more revisits to be cancelled. This does appear to have a small impact on the
generated code in few cases but it significantly reduces compile-time.
After this patch, our internal test that was most affected by the regression
dropped from ~2 million revisits to ~4k resulting in Reassociate having 0.46%
of the runtime it had before (99.54% improvement).
Here's the summaries reported by lnt for the LLVM test-suite with --benchmarking-only:
| metric | geomean before patch | geomean after patch | delta |
| ----- | ----- | ----- | ----- |
| compile time | 0.1956 | 0.1261 | -35.54% |
| execution time | 0.3240 | 0.3237 | - |
| code size | 7365.4459 | 7365.6079 | - |
The results have a few wins and losses on compile-time, mostly in the +/- 2.5% range. There was one outlier though:
| Performance Regressions - compile_time | Δ | Previous | Current |
| MultiSource/Benchmarks/ASC_Sequoia/CrystalMk/CrystalMk | 9.82% | 2.0473 | 2.2483 |
Reviewers: javed.absar, dberlin
Reviewed By: dberlin
Subscribers: kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D45734
llvm-svn: 331381
As suggested in D45842
(although still not sure if we're going to advance that),
we must invalidate references to instructions that have
been recycled (operands were changed, so result is different).
llvm-svn: 331083
Summary:
When Reassociate is rewriting an expression tree it may
reuse old binary expression nodes, for new expressions.
Whenever an expression node is reused, but with a non-trivial
change in the result, we need to invalidate any debug info
that is associated with the node.
If for example rewriting
x = mul a, b
y = mul c, x
into
x = mul c, b
y = mul a, x
we still get the same result for 'y', but 'x' is a new expression.
All debug info referring to 'x' must be invalidated (marked as
optimized out) since we no longer calculate the expected value.
As a side-effect this patch avoid (at least some) problems where
reassociate could end up creating IR with debug-use before def.
Earlier the dbg.value nodes where left untouched in the IR, while
the reused binary nodes where sinked to just before the root node
of the rewritten expression tree. See PR27273 for more info about
such problems.
Reviewers: dblaikie, aprantl, dexonsmith
Reviewed By: aprantl
Subscribers: JDevlieghere, llvm-commits
Tags: #debug-info
Differential Revision: https://reviews.llvm.org/D45975
llvm-svn: 330804
Similar to rL330086, I don't know if we want to do these
transforms here, but we might as well have the tests
here either way to show that this pass is missing
potential functionality (intentionally or not).
llvm-svn: 330368
These simplifications were previously enabled only with isFast(), but that
is more restrictive than required. Since r317488, FMF has 'reassoc' to
control these cases at a finer level.
llvm-svn: 330089
These are uncontroversial and independent of a proposed LangRef edits (D44216).
I tried to fix tests that would fold away:
rL327004
rL327028
rL327030
rL327034
I'm not sure if the Reassociate tests are meaningless yet, but they probably will be
as we add more folds, so if anyone has suggestions or wants to fix those, please do.
Differential Revision: https://reviews.llvm.org/D44258
llvm-svn: 327058
This is a retry of r326502 with updates to the reassociate
test file that I missed the first time.
@test15_reassoc in the supposed -reassociate test file
(except that it tests 2 other passes too...) shows that
there's no clear responsiblity for reassociation transforms.
Instcombine now gets that case, but only because the
constant values are identical. Otherwise, it would still
miss that pattern.
Reassociate doesn't get that case because it hasn't been
updated to use less than 'fast' FMF.
llvm-svn: 326513
This algorithm (explained more in the source code) takes into account
global redundancies by building a "pair map" to find common subexprs.
The primary motivation of this is to handle situations like
foo = (a * b) * c
bar = (a * d) * c
where we currently don't identify that "a * c" is redundant.
Accordingly, it prioritizes the emission of a * c so that CSE
can remove the redundant calculation later.
Does not change the actual reassociation algorithm -- only the
order in which the reassociated operand chain is reconstructed.
Gives ~1.5% floating point math instruction count reduction on
a large offline suite of graphics shaders.
llvm-svn: 320515
The toxic stew of created values named 'tmp' and tests that already have
values named 'tmp' and CHECK lines looking for values named 'tmp' causes
bad things to happen in our test line auto-generation scripts because it
wants to use 'TMP' as a prefix for unnamed values. Use less 'tmp' to
avoid that.
llvm-svn: 317818
The reassociate pass generates named values such as "%tmp2" which trips up the script's regex's
because the script uses a 'TMP' prefix for unnamed values (%2).
llvm-svn: 317810
Summary: The loop invariant add (i+j) is reassoicated, I think the FIXME can be removed, because this is what the test case tries to check (AFAIK). I also changed the test to use FileCheck.
Reviewers: mcrosier, davide
Reviewed By: mcrosier, davide
Subscribers: davide, llvm-commits
Differential Revision: https://reviews.llvm.org/D39424
llvm-svn: 317000
Summary:
When reassociating an expression, do not drop the instruction's
original debug location in case the replacement location is
missing.
The debug location must at least not be dropped for inlinable
callsites of debug-info-bearing functions in debug-info-bearing
functions. Failing to do so would result in an "inlinable function "
"call in a function with debug info must have a !dbg location"
error in the verifier.
As preserving the original debug location is not expected
to result in overly jumpy debug line information, it is
preserved for all other cases too.
This fixes PR34231:
https://bugs.llvm.org/show_bug.cgi?id=34231
Original patch by David Stenberg
Reviewers: davide, craig.topper, mcrosier, dblaikie, aprantl
Reviewed By: davide, aprantl
Subscribers: aprantl
Differential Revision: https://reviews.llvm.org/D36865
llvm-svn: 311642
..if the resulting subtract will be broken up later. This can cause us to get
into an infinite loop.
x + (-5.0 * y) -> x - (5.0 * y) ; Canonicalize neg const
x - (5.0 * y) -> x + (0 - (5.0 * y)) ; Break up subtract
x + (0 - (5.0 * y)) -> x + (-5.0 * y) ; Replace 0-X with X*-1.
PR34078
llvm-svn: 311554
Summary:
EraseInst didn't report that it made IR changes through MadeChange.
It is essential that changes to the IR are reported correctly,
since for example ReassociatePass::run() will indicate that all
analyses are preserved otherwise.
And the CGPassManager determines if the CallGraph is up-to-date
based on status from InstructionCombiningPass::runOnFunction().
Reviewers: craig.topper, rnk, davide
Reviewed By: rnk, davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D34616
llvm-svn: 306368
Summary:
Currently we don't try to do anything with vector xors.
This patch adds support for removing duplicate pairs from a chain of vector xors as its pretty easy to support. We still dont' try to combine the xors with and/ors, but I might try that in a future patch.
Reviewers: mcrosier, davide, resistor
Reviewed By: mcrosier
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D34338
llvm-svn: 305704
In OptimizeAdd, we scan the operand list to see if there are any common factors
between operands that can be factored out to reduce the number of multiplies
(e.g., 'A*A+A*B*C+D' -> 'A*(A+B*C)+D'). For each operand of the operand list, we
only consider unique factors (which is tracked by the Duplicate set). Now if we
find a factor that is a negative constant, we add the negated value as a factor
as well, because we can percolate the negate out. However, we mistakenly don't
add this negated constant to the Duplicates set.
Consider the expression A*2*-2 + B. Obviously, nothing to factor.
For the added value A*2*-2 we over count 2 as a factor without this change,
which causes the assert reported in PR30256. The problem is that this code is
assuming that all the multiply operands of the add are already reassociated.
This change avoids the issue by making OptimizeAdd tolerate multiplies which
haven't been completely optimized; this sort of works, but we're doing wasted
work: we'll end up revisiting the add later anyway.
Another possible approach would be to enforce RPO iteration order more strongly.
If we have RedoInsts, we process them immediately in RPO order, rather than
waiting until we've finished processing the whole function. Intuitively, it
seems like the natural approach: reassociation works on expression trees, so
the optimization only works in one direction. That said, I'm not sure how
practical that is given the current Reassociate; the "optimal" form for an
expression depends on its use list (see all the uses of "user_back()"), so
Reassociate is really an iterative optimization of sorts, so any changes here
would probably get messy.
PR30256
Differential Revision: https://reviews.llvm.org/D30228
llvm-svn: 296003
Summary:
It was detected that the reassociate pass could enter an inifite
loop when analysing dead code. Simply skipping to analyse basic
blocks that are dead avoids such problems (and as a side effect
we avoid spending time on optimising dead code).
The solution is using the same Reverse Post Order ordering of the
basic blocks when doing the optimisations, as when building the
precalculated rank map. A nice side-effect of this solution is
that we now know that we only try to do optimisations for blocks
with ranked instructions.
Fixes https://llvm.org/bugs/show_bug.cgi?id=30818
Reviewers: llvm-commits, davide, eli.friedman, mehdi_amini
Subscribers: dberlin
Differential Revision: https://reviews.llvm.org/D26154
llvm-svn: 285793
Fixes PR 30784. Discussed with Justin, who pointed out that
in the new PassManager infrastructure we can have more fine-grained
control on which analyses we want to preserve, but this is the
best we can do with the current infrastructure.
llvm-svn: 285380
Cameron McInally