is enabled.
We can save memory and reduce binary size significantly by enabling
ProfileSampleAccurate. However when ProfileSampleAccurate is true,
function without sample will be regarded as cold and this could
potentially cause performance regression.
To minimize the potential negative performance impact, we want to be
a little conservative here saying if a function shows up in the profile,
no matter as outline instance, inline instance or call targets, treat
the function as not being cold. This will handle the cases such as most
callsites of a function are inlined in sampled binary (thus outline copy
don't get any sample) but not inlined in current build (because of source
code drift, imprecise debug information, or the callsites are all cold
individually but not cold accumulatively...), so that the outline function
showing up as cold in sampled binary will actually not be cold after current
build. After the change, such function will be treated as not cold even
profile-sample-accurate is enabled.
At the same time we lower the hot criteria of callsiteIsHot check when
profile-sample-accurate is enabled. callsiteIsHot is used to determined
whether a callsite is hot and qualified for early inlining. When
profile-sample-accurate is enabled, functions without profile will be
regarded as cold and much less inlining will happen in CGSCC inlining pass,
so we can worry less about size increase and be aggressive to allow more
early inlining to happen for warm callsites and it is helpful for performance
overall.
Differential Revision: https://reviews.llvm.org/D67561
llvm-svn: 372232
Summary:
The PGO counter reading will add cold and inlinehint (hot) attributes
to functions that are very cold or hot. This was using hardcoded
thresholds, instead of the profile summary cutoffs which are used in
other hot/cold detection and are more dynamic and adaptable. Switch
to using the summary-based cold/hot detection.
The hardcoded limits were causing some code that had a medium level of
hotness (per the summary) to be incorrectly marked with a cold
attribute, blocking inlining.
Reviewers: davidxl
Subscribers: llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67673
llvm-svn: 372189
While we already fold that pattern if the sum of shift amounts is not
smaller than bitwidth, there's painfully obvious generalization:
https://rise4fun.com/Alive/F5R
I.e. the "sub of shift amounts" tells us how many bits will be left
in the output. If it's less than bitwidth, we simply need to
apply a mask, which is constant.
llvm-svn: 372170
Summary:
There were segfaults as we modified and iterated the instruction maps in
the cache at the same time. This was happening because we created new
instructions while we populated the cache. This fix changes the order
in which we perform these actions. First, the caches for the whole
module are created, then we start to create abstract attributes.
I don't have a unit test but the LLVM test suite exposes this problem.
Reviewers: uenoku, sstefan1
Subscribers: hiraditya, bollu, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67232
llvm-svn: 372105
We use `< UP.Threshold` later on, so we should use LoopSize + 1, to
allow unrolling if the result won't exceed to loop size.
Fixes PR43305.
Reviewers: efriedma, dmgreen, paquette
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D67594
llvm-svn: 372084
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
Related folds were added in:
rL125734
...the code comment about register pressure is discussed in
more detail in:
https://bugs.llvm.org/show_bug.cgi?id=2698
But 10 years later, perf testing bzip2 with this change now
shows a slight (0.2% average) improvement on Haswell although
that's probably within test noise.
Given that this is IR canonicalization, we shouldn't be worried
about register pressure though; the backend should be able to
adjust for that as needed.
This is part of solving PR43310 the theoretically right way:
https://bugs.llvm.org/show_bug.cgi?id=43310
...ie, if we don't cripple basic transforms, then we won't
need to add special-case code to detect larger patterns.
rL371940 and rL371981 are related patches in this series.
llvm-svn: 372007
Now that the vectorizer can do tail-folding (rL367592), and the ARM backend
understands MVE masked loads/stores (rL371932), it's time to add the MVE
tail-folding equivalent of the X86 tests that I added.
llvm-svn: 371996
This fold and several others were added in:
rL125734 <https://reviews.llvm.org/rL125734>
...with no explanation for the one-use checks other than the code
comments about register pressure.
Given that this is IR canonicalization, we shouldn't be worried
about register pressure though; the backend should be able to
adjust for that as needed.
This is part of solving PR43310 the theoretically right way:
https://bugs.llvm.org/show_bug.cgi?id=43310
...ie, if we don't cripple basic transforms, then we won't
need to add special-case code to detect larger patterns.
rL371940 is a related patch in this series.
llvm-svn: 371981
D53362 gives a prototype heap-to-stack conversion pass. With addition of new attributes in the attributor, this can now be revisted and improved. This will place it in the Attributor to make it easier to use new attributes (eg. nofree, nosync, willreturn, etc.) and other attributor features.
Reviewers: jdoerfert, uenoku, hfinkel, efriedma
Subscribers: lebedev.ri, xbolva00, hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D65408
llvm-svn: 371942
This fold and several others were added in:
rL125734
...with no explanation for the one-use checks other than the code
comments about register pressure.
Given that this is IR canonicalization, we shouldn't be worried
about register pressure though; the backend should be able to
adjust for that as needed.
There are similar checks as noted with the TODO comments. I'm
hoping to remove those restrictions too, but if any of these
does cause a regression, it should be easier to correct by making
small, individual commits.
This is part of solving PR43310 the theoretically right way:
https://bugs.llvm.org/show_bug.cgi?id=43310
...ie, if we don't cripple basic transforms, then we won't
need to add special-case code to detect larger patterns.
llvm-svn: 371940
Masked loads and store fit naturally with MVE, the instructions being easily
predicated. This adds lowering for the simple cases of masked loads and stores.
It does not yet deal with widening/narrowing or pre/post inc, and so is
currently behind an option.
The llvm masked load intrinsic will accept a "passthru" value, dictating the
values used for the zero masked lanes. In MVE the instructions write 0 to the
zero predicated lanes, so we need to match a passthru that isn't 0 (or undef)
with a select instruction to pull in the correct data after the load.
Differential Revision: https://reviews.llvm.org/D67186
llvm-svn: 371932
This is a fix for:
https://bugs.llvm.org/show_bug.cgi?id=33958
It seems universally true that we would not want to transform this kind of
sequence on any target, but if that's not correct, then we could view this
as a target-specific cost model problem. We could also white-list ConstantInt,
ConstantFP, etc. rather than blacklist Global and ConstantExpr.
Differential Revision: https://reviews.llvm.org/D67362
llvm-svn: 371931
Summary: This should be obsolete once the functionality in D66967 is integrated.
Reviewers: uenoku, sstefan1
Subscribers: hiraditya, bollu, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67231
llvm-svn: 371915
We were failing to compute trip counts (both exact and maximum) for any loop which involved a comparison against either an umin or smin. It looks like this simply got missed when we added smin/umin to SCEV. (Note: umin was submitted separately earlier today. Turned out two folks hit this at the same time.)
Differential Revision: https://reviews.llvm.org/D67514
llvm-svn: 371776
Expanding the folding of `nearbyint()`, `rint()` and `trunc()` to library
functions, in addition to the current support for intrinsics.
Differential revision: https://reviews.llvm.org/D67468
llvm-svn: 371774
Implement a TODO from rL371452, and handle loop invariant addresses in predicated blocks. If we can prove that the load is safe to speculate into the header, then we can avoid using a masked.load in favour of a normal load.
This is mostly about vectorization robustness. In the common case, it's generally expected that LICM/LoadStorePromotion would have eliminated such loads entirely.
Differential Revision: https://reviews.llvm.org/D67372
llvm-svn: 371745
In MVE, as of rL371218, we are attempting to sink chains of instructions such as:
%l1 = insertelement <8 x i8> undef, i8 %l0, i32 0
%broadcast.splat26 = shufflevector <8 x i8> %l1, <8 x i8> undef, <8 x i32> zeroinitializer
In certain situations though, we can end up breaking the dominance relations of
instructions. This happens when we sink the instruction into a loop, but cannot
remove the originals. The Use is updated, which might in fact be a Use from the
second instruction to the first.
This attempts to fix that by reversing the order of instruction that are sunk,
and ensuring that we update the uses on new instructions if they have already
been sunk, not the old ones.
Differential Revision: https://reviews.llvm.org/D67366
llvm-svn: 371743
Folding for fma/fmuladd was added here:
rL202914
...and as seen in existing/unchanged tests, that works to propagate NaN
if it's already an input, but we should fold an fma() that creates NaN too.
From IEEE-754-2008 7.2 "Invalid Operation", there are 2 clauses that apply
to fma, so I added tests for those patterns:
c) fusedMultiplyAdd: fusedMultiplyAdd(0, ∞, c) or fusedMultiplyAdd(∞, 0, c)
unless c is a quiet NaN; if c is a quiet NaN then it is implementation
defined whether the invalid operation exception is signaled
d) addition or subtraction or fusedMultiplyAdd: magnitude subtraction of
infinities, such as: addition(+∞, −∞)
Differential Revision: https://reviews.llvm.org/D67446
llvm-svn: 371735
Summary:
I don't have a direct motivational case for this,
but it would be good to have this for completeness/symmetry.
This pattern is basically the motivational pattern from
https://bugs.llvm.org/show_bug.cgi?id=43251
but with different predicate that requires that the offset is non-zero.
The completeness bit comes from the fact that a similar pattern (offset != zero)
will be needed for https://bugs.llvm.org/show_bug.cgi?id=43259,
so it'd seem to be good to not overlook very similar patterns..
Proofs: https://rise4fun.com/Alive/21b
Also, there is something odd with `isKnownNonZero()`, if the non-zero
knowledge was specified as an assumption, it didn't pick it up (PR43267)
Reviewers: spatel, nikic, xbolva00
Reviewed By: spatel
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67411
llvm-svn: 371718
This patch contains the basic functionality for reporting potentially
incorrect usage of __builtin_expect() by comparing the developer's
annotation against a collected PGO profile. A more detailed proposal and
discussion appears on the CFE-dev mailing list
(http://lists.llvm.org/pipermail/cfe-dev/2019-July/062971.html) and a
prototype of the initial frontend changes appear here in D65300
We revised the work in D65300 by moving the misexpect check into the
LLVM backend, and adding support for IR and sampling based profiles, in
addition to frontend instrumentation.
We add new misexpect metadata tags to those instructions directly
influenced by the llvm.expect intrinsic (branch, switch, and select)
when lowering the intrinsics. The misexpect metadata contains
information about the expected target of the intrinsic so that we can
check against the correct PGO counter when emitting diagnostics, and the
compiler's values for the LikelyBranchWeight and UnlikelyBranchWeight.
We use these branch weight values to determine when to emit the
diagnostic to the user.
A future patch should address the comment at the top of
LowerExpectIntrisic.cpp to hoist the LikelyBranchWeight and
UnlikelyBranchWeight values into a shared space that can be accessed
outside of the LowerExpectIntrinsic pass. Once that is done, the
misexpect metadata can be updated to be smaller.
In the long term, it is possible to reconstruct portions of the
misexpect metadata from the existing profile data. However, we have
avoided this to keep the code simple, and because some kind of metadata
tag will be required to identify which branch/switch/select instructions
are influenced by the use of llvm.expect
Patch By: paulkirth
Differential Revision: https://reviews.llvm.org/D66324
llvm-svn: 371635
This introduces additional rounding error in some cases. See D67434.
This reverts r371518 (git commit 18a1f0818b659cee13865b4fad2648d85984a4ed)
llvm-svn: 371634
This was actually the original intention in D67332,
but i messed up and forgot about it.
This patch was originally part of D67411, but precommitting this.
llvm-svn: 371630
adding new read attribute to an argument
Summary: Update optimization pass to prevent adding read-attribute to an
argument without removing its conflicting attribute.
A read attribute, based on the result of the attribute deduction
process, might be added to an argument. The attribute might be in
conflict with other read/write attribute currently associated with the
argument. To ensure the compatibility of attributes, conflicting
attribute, if any, must be removed before a new one is added.
The following snippet shows the current behavior of the compiler, where
the compilation process is aborted due to incompatible attributes.
$ cat x.ll
; ModuleID = 'x.bc'
%_type_of_d-ccc = type <{ i8*, i8, i8, i8, i8 }>
@d-ccc = internal global %_type_of_d-ccc <{ i8* null, i8 1, i8 13, i8 0,
i8 -127 }>, align 8
define void @foo(i32* writeonly %.aaa) {
foo_entry:
%_param_.aaa = alloca i32*, align 8
store i32* %.aaa, i32** %_param_.aaa, align 8
store i8 0, i8* getelementptr inbounds (%_type_of_d-ccc,
%_type_of_d-ccc* @d-ccc, i32 0, i32 3)
ret void
}
$ opt -O3 x.ll
Attributes 'readnone and writeonly' are incompatible!
void (i32*)* @foo
in function foo
LLVM ERROR: Broken function found, compilation aborted!
The purpose of this changeset is to fix the above error. This fix is
based on a suggestion from Johannes @jdoerfert (many thanks!!!)
Authored By: anhtuyen
Reviewer: nicholas, rnk, chandlerc, jdoerfert
Reviewed By: rnk
Subscribers: hiraditya, jdoerfert, llvm-commits, anhtuyen, LLVM
Tag: LLVM
Differential Revision: https://reviews.llvm.org/D58694
llvm-svn: 371622
