This is based on the same idea that I am using for the basic model implementation
and what I have partly already done for x86: throughput cost is number of
instructions/uops, so size/blended costs are identical except in special cases
(for example, fdiv or other known-expensive machine instructions or things like
MVE that may require cracking into >1 uop)).
Differential Revision: https://reviews.llvm.org/D90692
This reverts the revert commit 408c4408fa.
This version of the patch includes a fix for a crash caused by
treating ICmp/FCmp constant expressions as instructions.
Original message:
On some targets, like AArch64, vector selects can be efficiently lowered
if the vector condition is a compare with a supported predicate.
This patch adds a new argument to getCmpSelInstrCost, to indicate the
predicate of the feeding select condition. Note that it is not
sufficient to use the context instruction when querying the cost of a
vector select starting from a scalar one, because the condition of the
vector select could be composed of compares with different predicates.
This change greatly improves modeling the costs of certain
compare/select patterns on AArch64.
I am also planning on putting up patches to make use of the new argument in
SLPVectorizer & LV.
CallInst::updateProfWeight() creates branch_weights with i64 instead of i32.
To be more consistent everywhere and remove lots of casts from uint64_t
to uint32_t, use i64 for branch_weights.
Reviewed By: davidxl
Differential Revision: https://reviews.llvm.org/D88609
This patch changes MergeBlockIntoPredecessor to skip the call to
RemoveRedundantDbgInstrs, in effect partially reverting D71480 due to
some compile-time issues spotted in LoopUnroll and SimplifyCFG.
The call to RemoveRedundantDbgInstrs appears to have changed the
worst-case behavior of the merging utility. Loosely speaking, it seems
to have gone from O(#phis) to O(#insts).
It might not be possible to mitigate this by scanning a block to
determine whether there are any debug intrinsics to remove, since such a
scan costs O(#insts).
So: skip the call to RemoveRedundantDbgInstrs. There's surprisingly
little fallout from this, and most of it can be addressed by doing
RemoveRedundantDbgInstrs later. The exception is (the block-local
version of) SimplifyCFG, where it might just be too expensive to call
RemoveRedundantDbgInstrs.
Differential Revision: https://reviews.llvm.org/D88928
CallInst::updateProfWeight() creates branch_weights with i64 instead of i32.
To be more consistent everywhere and remove lots of casts from uint64_t
to uint32_t, use i64 for branch_weights.
Reviewed By: davidxl
Differential Revision: https://reviews.llvm.org/D88609
-debug-pass is a legacy PM only option.
Some tests checks that the pass returned that it made a change,
which is not relevant to the NPM, since passes return PreservedAnalyses.
Some tests check that passes are freed at the proper time, which is also
not relevant to the NPM.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D87945
SimplifyCFG's options should always be overridden by command line flags,
but they mistakenly weren't in the default constructor.
Reviewed By: ychen
Differential Revision: https://reviews.llvm.org/D87718
This was reverted in 503deec218
because it caused gigantic increase (3x) in branch mispredictions
in certain benchmarks on certain CPU's,
see https://reviews.llvm.org/D84108#2227365.
It has since been investigated and here are the results:
https://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20200907/827578.html
> It's an amazingly severe regression, but it's also all due to branch
> mispredicts (about 3x without this). The code layout looks ok so there's
> probably something else to deal with. I'm not sure there's anything we can
> reasonably do so we'll just have to take the hit for now and wait for
> another code reorganization to make the branch predictor a bit more happy :)
>
> Thanks for giving us some time to investigate and feel free to recommit
> whenever you'd like.
>
> -eric
So let's just reland this.
Original commit message:
I've been looking at missed vectorizations in one codebase.
One particular thing that stands out is that some of the loops
reach vectorizer in a rather mangled form, with weird PHI's,
and some of the loops aren't even in a rotated form.
After taking a more detailed look, that happened because
the loop's headers were too big by then. It is evident that
SimplifyCFG's common code hoisting transform is at fault there,
because the pattern it handles is precisely the unrotated
loop basic block structure.
Surprizingly, `SimplifyCFGOpt::HoistThenElseCodeToIf()` is enabled
by default, and is always run, unlike it's friend, common code sinking
transform, `SinkCommonCodeFromPredecessors()`, which is not enabled
by default and is only run once very late in the pipeline.
I'm proposing to harmonize this, and disable common code hoisting
until //late// in pipeline. Definition of //late// may vary,
here currently i've picked the same one as for code sinking,
but i suppose we could enable it as soon as right after
loop rotation happens.
Experimentation shows that this does indeed unsurprizingly help,
more loops got rotated, although other issues remain elsewhere.
Now, this undoubtedly seriously shakes phase ordering.
This will undoubtedly be a mixed bag in terms of both compile- and
run- time performance, codesize. Since we no longer aggressively
hoist+deduplicate common code, we don't pay the price of said hoisting
(which wasn't big). That may allow more loops to be rotated,
so we pay that price. That, in turn, that may enable all the transforms
that require canonical (rotated) loop form, including but not limited to
vectorization, so we pay that too. And in general, no deduplication means
more [duplicate] instructions going through the optimizations. But there's still
late hoisting, some of them will be caught late.
As per benchmarks i've run {F12360204}, this is mostly within the noise,
there are some small improvements, some small regressions.
One big regression i saw i fixed in rG8d487668d09fb0e4e54f36207f07c1480ffabbfd, but i'm sure
this will expose many more pre-existing missed optimizations, as usual :S
llvm-compile-time-tracker.com thoughts on this:
http://llvm-compile-time-tracker.com/compare.php?from=e40315d2b4ed1e38962a8f33ff151693ed4ada63&to=c8289c0ecbf235da9fb0e3bc052e3c0d6bff5cf9&stat=instructions
* this does regress compile-time by +0.5% geomean (unsurprizingly)
* size impact varies; for ThinLTO it's actually an improvement
The largest fallout appears to be in GVN's load partial redundancy
elimination, it spends *much* more time in
`MemoryDependenceResults::getNonLocalPointerDependency()`.
Non-local `MemoryDependenceResults` is widely-known to be, uh, costly.
There does not appear to be a proper solution to this issue,
other than silencing the compile-time performance regression
by tuning cut-off thresholds in `MemoryDependenceResults`,
at the cost of potentially regressing run-time performance.
D84609 attempts to move in that direction, but the path is unclear
and is going to take some time.
If we look at stats before/after diffs, some excerpts:
* RawSpeed (the target) {F12360200}
* -14 (-73.68%) loops not rotated due to the header size (yay)
* -272 (-0.67%) `"Number of live out of a loop variables"` - good for vectorizer
* -3937 (-64.19%) common instructions hoisted
* +561 (+0.06%) x86 asm instructions
* -2 basic blocks
* +2418 (+0.11%) IR instructions
* vanilla test-suite + RawSpeed + darktable {F12360201}
* -36396 (-65.29%) common instructions hoisted
* +1676 (+0.02%) x86 asm instructions
* +662 (+0.06%) basic blocks
* +4395 (+0.04%) IR instructions
It is likely to be sub-optimal for when optimizing for code size,
so one might want to change tune pipeline by enabling sinking/hoisting
when optimizing for size.
Reviewed By: mkazantsev
Differential Revision: https://reviews.llvm.org/D84108
This reverts commit 503deec218.
Modify FoldBranchToCommonDest to consider the cost of inserting
instructions when attempting to combine predicates to fold blocks.
The threshold can be controlled via a new option:
-simplifycfg-branch-fold-threshold which defaults to '2' to allow
the insertion of a not and another logical operator.
Differential Revision: https://reviews.llvm.org/D86526
When a switch case is folded into default's case, that's an IR change that
should be reported, update ConstantFoldTerminator accordingly.
Differential Revision: https://reviews.llvm.org/D87142
As discussed in
http://lists.llvm.org/pipermail/llvm-dev/2020-July/143801.html.
Currently no users outside of unit tests.
Replace all instances in tests of -constprop with -instsimplify.
Notable changes in tests:
* vscale.ll - @llvm.sadd.sat.nxv16i8 is evaluated by instsimplify, use a fake intrinsic instead
* InsertElement.ll - insertelement undef is removed by instsimplify in @insertelement_undef
llvm/test/Transforms/ConstProp moved to llvm/test/Transforms/InstSimplify/ConstProp
Reviewed By: lattner, nikic
Differential Revision: https://reviews.llvm.org/D85159
As disscussed in post-commit review starting with
https://reviews.llvm.org/D84108#2227365
while this appears to be mostly a win overall, especially code-size-wise,
this appears to shake //certain// code pattens in a way that is extremely
unfavorable for performance (+30% runtime regression)
on certain CPU's (i personally can't reproduce).
So until the behaviour is better understood, and a path forward is mapped,
let's back this out for now.
This reverts commit 1d51dc38d8.
Before we speculatively execute a basic block, query the cost of
inserting the necessary select instructions against the phi folding
threshold. For non-trivial insertions, a more accurate decision can
probably be made during machine if-conversion. With minsize we query
the CodeSize cost, otherwise we use SizeAndLatency.
Differential Revision: https://reviews.llvm.org/D82438
SimplifyCFG has two main folds for resumes - one when resume is directly
using the landingpad, and the other one where resume is using a PHI node.
While for the first case, we were already correctly ignoring all the
PHI nodes, and both the debug info intrinsics and lifetime intrinsics,
in the PHI-based-one, we weren't ignoring PHI's in the resume block,
and weren't ignoring lifetime intrinsics. That is clearly a bug.
On RawSpeed library, this results in +9.34% (+81) more invoke->call folds,
-0.19% (-39) landing pads, -0.24% (-81) invoke instructions
but +51 call instructions and -132 basic blocks.
Though, the run-time performance impact appears to be within the noise.
I've been looking at missed vectorizations in one codebase.
One particular thing that stands out is that some of the loops
reach vectorizer in a rather mangled form, with weird PHI's,
and some of the loops aren't even in a rotated form.
After taking a more detailed look, that happened because
the loop's headers were too big by then. It is evident that
SimplifyCFG's common code hoisting transform is at fault there,
because the pattern it handles is precisely the unrotated
loop basic block structure.
Surprizingly, `SimplifyCFGOpt::HoistThenElseCodeToIf()` is enabled
by default, and is always run, unlike it's friend, common code sinking
transform, `SinkCommonCodeFromPredecessors()`, which is not enabled
by default and is only run once very late in the pipeline.
I'm proposing to harmonize this, and disable common code hoisting
until //late// in pipeline. Definition of //late// may vary,
here currently i've picked the same one as for code sinking,
but i suppose we could enable it as soon as right after
loop rotation happens.
Experimentation shows that this does indeed unsurprizingly help,
more loops got rotated, although other issues remain elsewhere.
Now, this undoubtedly seriously shakes phase ordering.
This will undoubtedly be a mixed bag in terms of both compile- and
run- time performance, codesize. Since we no longer aggressively
hoist+deduplicate common code, we don't pay the price of said hoisting
(which wasn't big). That may allow more loops to be rotated,
so we pay that price. That, in turn, that may enable all the transforms
that require canonical (rotated) loop form, including but not limited to
vectorization, so we pay that too. And in general, no deduplication means
more [duplicate] instructions going through the optimizations. But there's still
late hoisting, some of them will be caught late.
As per benchmarks i've run {F12360204}, this is mostly within the noise,
there are some small improvements, some small regressions.
One big regression i saw i fixed in rG8d487668d09fb0e4e54f36207f07c1480ffabbfd, but i'm sure
this will expose many more pre-existing missed optimizations, as usual :S
llvm-compile-time-tracker.com thoughts on this:
http://llvm-compile-time-tracker.com/compare.php?from=e40315d2b4ed1e38962a8f33ff151693ed4ada63&to=c8289c0ecbf235da9fb0e3bc052e3c0d6bff5cf9&stat=instructions
* this does regress compile-time by +0.5% geomean (unsurprizingly)
* size impact varies; for ThinLTO it's actually an improvement
The largest fallout appears to be in GVN's load partial redundancy
elimination, it spends *much* more time in
`MemoryDependenceResults::getNonLocalPointerDependency()`.
Non-local `MemoryDependenceResults` is widely-known to be, uh, costly.
There does not appear to be a proper solution to this issue,
other than silencing the compile-time performance regression
by tuning cut-off thresholds in `MemoryDependenceResults`,
at the cost of potentially regressing run-time performance.
D84609 attempts to move in that direction, but the path is unclear
and is going to take some time.
If we look at stats before/after diffs, some excerpts:
* RawSpeed (the target) {F12360200}
* -14 (-73.68%) loops not rotated due to the header size (yay)
* -272 (-0.67%) `"Number of live out of a loop variables"` - good for vectorizer
* -3937 (-64.19%) common instructions hoisted
* +561 (+0.06%) x86 asm instructions
* -2 basic blocks
* +2418 (+0.11%) IR instructions
* vanilla test-suite + RawSpeed + darktable {F12360201}
* -36396 (-65.29%) common instructions hoisted
* +1676 (+0.02%) x86 asm instructions
* +662 (+0.06%) basic blocks
* +4395 (+0.04%) IR instructions
It is likely to be sub-optimal for when optimizing for code size,
so one might want to change tune pipeline by enabling sinking/hoisting
when optimizing for size.
Reviewed By: mkazantsev
Differential Revision: https://reviews.llvm.org/D84108
We do not thread blocks with convergent calls, but this check was missing
when we decide to insert PR Phis into it (which we only do for threading).
Differential Revision: https://reviews.llvm.org/D83936
Reviewed By: nikic
store (load float*) can be optimized to store(load i32*) in InstCombine pass.
Add store (load float*) to isProfitableToHoist to make sure we don't break
the opt in InstCombine pass.
Reviewed By: jsji
Differential Revision: https://reviews.llvm.org/D82341
Common code sinking is already guarded with a (with default-off!) flag,
so add a flag for hoisting, too.
D84108 will hopefully make hoisting off-by-default too.
SimplifyCFG was incorrectly reporting to the pass manager that it had not made
changes after folding away a PHI. This is detected in the EXPENSIVE_CHECKS
build when the function's hash changes.
Differential Revision: https://reviews.llvm.org/D83985
Summary:
The actual transform i was going after was:
https://rise4fun.com/Alive/Tp9H
```
Name: zz
Pre: isPowerOf2(C0) && isPowerOf2(C1) && C1 == C0
%t0 = and i8 %x, C0
%r = icmp eq i8 %t0, C1
=>
%t = icmp eq i8 %t0, 0
%r = xor i1 %t, -1
Name: zz
Pre: isPowerOf2(C0)
%t0 = and i8 %x, C0
%r = icmp ne i8 %t0, 0
=>
%t = icmp eq i8 %t0, 0
%r = xor i1 %t, -1
```
but as it can be seen from the current tests, we already canonicalize most of it,
and we are only missing handling multi-use non-canonical icmp predicates.
If we have both `!=0` and `==0`, even though we can CSE them,
we end up being stuck with them. We should canonicalize to the `==0`.
I believe this is one of the cleanup steps i'll need after `-scalarizer`
if i end up proceeding with my WIP alloca promotion helper pass.
Reviewers: spatel, jdoerfert, nikic
Reviewed By: nikic
Subscribers: zzheng, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D83139
Sometimes SimplifyCFG may decide to perform jump threading. In order
to do it, it follows the following algorithm:
1. Checks if the block is small enough for threading;
2. If yes, inserts a PR Phi relying that the next iteration will remove it
by performing jump threading;
3. The next iteration checks the block again and performs the threading.
This logic has a corner case: inserting the PR Phi increases block's size
by 1. If the block size at first check was max possible, one more Phi will
exceed this size, and we will neither perform threading nor remove the
created Phi node. As result, we will end up with worse IR than before.
This patch fixes this situation by excluding Phis from block size computation.
Excluding Phis from size computation for threading also makes sense by
itself because in case of threadign all those Phis will be removed.
Differential Revision: https://reviews.llvm.org/D81835
Reviewed By: asbirlea, nikic
I'm not sure if this is a regression from D81448 + D81643,
which moved at least the code cast from elsewhere,
or somehow no one triggered that before.
But now we can reach it with a non-instruction..
It is not straight-forward to write cost-model tests for constantexprs,
`-cost-model -analyze -cost-kind=` does not appear to look at them,
or maybe i'm doing it wrong.
I've encountered that via a SimplifyCFG crash,
so reduced (currently-crashing) test is added.
There are likely other instances.
For now, simply restore previous status quo of
not crashing and returning TTI::TCC_Basic.
Summary:
According to HowToUpdateDebugInfo.rst:
```
Preserving the debug locations of speculated instructions can make
it seem like a condition is true when it's not (or vice versa), which
leads to a confusing single-stepping experience
```
This patch follows the recommendation to drop debug locations on
speculated instructions.
Reviewers: aprantl, davide
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D82420
When an invoke instruction is converted to a call its
profile metadata is dropped because it has incompatible
format (see commit 16ad6eeb94).
This patch adds an attempt to convert profile data to
format of the call instruction. This used to work well
before the commit dcfa78a4cc.
Reviewers: reames
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D82071
The invoke instruction can have profile metadata with branch_weights,
which does not make sense for a call instruction and will be
rejected by the verifier.
Differential revision: https://reviews.llvm.org/D81996
If the only user of `Instr` is in a return or unreachable block, we can
sink `Instr` to the`User` safely (unless it reads/writes memory).
Return or unreachable blocks are guaranteed to execute zero
or one time, and `Instr` always dominates `User`, so they either will
be executed together (execution of `User` always implies execution
of `Instr`) or not executed at all.
Differential Revision: https://reviews.llvm.org/D80120
Reviewed By: asbirlea, jdoerfert
If we don't know anything about the alignment of a pointer, Align(1) is
still correct: all pointers are at least 1-byte aligned.
Included in this patch is a bugfix for an issue discovered during this
cleanup: pointers with "dereferenceable" attributes/metadata were
assumed to be aligned according to the type of the pointer. This
wasn't intentional, as far as I can tell, so Loads.cpp was fixed to
stop making this assumption. Frontends may need to be updated. I
updated clang's handling of C++ references, and added a release note for
this.
Differential Revision: https://reviews.llvm.org/D80072
The "null-pointer-is-valid" attribute needs to be checked by many
pointer-related combines. To make the check more efficient, convert
it from a string into an enum attribute.
In the future, this attribute may be replaced with data layout
properties.
Differential Revision: https://reviews.llvm.org/D78862
For IR generated by a compiler, this is really simple: you just take the
datalayout from the beginning of the file, and apply it to all the IR
later in the file. For optimization testcases that don't care about the
datalayout, this is also really simple: we just use the default
datalayout.
The complexity here comes from the fact that some LLVM tools allow
overriding the datalayout: some tools have an explicit flag for this,
some tools will infer a datalayout based on the code generation target.
Supporting this properly required plumbing through a bunch of new
machinery: we want to allow overriding the datalayout after the
datalayout is parsed from the file, but before we use any information
from it. Therefore, IR/bitcode parsing now has a callback to allow tools
to compute the datalayout at the appropriate time.
Not sure if I covered all the LLVM tools that want to use the callback.
(clang? lli? Misc IR manipulation tools like llvm-link?). But this is at
least enough for all the LLVM regression tests, and IR without a
datalayout is not something frontends should generate.
This change had some sort of weird effects for certain CodeGen
regression tests: if the datalayout is overridden with a datalayout with
a different program or stack address space, we now parse IR based on the
overridden datalayout, instead of the one written in the file (or the
default one, if none is specified). This broke a few AVR tests, and one
AMDGPU test.
Outside the CodeGen tests I mentioned, the test changes are all just
fixing CHECK lines and moving around datalayout lines in weird places.
Differential Revision: https://reviews.llvm.org/D78403
Simon Pilgrim