Add support for (if enabled) splitting cold functions into a separate section
in order to further boost locality of hot code.
Authored by: rjf (Ruijie Fang)
Reviewed by: hiraditya,rcorcs,vsk
Differential Revision: https://reviews.llvm.org/D85331
A function marked `noreturn` may contain unreachable terminators: these
should not be considered cold, as the function may be a trampoline.
rdar://58068594
Summary:
This cropped up in the Linux kernel where cold code was placed in an
incompatible section.
Reviewers: compnerd, vsk, tejohnson
Reviewed By: vsk
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70925
This file lists every pass in LLVM, and is included by Pass.h, which is
very popular. Every time we add, remove, or rename a pass in LLVM, it
caused lots of recompilation.
I found this fact by looking at this table, which is sorted by the
number of times a file was changed over the last 100,000 git commits
multiplied by the number of object files that depend on it in the
current checkout:
recompiles touches affected_files header
342380 95 3604 llvm/include/llvm/ADT/STLExtras.h
314730 234 1345 llvm/include/llvm/InitializePasses.h
307036 118 2602 llvm/include/llvm/ADT/APInt.h
213049 59 3611 llvm/include/llvm/Support/MathExtras.h
170422 47 3626 llvm/include/llvm/Support/Compiler.h
162225 45 3605 llvm/include/llvm/ADT/Optional.h
158319 63 2513 llvm/include/llvm/ADT/Triple.h
140322 39 3598 llvm/include/llvm/ADT/StringRef.h
137647 59 2333 llvm/include/llvm/Support/Error.h
131619 73 1803 llvm/include/llvm/Support/FileSystem.h
Before this change, touching InitializePasses.h would cause 1345 files
to recompile. After this change, touching it only causes 550 compiles in
an incremental rebuild.
Reviewers: bkramer, asbirlea, bollu, jdoerfert
Differential Revision: https://reviews.llvm.org/D70211
Factor out CodeExtractor's analysis of allocas (for shrinkwrapping
purposes), and allow the analysis to be reused.
This resolves a quadratic compile-time bug observed when compiling
AMDGPUDisassembler.cpp.o.
Pre-patch (Release + LTO clang):
```
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
176.5278 ( 57.8%) 0.4915 ( 18.5%) 177.0192 ( 57.4%) 177.4112 ( 57.3%) Hot Cold Splitting
```
Post-patch (ReleaseAsserts clang):
```
---User Time--- --System Time-- --User+System-- ---Wall Time--- --- Name ---
1.4051 ( 3.3%) 0.0079 ( 0.3%) 1.4129 ( 3.2%) 1.4129 ( 3.2%) Hot Cold Splitting
```
Testing: check-llvm, and comparing the AMDGPUDisassembler.cpp.o binary
pre- vs. post-patch.
An alternate approach is to hide CodeExtractorAnalysisCache from clients
of CodeExtractor, and to recompute the analysis from scratch inside of
CodeExtractor::extractCodeRegion(). This eliminates some redundant work
in the shrinkwrapping legality check. However, some clients continue to
exhibit O(n^2) compile time behavior as computing the analysis is O(n).
rdar://55912966
Differential Revision: https://reviews.llvm.org/D68616
llvm-svn: 374089
When a cold path is outlined, the value tracking in the assumption cache may be
invalidated due to the code motion. We would previously trip an assertion in
subsequent passes (but required the passes to happen in a single run as the
assumption cache is shared across the passes). Invalidating the cache ensures
that we get the correct information when needed with the legacy pass manager as
well.
llvm-svn: 372667
Now that we've moved to C++14, we no longer need the llvm::make_unique
implementation from STLExtras.h. This patch is a mechanical replacement
of (hopefully) all the llvm::make_unique instances across the monorepo.
llvm-svn: 369013
Create method `optForNone()` testing for the function level equivalent of
`-O0` and refactor appropriately.
Differential revision: https://reviews.llvm.org/D59852
llvm-svn: 357638
Part 2 of CSPGO changes (mostly related to ProfileSummary).
Note that I use a default parameter in setProfileSummary() and getSummary().
This is to break the dependency in clang. I will make the parameter explicit
after changing clang in a separated patch.
Differential Revision: https://reviews.llvm.org/D54175
llvm-svn: 355131
Splitting can make sanitizer errors harder to understand, as the
trapping instruction may not be in the function where the bug was
detected.
rdar://48142697
llvm-svn: 354931
Summary: Assumption cache's self-updating mechanism does not correctly handle the case when blocks are extracted from the function by the CodeExtractor. As a result function's assumption cache may have stale references to the llvm.assume calls that were moved to the outlined function. This patch fixes this problem by removing extracted llvm.assume calls from the function’s assumption cache.
Reviewers: hfinkel, vsk, fhahn, davidxl, sanjoy
Reviewed By: hfinkel, vsk
Subscribers: llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D57215
llvm-svn: 353500
Summary:
When compiling with profile data, ensure the split cold function gets
cold function_entry_count metadata (just use 0 since it should be cold).
Otherwise with function sections it will not be placed in the unlikely
text section with other cold code.
Reviewers: vsk
Subscribers: sebpop, hiraditya, davidxl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D57900
llvm-svn: 353434
Resumes that are not reachable from a cleanup landing pad are considered
to be unreachable. It’s not safe to split them out.
rdar://47808235
llvm-svn: 353242
The main goal of the model is to avoid *increasing* function size, as
that would eradicate any memory locality benefits from splitting. This
happens when:
- There are too many inputs or outputs to the cold region. Argument
materialization and reloads of outputs have a cost.
- The cold region has too many distinct exit blocks, causing a large
switch to be formed in the caller.
- The code size cost of the split code is less than the cost of a
set-up call.
A secondary goal is to prevent excessive overall binary size growth.
With the cost model in place, I experimented to find a splitting
threshold that works well in practice. To make warm & cold code easily
separable for analysis purposes, I moved split functions to a "cold"
section. I experimented with thresholds between [0, 4] and set the
default to the threshold which minimized geomean __text size.
Experiment data from building LNT+externals for X86 (N = 639 programs,
all sizes in bytes):
| Configuration | __text geom size | __cold geom size | TEXT geom size |
| **-Os** | 1736.3 | 0, n=0 | 10961.6 |
| -Os, thresh=0 | 1740.53 | 124.482, n=134 | 11014 |
| -Os, thresh=1 | 1734.79 | 57.8781, n=90 | 10978.6 |
| -Os, thresh=2 | ** 1733.85 ** | 65.6604, n=61 | 10977.6 |
| -Os, thresh=3 | 1733.85 | 65.3071, n=61 | 10977.6 |
| -Os, thresh=4 | 1735.08 | 67.5156, n=54 | 10965.7 |
| **-Oz** | 1554.4 | 0, n=0 | 10153 |
| -Oz, thresh=2 | ** 1552.2 ** | 65.633, n=61 | 10176 |
| **-O3** | 2563.37 | 0, n=0 | 13105.4 |
| -O3, thresh=2 | ** 2559.49 ** | 71.1072, n=61 | 13162.4 |
Picking thresh=2 reduces the geomean __text section size by 0.14% at
-Os, -Oz, and -O3 and causes ~0.2% growth in the TEXT segment. Note that
TEXT size is page-aligned, whereas section sizes are byte-aligned.
Experiment data from building LNT+externals for ARM64 (N = 558 programs,
all sizes in bytes):
| Configuration | __text geom size | __cold geom size | TEXT geom size |
| **-Os** | 1763.96 | 0, n=0 | 42934.9 |
| -Os, thresh=2 | ** 1760.9 ** | 76.6755, n=61 | 42934.9 |
Picking thresh=2 reduces the geomean __text section size by 0.17% at
-Os and causes no growth in the TEXT segment.
Measurements were done with D57082 (r352080) applied.
Differential Revision: https://reviews.llvm.org/D57125
llvm-svn: 352228
While a cold invoke itself and its unwind destination can't be
extracted, code which unconditionally executes before/after the invoke
may still be profitable to extract.
With cost model changes from D57125 applied, this gives a 3.5% increase
in split text across LNT+externals on arm64 at -Os.
llvm-svn: 352160
Instead of manually computing DT and PDT, we can get the from the pass
manager, which ideally has them already cached. With the new pass
manager, we could even preserve DT/PDT on a per function basis in a
module pass.
I think this also addresses the TODO about re-using the computed DTs for
BFI. IIUC, GetBFI will fetch the DT from the pass manager and when we
will fetch the cached version later.
Reviewers: vsk, hiraditya, tejohnson, thegameg, sebpop
Reviewed By: vsk
Differential Revision: https://reviews.llvm.org/D57092
llvm-svn: 352036
The splitting pass does not need BFI unless the Module actually has a profile
summary. Do not calcualte BFI unless the summary is present.
For the sqlite3 amalgamation, this reduces time spent in the splitting pass
from 0.4% of the total to under 0.1%.
llvm-svn: 351894
The splitting pass does not need (post)domtrees until after it's found a
cold block. Defer domtree calculation until a cold block is found.
For the sqlite3 amalgamation, this reduces time spent in the splitting
pass from 0.8% of the total to 0.4%.
llvm-svn: 351892
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
If an inherently cold function is found, mark it as cold. For now this
means applying the `cold` and `minsize` attributes.
As a drive-by, revisit and clean up the criteria for considering a
function for splitting. Add tests.
llvm-svn: 351623
Prior to r348205, extracting code regions with live output values was
disabled because of a miscompilation (PR39433). Lift the restriction as
PR39433 has been addressed.
Tested on LNT+externals, on a run of check-llvm in a stage2 build, and
with a full build of iOS (with hot/cold splitting enabled).
As a drive-by, remove an errant TODO.
llvm-svn: 351492
Resuming exception unwinding is roughly as unlikely as throwing an
exception.
Tested on LNT+externals (in particular, the C++ EH regression tests
provide end-to-end test coverage), as well as with a full build of iOS.
llvm-svn: 351491
This gets rid of the brittle/mysterious calls to @sink()/@sideeffect()
peppered throughout the test cases. They are no longer needed to force
splitting to occur.
llvm-svn: 351480
It's currently not safe to outline landingpad instructions (see
llvm.org/PR39917). Like @llvm.eh.typeid.for, the order and content of
previous landingpad instructions in a function alters the lowering of
subsequent landingpads by renumbering type info ID's. Outlining a
landingpad therefore breaks exception handling & unwinding.
llvm-svn: 348870
The splitting pass uses its 'unlikelyExecuted' predicate to statically
decide which blocks are cold.
- Do not treat noreturn calls as if they are cold unless they are actually
marked cold. This is motivated by functions like exit() and longjmp(), which
are not beneficial to outline.
- Do not treat inline asm as an outlining barrier. In practice asm("") is
frequently used to inhibit basic block merging; enabling outlining in this case
results in substantial memory savings.
- Treat invokes of cold functions as cold.
As a drive-by, remove the 'exceptionHandlingFunctions' predicate, because it's
no longer needed. The pass can identify & outline blocks dominated by EH pads,
so there's no need to special-case __cxa_begin_catch etc.
Differential Revision: https://reviews.llvm.org/D54244
llvm-svn: 348640
Algorithm: Identify maximal cold regions and put them in a worklist. If
a candidate region overlaps with another, discard it. While the worklist
is full, remove a single-entry sub-region from the worklist and attempt
to outline it. By the non-overlap property, this should not invalidate
parts of the domtree pertaining to other outlining regions.
Testing: LNT results on X86 are clean. With test-suite + externals, llvm
outlines 134KB pre-patch, and 352KB post-patch (+ ~2.6x). The file
483.xalancbmk/src/Constants.cpp stands out as an extreme case where llvm
outlines over 100 times in some functions (mostly EH paths). There was
not a significant performance impact pre vs. post-patch.
Differential Revision: https://reviews.llvm.org/D53887
llvm-svn: 348639
Every Analysis pass has a get method that returns a reference of the Result of
the Analysis, for example, BlockFrequencyInfo
&BlockFrequencyInfoWrapperPass::getBFI(). I believe that
ProfileSummaryInfo::getPSI() is the only exception to that, as it was returning
a pointer.
Another change is renaming isHotBB and isColdBB to isHotBlock and isColdBlock,
respectively. Most methods use BB as the argument of variable names while
methods usually refer to Basic Blocks as Blocks, instead of BB. For example,
Function::getEntryBlock, Loop:getExitBlock, etc.
I also fixed one of the comments.
Patch by Rodrigo Caetano Rocha!
Differential Revision: https://reviews.llvm.org/D54669
llvm-svn: 347182
Using TargetTransformInfo allows the splitting pass to factor in the
code size cost of instructions as it decides whether or not outlining is
profitable.
This did not regress the overall amount of outlining seen on the handful
of internal frameworks I tested.
Thanks to Jun Bum Lim for suggesting this!
Differential Revision: https://reviews.llvm.org/D53835
llvm-svn: 346108
It can be profitable to outline single-block cold regions because they
may be large.
Allow outlining single-block regions if they have over some threshold of
non-debug, non-terminator instructions. I chose 3 as the threshold after
experimenting with several internal frameworks.
In practice, reducing the threshold further did not give much
improvement, whereas increasing it resulted in substantial regressions.
Differential Revision: https://reviews.llvm.org/D53824
llvm-svn: 345524
The current splitting algorithm works in three stages:
1) Identify cold blocks, then
2) Use forward/backward propagation to mark hot blocks, then
3) Grow a SESE region of blocks *outside* of the set of hot blocks and
start outlining.
While testing this pass on Apple internal frameworks I noticed that some
kinds of control flow (e.g. loops) are never outlined, even though they
unconditionally lead to / follow cold blocks. I noticed two other issues
related to how cold regions are identified:
- An inconsistency can arise in the internal state of the hotness
propagation stage, as a block may end up in both the ColdBlocks set
and the HotBlocks set. Further inconsistencies can arise as these sets
do not match what's in ProfileSummaryInfo.
- It isn't necessary to limit outlining to single-exit regions.
This patch teaches the splitting algorithm to identify maximal cold
regions and outline them. A maximal cold region is defined as the set of
blocks post-dominated by a cold sink block, or dominated by that sink
block. This approach can successfully outline loops in the cold path. As
a side benefit, it maintains less internal state than the current
approach.
Due to a limitation in CodeExtractor, blocks within the maximal cold
region which aren't dominated by a single entry point (a so-called "max
ancestor") are filtered out.
Results:
- X86 (LNT + -Os + externals): 134KB of TEXT were outlined compared to
47KB pre-patch, or a ~3x improvement. Did not see a performance impact
across two runs.
- AArch64 (LNT + -Os + externals + Apple-internal benchmarks): 149KB
of TEXT were outlined. Ditto re: performance impact.
- Outlining results improve marginally in the internal frameworks I
tested.
Follow-ups:
- Outline more than once per function, outline large single basic
blocks, & try to remove unconditional branches in outlined functions.
Differential Revision: https://reviews.llvm.org/D53627
llvm-svn: 345209
Summary:
The current default of appending "_"+entry block label to the new
extracted cold function breaks demangling. Change the deliminator from
"_" to "." to enable demangling. Because the header block label will
be empty for release compile code, use "extracted" after the "." when
the label is empty.
Additionally, add a mechanism for the client to pass in an alternate
suffix applied after the ".", and have the hot cold split pass use
"cold."+Count, where the Count is currently 1 but can be used to
uniquely number multiple cold functions split out from the same function
with D53588.
Reviewers: sebpop, hiraditya
Subscribers: llvm-commits, erik.pilkington
Differential Revision: https://reviews.llvm.org/D53534
llvm-svn: 345178
Outlined code is cold by assumption, so it makes sense to optimize it
for minimal code size rather than performance.
After r344869 moved the splitting pass to the end of the IR pipeline,
this does not result in much of a code size reduction. This is probably
because a comparatively small number backend transforms make use of the
MinSize hint.
Running LNT on x86_64, I see that 33/1020 binaries shrink for a total of
919 bytes of TEXT reduction. I didn't measure a significant performance
impact.
Differential Revision: https://reviews.llvm.org/D53518
llvm-svn: 345072
r344558 added an assignment to a TerminatorInst* from
BasicBlock::getTerminatorInst(), but BasicBlock::getTerminatorInst() returns an
Instruction* rather than a TerminatorInst* since r344504 so this fails to
compile.
Changing the variable to an Instruction* should get the bots building again.
llvm-svn: 344566
Make the code of blockEndsInUnreachable to match the function
blockEndsInUnreachable in CodeGen/BranchFolding.cpp. I also have
added a note to make sure the code of this function will not be
modified unless the back-end version is also modified.
An early return before outlining has been added to avoid
outlining the full function body when the first block in the
function is marked cold.
The static analysis of cold code has been amended to avoid
marking the whole function as cold by back-propagation
because the back-propagation would mark blocks with return
statements as cold.
The patch adds debug statements to help discover these problems.
Differential Revision: https://reviews.llvm.org/D52904
llvm-svn: 344558
by `getTerminator()` calls instead be declared as `Instruction`.
This is the biggest remaining chunk of the usage of `getTerminator()`
that insists on the narrow type and so is an easy batch of updates.
Several files saw more extensive updates where this would cascade to
requiring API updates within the file to use `Instruction` instead of
`TerminatorInst`. All of these were trivial in nature (pervasively using
`Instruction` instead just worked).
llvm-svn: 344502
Aditya Kumar