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
2nd part of D57095 with the same reason, just in another place. We never
fold branches that are not immediately in the current loop, but this check
is missing in `IsEdgeLive` As result, it may think that the edge in subloop is
dead while it's live. It's a pessimization in the current stance.
Differential Revision: https://reviews.llvm.org/D57147
Reviewed By: rupprecht
llvm-svn: 352170
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
Otherwise they are treated as dynamic allocas, which ends up increasing
code size significantly. This reduces size of Chromium base_unittests
by 2MB (6.7%).
Differential Revision: https://reviews.llvm.org/D57205
llvm-svn: 352152
Summary:
MemorySSA needs updating each time an instruction is moved.
LICM and control flow hoisting re-hoists instructions, thus needing another update when re-moving those instructions.
Pending cleanup: the MSSA update is duplicated, should be moved inside moveInstructionBefore.
Reviewers: jnspaulsson
Subscribers: sanjoy, jlebar, Prazek, george.burgess.iv, llvm-commits
Differential Revision: https://reviews.llvm.org/D57176
llvm-svn: 352092
Performing splitting early has several advantages:
- Inhibiting inlining of cold code early improves code size. Compared
to scheduling splitting at the end of the pipeline, this cuts code
size growth in half within the iOS shared cache (0.69% to 0.34%).
- Inhibiting inlining of cold code improves compile time. There's no
need to inline split cold functions, or to inline as much *within*
those split functions as they are marked `minsize`.
- During LTO, extra work is only done in the pre-link step. Less code
must be inlined during cross-module inlining.
An additional motivation here is that the most common cold regions
identified by the static/conservative splitting heuristic can (a) be
found before inlining and (b) do not grow after inlining. E.g.
__assert_fail, os_log_error.
The disadvantages are:
- Some opportunities for splitting out cold code may be missed. This
gap can potentially be narrowed by adding a worklist algorithm to the
splitting pass.
- Some opportunities to reduce code size may be lost (e.g. store
sinking, when one side of the CFG diamond is split). This does not
outweigh the code size benefits of splitting earlier.
On net, splitting early in the pipeline has substantial code size
benefits, and no major effects on memory locality or performance. We
measured memory locality using ktrace data, and consistently found that
10% fewer pages were needed to capture 95% of text page faults in key
iOS benchmarks. We measured performance on frequency-stabilized iOS
devices using LNT+externals.
This reverses course on the decision made to schedule splitting late in
r344869 (D53437).
Differential Revision: https://reviews.llvm.org/D57082
llvm-svn: 352080
After submitting https://reviews.llvm.org/D57138, I realized it was slightly more conservative than needed. The scalar indices don't appear to be a problem on a vector gep, we even had a test for that.
Differential Revision: https://reviews.llvm.org/D57161
llvm-svn: 352061
This is an alternative to https://reviews.llvm.org/D57103. After discussion, we dedicided to check this in as a temporary workaround, and pursue a true fix under the original thread.
The issue at hand is that the base rewriting algorithm doesn't consider the fact that GEPs can turn a scalar input into a vector of outputs. We had handling for scalar GEPs and fully vector GEPs (i.e. all vector operands), but not the scalar-base + vector-index forms. A true fix here requires treating GEP analogously to extractelement or shufflevector.
This patch is merely a workaround. It simply hides the crash at the cost of some ugly code gen for this presumable very rare pattern.
Differential Revision: https://reviews.llvm.org/D57138
llvm-svn: 352059
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
When we choose whether or not we should mark block as dead, we have an
inconsistent logic in markup of live blocks.
- We take candidate IF its terminator branches on constant AND it is immediately
in current loop;
- We mark successor live IF its terminator doesn't branch by constant OR it branches
by constant and the successor is its always taken block.
What we are missing here is that when the terminator branches on a constant but is
not taken as a candidate because is it not immediately in the current loop, we will
mark only one (always taken) successor as live. Therefore, we do NOT do the actual
folding but may NOT mark one of the successors as live. So the result of markup is
wrong in this case, and we may then hit various asserts.
Thanks Jordan Rupprech for reporting this!
Differential Revision: https://reviews.llvm.org/D57095
Reviewed By: rupprecht
llvm-svn: 352024
Summary:
UBSan wants to detect when unreachable code is actually reached, so it
adds instrumentation before every `unreachable` instruction. However,
the optimizer will remove code after calls to functions marked with
`noreturn`. To avoid this UBSan removes `noreturn` from both the call
instruction as well as from the function itself. Unfortunately, ASan
relies on this annotation to unpoison the stack by inserting calls to
`_asan_handle_no_return` before `noreturn` functions. This is important
for functions that do not return but access the the stack memory, e.g.,
unwinder functions *like* `longjmp` (`longjmp` itself is actually
"double-proofed" via its interceptor). The result is that when ASan and
UBSan are combined, the `noreturn` attributes are missing and ASan
cannot unpoison the stack, so it has false positives when stack
unwinding is used.
Changes:
# UBSan now adds the `expect_noreturn` attribute whenever it removes
the `noreturn` attribute from a function
# ASan additionally checks for the presence of this attribute
Generated code:
```
call void @__asan_handle_no_return // Additionally inserted to avoid false positives
call void @longjmp
call void @__asan_handle_no_return
call void @__ubsan_handle_builtin_unreachable
unreachable
```
The second call to `__asan_handle_no_return` is redundant. This will be
cleaned up in a follow-up patch.
rdar://problem/40723397
Reviewers: delcypher, eugenis
Tags: #sanitizers
Differential Revision: https://reviews.llvm.org/D56624
llvm-svn: 352003
Summary:
Profile sample files include the number of times each entry or inlined
call site is sampled. This is translated into the entry count metadta
on functions.
When sample data is being read, if a call site that was inlined
in the sample program is considered cold and not inlined, then
the entry count of the out-of-line functions does not reflect
the current compilation.
In this patch, we note call sites where the function was not inlined
and as a last action of the sample profile loading, we update the
called function's entry count to reflect the calls from these
call sites which are not included in the profile file.
Reviewers: danielcdh, wmi, Kader, modocache
Reviewed By: wmi
Subscribers: davidxl, eraman, llvm-commits
Differential Revision: https://reviews.llvm.org/D52845
llvm-svn: 352001
Summary:
Renamed setBaseDiscriminator to cloneWithBaseDiscriminator, to match
similar APIs. Also changed its behavior to copy over the other
discriminator components, instead of eliding them.
Renamed cloneWithDuplicationFactor to
cloneByMultiplyingDuplicationFactor, which more closely matches what
this API does.
Reviewers: dblaikie, wmi
Reviewed By: dblaikie
Subscribers: zzheng, llvm-commits
Differential Revision: https://reviews.llvm.org/D56220
llvm-svn: 351996
VPlan-native path
Context: Patch Series #2 for outer loop vectorization support in LV
using VPlan. (RFC:
http://lists.llvm.org/pipermail/llvm-dev/2017-December/119523.html).
Patch series #2 checks that inner loops are still trivially lock-step
among all vector elements. Non-loop branches are blindly assumed as
divergent.
Changes here implement VPlan based predication algorithm to compute
predicates for blocks that need predication. Predicates are computed
for the VPLoop region in reverse post order. A block's predicate is
computed as OR of the masks of all incoming edges. The mask for an
incoming edge is computed as AND of predecessor block's predicate and
either predecessor's Condition bit or NOT(Condition bit) depending on
whether the edge from predecessor block to the current block is true
or false edge.
Reviewers: fhahn, rengolin, hsaito, dcaballe
Reviewed By: fhahn
Patch by Satish Guggilla, thanks!
Differential Revision: https://reviews.llvm.org/D53349
llvm-svn: 351990
This saves a cbz+cold call in the interceptor ABI, as well as a realign
in both ABIs, trading off a dcache entry against some branch predictor
entries and some code size.
Unfortunately the functionality is hidden behind a flag because ifunc is
known to be broken on static binaries on Android.
Differential Revision: https://reviews.llvm.org/D57084
llvm-svn: 351989
This patch relaxes restrictions on types of latch condition and range check.
In current implementation, they should match. This patch allows to handle
wide range checks against narrow condition. The motivating example is the
following:
int N = ...
for (long i = 0; (int) i < N; i++) {
if (i >= length) deopt;
}
In this patch, the option that enables this support is turned off by
default. We'll wait until it is switched to true.
Differential Revision: https://reviews.llvm.org/D56837
Reviewed By: reames
llvm-svn: 351926
Each hwasan check requires emitting a small piece of code like this:
https://clang.llvm.org/docs/HardwareAssistedAddressSanitizerDesign.html#memory-accesses
The problem with this is that these code blocks typically bloat code
size significantly.
An obvious solution is to outline these blocks of code. In fact, this
has already been implemented under the -hwasan-instrument-with-calls
flag. However, as currently implemented this has a number of problems:
- The functions use the same calling convention as regular C functions.
This means that the backend must spill all temporary registers as
required by the platform's C calling convention, even though the
check only needs two registers on the hot path.
- The functions take the address to be checked in a fixed register,
which increases register pressure.
Both of these factors can diminish the code size effect and increase
the performance hit of -hwasan-instrument-with-calls.
The solution that this patch implements is to involve the aarch64
backend in outlining the checks. An intrinsic and pseudo-instruction
are created to represent a hwasan check. The pseudo-instruction
is register allocated like any other instruction, and we allow the
register allocator to select almost any register for the address to
check. A particular combination of (register selection, type of check)
triggers the creation in the backend of a function to handle the check
for specifically that pair. The resulting functions are deduplicated by
the linker. The pseudo-instruction (really the function) is specified
to preserve all registers except for the registers that the AAPCS
specifies may be clobbered by a call.
To measure the code size and performance effect of this change, I
took a number of measurements using Chromium for Android on aarch64,
comparing a browser with inlined checks (the baseline) against a
browser with outlined checks.
Code size: Size of .text decreases from 243897420 to 171619972 bytes,
or a 30% decrease.
Performance: Using Chromium's blink_perf.layout microbenchmarks I
measured a median performance regression of 6.24%.
The fact that a perf/size tradeoff is evident here suggests that
we might want to make the new behaviour conditional on -Os/-Oz.
But for now I've enabled it unconditionally, my reasoning being that
hwasan users typically expect a relatively large perf hit, and ~6%
isn't really adding much. We may want to revisit this decision in
the future, though.
I also tried experimenting with varying the number of registers
selectable by the hwasan check pseudo-instruction (which would result
in fewer variants being created), on the hypothesis that creating
fewer variants of the function would expose another perf/size tradeoff
by reducing icache pressure from the check functions at the cost of
register pressure. Although I did observe a code size increase with
fewer registers, I did not observe a strong correlation between the
number of registers and the performance of the resulting browser on the
microbenchmarks, so I conclude that we might as well use ~all registers
to get the maximum code size improvement. My results are below:
Regs | .text size | Perf hit
-----+------------+---------
~all | 171619972 | 6.24%
16 | 171765192 | 7.03%
8 | 172917788 | 5.82%
4 | 177054016 | 6.89%
Differential Revision: https://reviews.llvm.org/D56954
llvm-svn: 351920
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
This patch adds support of guards expressed as branches by widenable
conditions in Loop Predication.
Differential Revision: https://reviews.llvm.org/D56081
Reviewed By: reames
llvm-svn: 351805
Deopt operands are generally intended to record information about a site in code with minimal perturbation of the surrounding code. Idiomatically, they also tend to appear down rare paths. Putting these together, we have an obvious case for extending CVP w/deopt operand constant folding. Arguably, we should be doing this for all operands on all instructions, but that's definitely a much larger and risky change.
Differential Revision: https://reviews.llvm.org/D55678
llvm-svn: 351774
As noted in https://bugs.llvm.org/show_bug.cgi?id=36651, the specialization for
isPodLike<std::pair<...>> did not match the expectation of
std::is_trivially_copyable which makes the memcpy optimization invalid.
This patch renames the llvm::isPodLike trait into llvm::is_trivially_copyable.
Unfortunately std::is_trivially_copyable is not portable across compiler / STL
versions. So a portable version is provided too.
Note that the following specialization were invalid:
std::pair<T0, T1>
llvm::Optional<T>
Tests have been added to assert that former specialization are respected by the
standard usage of llvm::is_trivially_copyable, and that when a decent version
of std::is_trivially_copyable is available, llvm::is_trivially_copyable is
compared to std::is_trivially_copyable.
As of this patch, llvm::Optional is no longer considered trivially copyable,
even if T is. This is to be fixed in a later patch, as it has impact on a
long-running bug (see r347004)
Note that GCC warns about this UB, but this got silented by https://reviews.llvm.org/D50296.
Differential Revision: https://reviews.llvm.org/D54472
llvm-svn: 351701
This causes a couple of changes in the upgrade tests as signed/unsigned eq/ne are equivalent and we constant fold true/false codes, these changes are the same as what we already do for avx512 cmp/ucmp.
Noticed while cleaning up vector integer comparison costs for PR40376.
llvm-svn: 351697
Followup to D55745, this time handling comparisons with ugt and ult
predicates (which are the canonical forms for non-equality predicates).
For ctlz we can convert into a simple icmp, for cttz we can convert
into a mask check.
Differential Revision: https://reviews.llvm.org/D56355
llvm-svn: 351645
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
This modification of the currently unused inter-procedural constant
propagation pass (IPConstantPropagation) shows how abstract call sites
enable optimization of callback calls alongside direct and indirect
calls. Through minimal changes, mostly dealing with the partial mapping
of callbacks, inter-procedural constant propagation was enabled for
callbacks, e.g., OpenMP runtime calls or pthreads_create.
Differential Revision: https://reviews.llvm.org/D56447
llvm-svn: 351628
Thanks to Nikita Popov for pointing out this missed case.
This is a follow-up to r351411, which disabled function merging for
vararg functions outright due to a miscompile (see llvm.org/PR40345).
Differential Revision: https://reviews.llvm.org/D56865
llvm-svn: 351624
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
CodeExtractor permits extracting a region of blocks from a function even
when values defined within the region are used outside of it.
This is typically done by creating an alloca in the original function
and reloading the alloca after a call to the extracted function.
Wrap the reload in lifetime start/end markers to promote stack coloring.
Suggested by Sergei Kachkov!
Differential Revision: https://reviews.llvm.org/D56045
llvm-svn: 351621
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
If the sample profile has no inlining hierachy information included, we call
the sample profile is flattened. For flattened profile, in ThinLTO postlink
phase, SampleProfileLoader's hot function inlining and profile annotation will
do nothing, so it is better to save the effort to read in the profile and run
the sample profile loader pass. It is helpful for reducing compile time when
the flattened profile is huge.
Differential Revision: https://reviews.llvm.org/D54819
llvm-svn: 351476
Summary:
InstCombine's sinking algorithm only thinks about memory. It doesn't
think about non-memory constraints like stack object lifetime. It can
sink dynamic allocas across a stacksave call, which may be used with
stackrestore, which can incorrectly reduce the lifetime of the dynamic
alloca.
Fixes PR40365
Reviewers: hfinkel, efriedma
Subscribers: hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D56872
llvm-svn: 351475
Summary:
If LTOUnit splitting is disabled, the module summary analysis computes
the summary information necessary to perform single implementation
devirtualization during the thin link with the index and no IR. The
information collected from the regular LTO IR in the current hybrid WPD
algorithm is summarized, including:
1) For vtable definitions, record the function pointers and their offset
within the vtable initializer (subsumes the information collected from
IR by tryFindVirtualCallTargets).
2) A record for each type metadata summarizing the vtable definitions
decorated with that metadata (subsumes the TypeIdentiferMap collected
from IR).
Also added are the necessary bitcode records, and the corresponding
assembly support.
The index-based WPD will be sent as a follow-on.
Depends on D53890.
Reviewers: pcc
Subscribers: mehdi_amini, Prazek, inglorion, eraman, steven_wu, dexonsmith, arphaman, llvm-commits
Differential Revision: https://reviews.llvm.org/D54815
llvm-svn: 351453
During the transforms in LoopSimplifyCFG, when we remove a dead exiting edge, the
parent loop may stop being reachable from the child loop, and therefore they become
siblings. If the former child loop had uses of some values from its former parent loop,
now such uses will require LCSSA Phis, even if they weren't needed before. So we must
form LCSSA for all loops that stopped being ancestors of the current loop in this case.
Differential Revision: https://reviews.llvm.org/D56144
Reviewed By: fedor.sergeev
llvm-svn: 351434
Vedant Kumar