This change slightly relaxed the current ICP threshold in top-down inliner, specifically always allow one ICP for it. It shows some perf improvements on SPEC and our internal benchmarks. Also renamed the previous flag. We can also try to turn off PGO ICP in the future.
Reviewed By: wenlei, hoy, wmi
Differential Revision: https://reviews.llvm.org/D106588
This patch teaches the sample profile loader to merge function
attributes after inlining functions.
Without this patch, the compiler could inline a function requiring the
512-bit vector width into its caller without merging function
attributes, triggering a failure during instruction selection.
Differential Revision: https://reviews.llvm.org/D105729
As a follow-up to https://reviews.llvm.org/D104129, I'm cleaning up the danling probe related code in both the compiler and llvm-profgen.
I'm seeing a 5% size win for the pseudo_probe section for SPEC2017 and 10% for Ciner. Certain benchmark such as 602.gcc has a 20% size win. No obvious difference seen on build time for SPEC2017 and Cinder.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D104477
Previously dangling samples were represented by INT64_MAX in sample profile while probes never executed were not reported. This was based on an observation that dangling probes were only at a smaller portion than zero-count probes. However, with compiler optimizations, dangling probes end up becoming at large portion of all probes in general and reporting them does not make sense from profile size point of view. This change flips sample reporting by reporting zero-count probes instead. This enabled dangling probe to be represented by none (missing entry in profile). This has a couple benefits:
1. Reducing sample profile size in optimize mode, even when the number of non-executed probes outperform the number of dangling probes, since INT64_MAX takes more space over 0 to encode.
2. Binary size savings. No need to encode dangling probe anymore, since missing probes are treated as dangling in the profile reader.
3. Reducing compiler work to track dangling probes. However, for probes that are real dead and removed, we still need the compiler to identify them so that they can be reported as zero-count, instead of mistreated as dangling probes.
4. Improving counts quality by respecting the counts already collected on the non-dangling copy of a probe. A probe, when duplicated, gets two copies at runtime. If one of them is dangling while the other is not, merging the two probes at profile generation time will cause the real samples collected on the non-dangling one to be discarded. Not reporting the dangling counterpart will keep the real samples.
5. Better readability.
6. Be consistent with non-CS dwarf line number based profile. Zero counts are trusted by the compiler counts inferencer while missing counts will be inferred by the compiler.
Note that the current patch does include any work for #3. There will be follow-up changes.
For #1, I've seen for a large Facebook service, the text profile is reduced by 7%. For extbinary profile, the size of LBRProfileSection is reduced by 35%.
For #4, I have seen general counts quality for SPEC2017 is improved by 10%.
Reviewed By: wenlei, wlei, wmi
Differential Revision: https://reviews.llvm.org/D104129
The current implementation for computing relative block frequencies does
not handle correctly control-flow graphs containing irreducible loops. This
results in suboptimally generated binaries, whose perf can be up to 5%
worse than optimal.
To resolve the problem, we apply a post-processing step, which iteratively
updates block frequencies based on the frequencies of their predesessors.
This corresponds to finding the stationary point of the Markov chain by
an iterative method aka "PageRank computation". The algorithm takes at
most O(|E| * IterativeBFIMaxIterations) steps but typically converges faster.
It is turned on by passing option `use-iterative-bfi-inference`
and applied only for functions containing profile data and irreducible loops.
Tested on SPEC06/17, where it is helping to get correct profile counts for one of
the binaries (403.gcc). In prod binaries, we've seen a speedup of up to 2%-5%
for binaries containing functions with hot irreducible loops.
Reviewed By: hoy, wenlei, davidxl
Differential Revision: https://reviews.llvm.org/D103289
This patch was split from https://reviews.llvm.org/D102246
[SampleFDO] New hierarchical discriminator for Flow Sensitive SampleFDO
This is mainly for ProfileData part of change. It will load
FS Profile when such profile is detected. For an extbinary format profile,
create_llvm_prof tool will add a flag to profile summary section.
For other format profiles, the users need to use an internal option
(-profile-isfs) to tell the compiler that the profile uses FS discriminators.
This patch also simplified the bit API used by FS discriminators.
Differential Revision: https://reviews.llvm.org/D103041
This change tries to fix a place missing `moveAndDanglePseudoProbes `. In FoldValueComparisonIntoPredecessors, it folds the BB into predecessors and then marked the BB unreachable. However, the original logic from the BB is still alive, deleting the probe will mislead the SampleLoader mark it as zero count sample.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D102721
Sample profile loader can be run in both LTO prelink and postlink. Currently the counts annoation in postilnk doesn't fully overwrite what's done in prelink. I'm adding a switch (`-overwrite-existing-weights=1`) to enable a full overwrite, which includes:
1. Clear old metadata for calls when their parent block has a zero count. This could be caused by prelink code duplication.
2. Clear indirect call metadata if somehow all the rest targets have a sum of zero count.
3. Overwrite branch weight for basic blocks.
With a CS profile, I was seeing #1 and #2 help reduce code size by preventing post-sample ICP and CGSCC inliner working on obsolete metadata, which come from a partial global inlining in prelink. It's not expected to work well for non-CS case with a less-accurate post-inline count quality.
It's worth calling out that some prelink optimizations can damage counts quality in an irreversible way. One example is the loop rotate optimization. Due to lack of exact loop entry count (profiling can only give loop iteration count and loop exit count), moving one iteration out of the loop body leaves the rest iteration count unknown. We had to turn off prelink loop rotate to achieve a better postlink counts quality. A even better postlink counts quality can be archived by turning off prelink CGSCC inlining which is not context-sensitive.
Reviewed By: wenlei, wmi
Differential Revision: https://reviews.llvm.org/D102537
With prelink inlining, pseudo probes with same ID can come from different inline contexts. Such probes should not share samples and their factors should be fixed up separately.
I'm seeing 0.3% speedup for SPEC2017 overall. Benchmark 631.deepsjeng_s benefits the most, about 4%.
Reviewed By: wenlei, wmi
Differential Revision: https://reviews.llvm.org/D102429
As a follow-up to D95982, this patch continues unblocking optimizations that are blocked by pseudu probe instrumention.
The optimizations unblocked are:
- In-block load propagation.
- In-block dead store elimination
- Memory copy optimization that turns stores to consecutive memories into a memset.
These optimizations are local to a block, so they shouldn't affect the profile quality.
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D100075
While doing speculative execution opt, it conservatively drops all insn's debug info in the merged `ThenBB`(see the loop at line 2384) including the dangling probe. The missing debug info of the dangling probe will cause the wrong inference computation.
So we should avoid dropping the debug info from pseudo probe, this change try to fix this by moving the to-be dangling probe to the merging target BB before the debug info is dropped.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D101195
Pseudo probe distribution factor is used to scale down profile samples to avoid misleading the counts inference due to the usage of "maximum" in `getBlockWeight`. For callsites, the scaling down can come from code duplication prior to the sample profile loader (prelink or postlink), or due to the indirect call promotion in sample loader inliner. This patch fixes an issue in sample loader ICP where the leftover indirect callsite scaling down causes the loss of non-promoted call target samples unexpectedly. While the scaling down is to favor BFI/BPI with accurate an callsite count, it doesn't fit in the current distribution factor that represents code duplication changes. Ideally, we would need two factors, one is for code duplication, the other is for ICP. However this seems over complicated. I'm going to trade one usage (callsite counts) for the other (call target counts).
Seeing perf win on one benchmark (mcf) of SPEC2017 with others unchanged.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D100993
Pseudo probe are currently given a slot index like other regular instructions. This affects register pressure and lifetime weight computation because of enlarged lifetime length with pseudo probe instructions. As a consequence, program could get different code generated w/ and w/o pseudo probes. I'm closing the gap by excluding pseudo probes from stack index and downstream register allocation related passes.
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D100334
As a follow-up to D99815, this patch enables the test by using a DAG order instead of a sequential order to mitigate the platform portability issue due to `std:: _Hash_bytes`.
Reviewed By: wenlei, jsji
Differential Revision: https://reviews.llvm.org/D100351
Pseudo probes, when scattered in a block, can be chained dependencies of other regular DAG nodes and block DAG combine optimizations. To fix this, scattered probes in a block are grouped and placed at the beginning of the block. This shouldn't affect the profile quality.
Test Plan:
Reviewed By: wenlei, wmi
Differential Revision: https://reviews.llvm.org/D100002
We see a regression related to low probe factor(0.01) which prevents some callsites being promoted in ICPPass and later cause the missing inline in CGSCC inliner. The root cause is due to redundant(the second) multiplication of the probe factor and this change try to fix it.
`Sum` does multiply a factor right after findCallSamples but later when using as the parameter in setProbeDistributionFactor, it multiplies one again.
This change could get ~2% perf back on mcf benchmark. In mcf, previously the corresponding factor is 1 and it's the recent feature introducing the <1 factor then trigger this bug.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D99787
The case start to fail since https://reviews.llvm.org/D99351.
Looks like to me that the node order within Context Profile Tree depends
on the implmementation of std::hash<std::string>.
Unfortunately, the current clang implementation generate different values on
AIX (or for all big-endian systems?)
On Linux:
main: 2408804140(0x8f936f2c)
external: 896680882(0x357243b2)
externalA: 620231129(0x24f7f9d9)
On AIX:
main: 994322777(0x3b442959)
external: 3548191215(0xd37d19ef)
externalA: 1390365101(0x52df49ad)
XFAIL it first while we discuss and seek for a fix.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D99815
Currently prof metadata with branch counts is added only for BranchInst and SwitchInst, but not for IndirectBrInst. As a result, BPI/BFI make incorrect inferences for indirect branches, which can be very hot.
This diff adds metadata for IndirectBrInst, in addition to BranchInst and SwitchInst.
Reviewed By: wmi, wenlei
Differential Revision: https://reviews.llvm.org/D99550
Use profiled call edges to augment the top-down order. There are cases that the top-down order computed based on the static call graph doesn't reflect real execution order. For example:
1. Incomplete static call graph due to unknown indirect call targets. Adjusting the order by considering indirect call edges from the profile can enable the inlining of indirect call targets by allowing the caller processed before them.
2. Mutual call edges in an SCC. The static processing order computed for an SCC may not reflect the call contexts in the context-sensitive profile, thus may cause potential inlining to be overlooked. The function order in one SCC is being adjusted to a top-down order based on the profile to favor more inlining.
3. Transitive indirect call edges due to inlining. When a callee function is inlined into into a caller function in LTO prelink, every call edge originated from the callee will be transferred to the caller. If any of the transferred edges is indirect, the original profiled indirect edge, even if considered, would not enforce a top-down order from the caller to the potential indirect call target in LTO postlink since the inlined callee is gone from the static call graph.
4. #3 can happen even for direct call targets, due to functions defined in header files. Header functions, when included into source files, are defined multiple times but only one definition survives due to ODR. Therefore, the LTO prelink inlining done on those dropped definitions can be useless based on a local file scope. More importantly, the inlinee, once fully inlined to a to-be-dropped inliner, will have no profile to consume when its outlined version is compiled. This can lead to a profile-less prelink compilation for the outlined version of the inlinee function which may be called from external modules. while this isn't easy to fix, we rely on the postlink AutoFDO pipeline to optimize the inlinee. Since the survived copy of the inliner (defined in headers) can be inlined in its local scope in prelink, it may not exist in the merged IR in postlink, and we'll need the profiled call edges to enforce a top-down order for the rest of the functions.
Considering those cases, a profiled call graph completely independent of the static call graph is constructed based on profile data, where function objects are not even needed to handle case #3 and case 4.
I'm seeing an average 0.4% perf win out of SPEC2017. For certain benchmark such as Xalanbmk and GCC, the win is bigger, above 2%.
The change is an enhancement to https://reviews.llvm.org/D95988.
Reviewed By: wmi, wenlei
Differential Revision: https://reviews.llvm.org/D99351
during profile update.
When we inline a function and update the profile, the value profiles of the
indirect call in the inliner and inlinee will be scaled. In
https://reviews.llvm.org/D96806 and https://reviews.llvm.org/D97350, we start
using the magic number NOMORE_ICP_MAGICNUM (-1) to mark targets which have
been promoted. The magic number shouldn't be scaled during the profile update.
Although the problem has been suppressed by https://reviews.llvm.org/D98187
for SampleFDO, which stops profile update for inlining in sampleFDO, the patch
is still wanted since it will be more consistent to handle the magic number
properly in profile update.
Differential Revision: https://reviews.llvm.org/D99394
value profile annotated after inlining.
In https://reviews.llvm.org/D96806 and https://reviews.llvm.org/D97350, we
use the magic number -1 in the value profile to avoid repeated indirect call
promotion to the same target for an indirect call. Function updateIDTMetaData
is used to mark an target as being promoted in the value profile with the
magic number. updateIDTMetaData is also used to update the value profile
when an indirect call is inlined and new inline instance profile should be
applied. For the second case, currently updateIDTMetaData mixes up the
existing value profile of the indirect call with the new profile, leading
to the problematic senario that a target count is larger than the total count
in the value profile.
The patch fixes the problem. When updateIDTMetaData is used to update the
value profile after inlining, all the values in the existing value profile
will be dropped except the values with the magic number counts.
Differential Revision: https://reviews.llvm.org/D98835
For ThinLTO's prelink compilation, we need to put external inline candidates into an import list attached to function's entry count metadata. This enables ThinLink to treat such cross module callee as hot in summary index, and later helps postlink to import them for profile guided cross module inlining.
For AutoFDO, the import list is retrieved by traversing the nested inlinee functions. For CSSPGO, since profile is flatterned, a few things need to happen for it to work:
- When loading input profile in extended binary format, we need to load all child context profile whose parent is in current module, so context trie for current module includes potential cross module inlinee.
- In order to make the above happen, we need to know whether input profile is CSSPGO profile before start reading function profile, hence a flag for profile summary section is added.
- When searching for cross module inline candidate, we need to walk through the context trie instead of nested inlinee profile (callsite sample of AutoFDO profile).
- Now that we have more accurate counts with CSSPGO, we swtiched to use entry count instead of total count to decided if an external callee is potentially beneficial to inline. This make it consistent with how we determine whether call tagert is potential inline candidate.
Differential Revision: https://reviews.llvm.org/D98590
now -funique-internal-linkage-name flag is available, and we want to flip
it on by default since it is beneficial to have separate sample profiles
for different internal symbols with the same name. As a preparation, we
want to avoid regression caused by the flip.
When we flip -funique-internal-linkage-name on, the profile is collected
from binary built without -funique-internal-linkage-name so it has no uniq
suffix, but the IR in the optimized build contains the suffix. This kind of
mismatch may introduce transient regression.
To avoid such mismatch, we introduce a NameTable section flag indicating
whether there is any name in the profile containing uniq suffix. Compiler
will decide whether to keep uniq suffix during name canonicalization
depending on the NameTable section flag. The flag is only available for
extbinary format. For other formats, by default compiler will keep uniq
suffix so they will only experience transient regression when
-funique-internal-linkage-name is just flipped.
Another type of regression is caused by places where we miss to call
getCanonicalFnName. Those places are fixed.
Differential Revision: https://reviews.llvm.org/D96932
sample loader pass.
In https://reviews.llvm.org/rG5fb65c02ca5e91e7e1a00e0efdb8edc899f3e4b9,
to prevent repeated indirect call promotion for the same indirect call
and the same target, we used zero-count value profile to indicate an
indirect call has been promoted for a certain target. We removed
PromotedInsns cache in the same patch. However, there was a problem in
that patch described below, and that problem led me to add PromotedInsns
back as a mitigation in
https://reviews.llvm.org/rG4ffad1fb489f691825d6c7d78e1626de142f26cf.
When we get value profile from metadata by calling getValueProfDataFromInst,
we need to specify the maximum possible number of values we expect to read.
We uses MaxNumPromotions in the last patch so the maximum number of value
information extracted from metadata is MaxNumPromotions. If we have many
values including zero-count values when we write the metadata, some of them
will be dropped when we read them because we only read MaxNumPromotions
values. It will allow repeated indirect call promotion again. We need to
make sure if there are values indicating promoted targets, those values need
to be saved in metadata with higher priority than other values.
The patch fixed that problem. We change to use -1 to represent the count
of a promoted target instead of 0 so it is easier to sort the values.
When we prepare to update the metadata in updateIDTMetaData, we will sort
the values in the descending count order and extract only MaxNumPromotions
values to write into metadata. Since -1 is the max uint64_t number, if we
have equal to or less than MaxNumPromotions of -1 count values, they will
all be kept in metadata. If we have more than MaxNumPromotions of -1 count
values, we will only save MaxNumPromotions such values maximally. In such
case, we have logic in place in doesHistoryAllowICP to guarantee no more
promotion in sample loader pass will happen for the indirect call, because
it has been promoted enough.
With this change, now we can remove PromotedInsns without problem.
Differential Revision: https://reviews.llvm.org/D97350
Same dangling probes are redundant since they all have the same semantic that is to rely on the counts inference tool to get reasonable count for the same original block. Therefore, there's no need to keep multiple copies of them. I've seen jump threading created tons of redundant dangling probes that slowed down the compiler dramatically. Other optimization passes can also result in redundant probes though without an observed impact so far.
This change removes block-wise redundant dangling probes specifically introduced by jump threading. To support removing redundant dangling probes caused by all other passes, a final function-wise deduplication is also added.
An 18% size win of the .pseudo_probe section was seen for SPEC2017. No performance difference was observed.
Differential Revision: https://reviews.llvm.org/D97482
This change fixes a couple places where the pseudo probe intrinsic blocks optimizations because they are not naturally removable. To unblock those optimizations, the blocking pseudo probes are moved out of the original blocks and tagged dangling, instead of allowing pseudo probes to be literally removed. The reason is that when the original block is removed, we won't be able to sample it. Instead of assigning it a zero weight, moving all its pseudo probes into another block and marking them dangling should allow the counts inference a chance to assign them a more reasonable weight. We have not seen counts quality degradation from our experiments.
The optimizations being unblocked are:
1. Removing conditional probes for if-converted branches. Conditional probes are tagged dangling when their homing branch arms are folded so that they will not be over-counted.
2. Unblocking jump threading from removing empty blocks. Pseudo probe prevents jump threading from removing logically empty blocks that only has one unconditional jump instructions.
3. Unblocking SimplifyCFG and MIR tail duplicate to thread empty blocks and blocks with redundant branch checks.
Since dangling probes are logically deleted, they should not consume any samples in LTO postLink. This can be achieved by setting their distribution factors to zero when dangled.
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D97481
Dangling probes are the probes associated to an empty block. This usually happens when all real instructions are optimized away from the block. There is a problem with dangling probes during the offline counts processing. The way the sample profiler works is that samples collected on the first physical instruction following a probe will be counted towards the probe. This logically equals to treating the instruction next to a probe as if it is from the same block of the probe. In the dangling probe case, the real instruction following a dangling probe actually starts a new block, and samples collected on the new block may cause issues when counted towards the empty block.
To mitigate this issue, we first try to move around a dangling probe inside its owning block. If there are still native instructions preceding the probe in the same block, we can then use them as a place holder to collect samples for the probe. A pass is added to walk each block backwards looking for probes not followed by any real instruction and moving them before the first real instruction. This is done right before the object emission.
If we are unlucky to find such in-block preceding instructions for a probe, the solution we are taking is to tag such probe as dangling so that the samples reported for them will not be trusted by the compiler. We leave it up to the counts inference algorithm to get such probes a reasonable count. The number `UINT64_MAX` is used to mark sample count as collected for a dangling probe.
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D95962
into profile symbol list.
When test is unrepresentative to production behavior, sample profile
collected from production can cause unexpected performance behavior
in test. To triage such issue, it is useful to have a cutoff flag
to control how many symbols will be included into profile symbol list
in order to do binary search.
Differential Revision: https://reviews.llvm.org/D97623
Found a problem in indirect call promotion in sample loader pass. Currently
if an indirect call is promoted for a target, and if the parent function is
inlined into some other function, the indirect call can be promoted for the
same target again. That is redundent which can harm performance and can cause
excessive compile time in some extreme case.
The patch fixes the issue. If a target is promoted for an indirect call, the
patch will write ICP metadata with the target call count being set to 0.
In the later ICP in sample profile loader, if it sees a target has 0 count
for an indirect call, it knows the target has been promoted and won't do
indirect call promotion for the indirect call.
The fix brings 0.1~0.2% performance on our search benchmark.
Differential Revision: https://reviews.llvm.org/D96806
Summary:
The negative test (with the feature being added disabled) caused MSAN failure and that's the added feature is supposed to fix. Therefore the negative test code is being removed.
Functions are currently processed by the sample profiler loader in a top-down order defined by the static call graph. The order is being adjusted to be a top-down order based on the input context-sensitive profile. One benefit is that the processing order of caller and callee in one SCC would follow the context order in the profile to favor more inlining. Another benefit is that the processing order of caller and callee through an indirect call (which is not on the static call graph) can be honored which in turn allows for more inlining.
The profile top-down order for SCC is also extended to support non-CS profiles.
Two switches `-mllvm -use-profile-indirect-call-edges` and `-mllvm -use-profile-top-down-order` are being introduced.
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D95988
The IR/MIR pseudo probe intrinsics don't get materialized into real machine instructions and therefore they don't incur runtime cost directly. However, they come with indirect cost by blocking certain optimizations. Some of the blocking are intentional (such as blocking code merge) for better counts quality while the others are accidental. This change unblocks perf-critical optimizations that do not affect counts quality. They include:
1. IR InstCombine, sinking load operation to shorten lifetimes.
2. MIR LiveRangeShrink, similar to #1
3. MIR TwoAddressInstructionPass, i.e, opeq transform
4. MIR function argument copy elision
5. IR stack protection. (though not perf-critical but nice to have).
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D95982
Context-sensitive profile effectively split a function profile into many copies each representing the CFG profile of a particular calling context. That makes the count distribution looks more flat as we now have more function profiles each with lower counts, which in turn leads to lower hot thresholds. Now we tells threshold computation to merge context profile first before calculating percentile based cutoffs to compensate for seemingly flat context profile. This can be controlled by swtich `sample-profile-contextless-threshold`.
Earlier measurement showed ~0.4% perf boost with this tuning on spec2k6 for CSSPGO (with pseudo-probe and new inliner).
Differential Revision: https://reviews.llvm.org/D95980
Sample re-annotation is required in LTO time to achieve a reasonable post-inline profile quality. However, we have seen that such LTO-time re-annotation degrades profile quality. This is mainly caused by preLTO code duplication that is done by passes such as loop unrolling, jump threading, indirect call promotion etc, where samples corresponding to a source location are aggregated multiple times due to the duplicates. In this change we are introducing a concept of distribution factor for pseudo probes so that samples can be distributed for duplicated probes scaled by a factor. We hope that optimizations duplicating code well-maintain the branch frequency information (BFI) based on which probe distribution factors are calculated. Distribution factors are updated at the end of preLTO pipeline to reflect an estimated portion of the real execution count.
This change also introduces a pseudo probe verifier that can be run after each IR passes to detect duplicated pseudo probes.
A saturated distribution factor stands for 1.0. A pesudo probe will carry a factor with the value ranged from 0.0 to 1.0. A 64-bit integral distribution factor field that represents [0.0, 1.0] is associated to each block probe. Unfortunately this cannot be done for callsite probes due to the size limitation of a 32-bit Dwarf discriminator. A 7-bit distribution factor is used instead.
Changes are also needed to the sample profile inliner to deal with prorated callsite counts. Call sites duplicated by PreLTO passes, when later on inlined in LTO time, should have the callees’s probe prorated based on the Prelink-computed distribution factors. The distribution factors should also be taken into account when computing hotness for inline candidates. Also, Indirect call promotion results in multiple callisites. The original samples should be distributed across them. This is fixed by adjusting the callisites' distribution factors.
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D93264
Refactoring SampleProfileLoader::inlineHotFunctions to use helpers from CSSPGO inlining and reduce similar code in the inlining loop, plus minor cleanup for AFDO path.
This is resubmit of D95024, with build break and overtighten assertion fixed.
Test Plan:
Refactoring SampleProfileLoader::inlineHotFunctions to use helpers from CSSPGO inlining and reduce similar code in the inlining loop, plus minor cleanup for AFDO path.
Test Plan:
Differential Revision: https://reviews.llvm.org/D95024
This change implemented call site prioritized BFS profile guided inlining for sample profile loader. The new inlining strategy maximize the benefit of context-sensitive profile as mentioned in the follow up discussion of CSSPGO RFC. The change will not affect today's AutoFDO as it's opt-in. CSSPGO now defaults to the new FDO inliner, but can fall back to today's replay inliner using a switch (`-sample-profile-prioritized-inline=0`).
Motivation
With baseline AutoFDO, the inliner in sample profile loader only replays previous inlining, and the use of profile is only for pruning previous inlining that turned out to be cold. Due to the nature of replay, the FDO inliner is simple with hotness being the only decision factor. It has the following limitations that we're improving now for CSSPGO.
- It doesn't take inline candidate size into account. Since it's doing replay, the size growth is bounded by previous CGSCC inlining. With context-sensitive profile, FDO inliner is no longer limited by previous inlining, so we need to take size into account to avoid significant size bloat.
- The way it looks at hotness is not accurate. It uses total samples in an inlinee as proxy for hotness, while what really matters for an inline decision is the call site count. This is an unfortunate fall back because call site count and callee entry count are not reliable due to dwarf based correlation, especially for inlinees. Now paired with pseudo-probe, we have accurate call site count and callee's entry count, so we can use that to gauge hotness more accurately.
- It treats all call sites from a block as hot as long as there's one call site considered hot. This is normally true, but since total samples is used as hotness proxy, this transitiveness within block magnifies the inacurate hotness heuristic. With pseduo-probe and the change above, this is no longer an issue for CSSPGO.
New FDO Inliner
Putting all the requirement for CSSPGO together, we need a top-down call site prioritized BFS inliner. Here're reasons why each component is needed.
- Top-down: We need a top-down inliner to better leverage context-sensitive profile, so inlining is driven by accurate context profile, and post-inline is also accurate. This is already implemented in https://reviews.llvm.org/D70655.
- Size Cap: For top-down inliner, taking function size into account for inline decision alone isn't sufficient to control size growth. We also need to explicitly cap size growth because with top-down inlining, we can grow inliner size significantly with large number of smaller inlinees even if each individually passes the cost/size check.
- Prioritize call sites: With size cap, inlining order also becomes important, because if we stop inlining due to size budget limit, we'd want to use budget towards the most beneficial call sites.
- BFS inline: Same as call site prioritization, if we stop inlining due to size budget limit, we want a balanced inline tree, rather than going deep on one call path.
Note that the new inliner avoids repeatedly evaluating same set of call site, so it should help with compile time too. For this reason, we could transition today's FDO inliner to use a queue with equal priority to avoid wasted reevaluation of same call site (TODO).
Speculative indirect call promotion and inlining is also supported now with CSSPGO just like baseline AutoFDO.
Tunings and knobs
I created tuning knobs for size growth/cap control, and for hot threshold separate from CGSCC inliner. The default values are selected based on initial tuning with CSSPGO.
Results
Evaluated with an internal LLVM fork couple months ago, plus another change to adjust hot-threshold cutoff for context profile (will send up after this one), the new inliner show ~1% geomean perf win on spec2006 with CSSPGO, while reducing code size too. The measurement was done using train-train setup, MonoLTO w/ new pass manager and pseudo-probe. Note that this is just a starting point - we hope that the new inliner will open up more opportunity with CSSPGO, but it will certainly take more time and effort to make it fully calibrated and ready for bigger workloads (we're working on it).
Differential Revision: https://reviews.llvm.org/D94001
Fixing up a couple places where `getCallSiteIdentifier` is needed to support pseudo-probe-based callsites.
Also fixing an issue in the extbinary profile reader where the metadata section is not fully scanned based on the number of profiles loaded only for the current module.
Reviewed By: wmi, wenlei
Differential Revision: https://reviews.llvm.org/D95791
This change brings up support of context-sensitive profiles in the format of extended binary. Existing sample profile reader/writer/merger code is being tweaked to reflect the fact of bracketed input contexts, like (`[...]`). The paired brackets are also needed in extbinary profiles because we don't yet have an otherwise good way to tell calling contexts apart from regular function names since the context delimiter `@` can somehow serve as a part of the C++ mangled names.
Reviewed By: wmi, wenlei
Differential Revision: https://reviews.llvm.org/D95547
turning off SampleFDO silently.
Currently sample loader pass turns off SampleFDO optimization silently when
it sees error in reading the profile. This behavior will defeat the tests
which could have caught those bad/incompatible profile problems. This patch
change the behavior to report error.
Differential Revision: https://reviews.llvm.org/D95269
separate sections.
For ThinLTO, all the function profiles without context has been annotated to
outline functions if possible in prelink phase. In postlink phase, profile
annotation in postlink phase is only meaningful for function profile with
context. If the profile is large, it is better to split the profile into two
parts, one with context and one without, so the profile reading in postlink
phase only has to read the part with context. To have the profile splitting,
we extend the ExtBinary format to support different section arrangement. It
will be flexible to add other section layout in the future without the need
to create new class inheriting from ExtBinary class.
Differential Revision: https://reviews.llvm.org/D94435
This change modifies the source location formatting from:
LineNumber.Discriminator
to:
LineNumber:ColumnNumber.Discriminator
The motivation here is to enhance location information for inline replay that currently exists for the SampleProfile inliner. This will be leveraged further in inline replay for the CGSCC inliner in the related diff.
The ReplayInlineAdvisor is also modified to read the new format and now takes into account the callee for greater accuracy.
Testing:
ninja check-llvm
Reviewed By: mtrofin
Differential Revision: https://reviews.llvm.org/D94333
This change enables pseudo-probe-based sample counts to be consumed by the sample profile loader under the regular `-fprofile-sample-use` switch with minimal adjustments to the existing sample file formats. After the counts are imported, a probe helper, aka, a `PseudoProbeManager` object, is automatically launched to verify the CFG checksum of every function in the current compilation against the corresponding checksum from the profile. Mismatched checksums will cause a function profile to be slipped. A `SampleProfileProber` pass is scheduled before any of the `SampleProfileLoader` instances so that the CFG checksums as well as probe mappings are available during the profile loading time. The `PseudoProbeManager` object is set up right after the profile reading is done. In the future a CFG-based fuzzy matching could be done in `PseudoProbeManager`.
Samples will be applied only to pseudo probe instructions as well as probed callsites once the checksum verification goes through. Those instructions are processed in the same way that regular instructions would be processed in the line-number-based scenario. In other words, a function is processed in a regular way as if it was reduced to just containing pseudo probes (block probes and callsites).
**Adjustment to profile format **
A CFG checksum field is being added to the existing AutoFDO profile formats. So far only the text format and the extended binary format are supported. For the text format, a new line like
```
!CFGChecksum: 12345
```
is added to the end of the body sample lines. For the extended binary profile format, we introduce a metadata section to store the checksum map from function names to their CFG checksums.
Differential Revision: https://reviews.llvm.org/D92347
This change implements pseudo probe encoding and emission for CSSPGO. Please see RFC here for more context: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s
Pseudo probes are in the form of intrinsic calls on IR/MIR but they do not turn into any machine instructions. Instead they are emitted into the binary as a piece of data in standalone sections. The probe-specific sections are not needed to be loaded into memory at execution time, thus they do not incur a runtime overhead.
**ELF object emission**
The binary data to emit are organized as two ELF sections, i.e, the `.pseudo_probe_desc` section and the `.pseudo_probe` section. The `.pseudo_probe_desc` section stores a function descriptor for each function and the `.pseudo_probe` section stores the actual probes, each fo which corresponds to an IR basic block or an IR function callsite. A function descriptor is stored as a module-level metadata during the compilation and is serialized into the object file during object emission.
Both the probe descriptors and pseudo probes can be emitted into a separate ELF section per function to leverage the linker for deduplication. A `.pseudo_probe` section shares the same COMDAT group with the function code so that when the function is dead, the probes are dead and disposed too. On the contrary, a `.pseudo_probe_desc` section has its own COMDAT group. This is because even if a function is dead, its probes may be inlined into other functions and its descriptor is still needed by the profile generation tool.
The format of `.pseudo_probe_desc` section looks like:
```
.section .pseudo_probe_desc,"",@progbits
.quad 6309742469962978389 // Func GUID
.quad 4294967295 // Func Hash
.byte 9 // Length of func name
.ascii "_Z5funcAi" // Func name
.quad 7102633082150537521
.quad 138828622701
.byte 12
.ascii "_Z8funcLeafi"
.quad 446061515086924981
.quad 4294967295
.byte 9
.ascii "_Z5funcBi"
.quad -2016976694713209516
.quad 72617220756
.byte 7
.ascii "_Z3fibi"
```
For each `.pseudoprobe` section, the encoded binary data consists of a single function record corresponding to an outlined function (i.e, a function with a code entry in the `.text` section). A function record has the following format :
```
FUNCTION BODY (one for each outlined function present in the text section)
GUID (uint64)
GUID of the function
NPROBES (ULEB128)
Number of probes originating from this function.
NUM_INLINED_FUNCTIONS (ULEB128)
Number of callees inlined into this function, aka number of
first-level inlinees
PROBE RECORDS
A list of NPROBES entries. Each entry contains:
INDEX (ULEB128)
TYPE (uint4)
0 - block probe, 1 - indirect call, 2 - direct call
ATTRIBUTE (uint3)
reserved
ADDRESS_TYPE (uint1)
0 - code address, 1 - address delta
CODE_ADDRESS (uint64 or ULEB128)
code address or address delta, depending on ADDRESS_TYPE
INLINED FUNCTION RECORDS
A list of NUM_INLINED_FUNCTIONS entries describing each of the inlined
callees. Each record contains:
INLINE SITE
GUID of the inlinee (uint64)
ID of the callsite probe (ULEB128)
FUNCTION BODY
A FUNCTION BODY entry describing the inlined function.
```
To support building a context-sensitive profile, probes from inlinees are grouped by their inline contexts. An inline context is logically a call path through which a callee function lands in a caller function. The probe emitter builds an inline tree based on the debug metadata for each outlined function in the form of a trie tree. A tree root is the outlined function. Each tree edge stands for a callsite where inlining happens. Pseudo probes originating from an inlinee function are stored in a tree node and the tree path starting from the root all the way down to the tree node is the inline context of the probes. The emission happens on the whole tree top-down recursively. Probes of a tree node will be emitted altogether with their direct parent edge. Since a pseudo probe corresponds to a real code address, for size savings, the address is encoded as a delta from the previous probe except for the first probe. Variant-sized integer encoding, aka LEB128, is used for address delta and probe index.
**Assembling**
Pseudo probes can be printed as assembly directives alternatively. This allows for good assembly code readability and also provides a view of how optimizations and pseudo probes affect each other, especially helpful for diff time assembly analysis.
A pseudo probe directive has the following operands in order: function GUID, probe index, probe type, probe attributes and inline context. The directive is generated by the compiler and can be parsed by the assembler to form an encoded `.pseudoprobe` section in the object file.
A example assembly looks like:
```
foo2: # @foo2
# %bb.0: # %bb0
pushq %rax
testl %edi, %edi
.pseudoprobe 837061429793323041 1 0 0
je .LBB1_1
# %bb.2: # %bb2
.pseudoprobe 837061429793323041 6 2 0
callq foo
.pseudoprobe 837061429793323041 3 0 0
.pseudoprobe 837061429793323041 4 0 0
popq %rax
retq
.LBB1_1: # %bb1
.pseudoprobe 837061429793323041 5 1 0
callq *%rsi
.pseudoprobe 837061429793323041 2 0 0
.pseudoprobe 837061429793323041 4 0 0
popq %rax
retq
# -- End function
.section .pseudo_probe_desc,"",@progbits
.quad 6699318081062747564
.quad 72617220756
.byte 3
.ascii "foo"
.quad 837061429793323041
.quad 281547593931412
.byte 4
.ascii "foo2"
```
With inlining turned on, the assembly may look different around %bb2 with an inlined probe:
```
# %bb.2: # %bb2
.pseudoprobe 837061429793323041 3 0
.pseudoprobe 6699318081062747564 1 0 @ 837061429793323041:6
.pseudoprobe 837061429793323041 4 0
popq %rax
retq
```
**Disassembling**
We have a disassembling tool (llvm-profgen) that can display disassembly alongside with pseudo probes. So far it only supports ELF executable file.
An example disassembly looks like:
```
00000000002011a0 <foo2>:
2011a0: 50 push rax
2011a1: 85 ff test edi,edi
[Probe]: FUNC: foo2 Index: 1 Type: Block
2011a3: 74 02 je 2011a7 <foo2+0x7>
[Probe]: FUNC: foo2 Index: 3 Type: Block
[Probe]: FUNC: foo2 Index: 4 Type: Block
[Probe]: FUNC: foo Index: 1 Type: Block Inlined: @ foo2:6
2011a5: 58 pop rax
2011a6: c3 ret
[Probe]: FUNC: foo2 Index: 2 Type: Block
2011a7: bf 01 00 00 00 mov edi,0x1
[Probe]: FUNC: foo2 Index: 5 Type: IndirectCall
2011ac: ff d6 call rsi
[Probe]: FUNC: foo2 Index: 4 Type: Block
2011ae: 58 pop rax
2011af: c3 ret
```
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D91878
