Provide backwards compatibility for YAML profile that uses `std::hash`:
xxh3 hash is the default for newly produced profile (sets `std-hash:
false`),
whereas the profile that doesn't specify `std-hash` will be treated as
`std-hash: true`, preserving old behavior.
std::hash and ADT/Hashing::hash_value are non-deterministic functions
whose
results might vary across implementation/process/execution. Using xxh3
instead
for computing hashes of BinaryFunctions and BinaryBasicBlock for stale
profile
matching.
(A possible alternative is to use ADT/StableHashing.h based on FNV
hashing but
xxh3 seems to be more popular in LLVM)
This is to address https://github.com/llvm/llvm-project/issues/65241.
Two (minor) improvements for stale matching:
- always match entry blocks to each other, even if there is a hash mismatch;
- ignore nops in (loose) hash computation.
I record a small improvement in inference quality on my benchmarks. Tests are not affected
Reviewed By: Amir
Differential Revision: https://reviews.llvm.org/D159488
Fine-tuning hash computation for stale matching:
- introducing a new "loose" basic block hash that allows to match many more blocks than before;
- tweaking params of the inference algorithm that find (slightly) better solutions;
- added more meaningful tests for stale matching.
Tested the changes on several open-source benchmarks (clang, rocksdb, chrome)
and one prod workload using different compiler modes (LTO/PGO etc). There is
always an improvement in the quality of inferred profiles.
(The current implementation is still not optimal but the diff is a step forward;
I am open to further suggestions)
Reviewed By: Amir
Differential Revision: https://reviews.llvm.org/D156278
1. Using ADT/Bitfields.h for hash computation; this is equivalent but shorter than the existing implementation
2. Getting rid of Layout indices for stale matching; using BB->getIndex for indexing
Reviewed By: Amir
Differential Revision: https://reviews.llvm.org/D155748
Adding some logs related to stale profile matching. The new data can be helpful
to understand how "stale" the input profile is and how well the inference is
able to utilize the stale data.
Example of outputs on clang-10 built with LTO (profile collected on a year-old release):
```
BOLT-INFO: inferred profile for 2101 (18.52% of profiled, 100.00% of stale) functions responsible for 30.95% samples (14754697 out of 47670654)
BOLT-INFO: stale inference matched 89.42% of basic blocks (79052 out of 88402 stale) responsible for 76.99% samples (645737 out of 838719 stale)
```
LTO+AutoFDO:
```
BOLT-INFO: inferred profile for 6146 (57.57% of profiled, 100.00% of stale) functions responsible for 90.34% samples (50891403 out of 56330313)
BOLT-INFO: stale inference matched 74.55% of basic blocks (191295 out of 256589 stale) responsible for 57.30% samples (1288632 out of 2248799 stale)
```
Reviewed By: Amir, maksfb
Differential Revision: https://reviews.llvm.org/D154737
This is a first "serious" version of stale profile matching in BOLT. This diff
extends the hash computation for basic blocks so that we can apply a fuzzy
hash-based matching. The idea is to compute several "versions" of a hash value
for a basic block. A loose version of a hash (computed by ignoring instruction
operands) allows to match blocks in functions whose content has been changed,
while stricter hash values (considering instruction opcodes with operands and
even based on hashes of block's successors/predecessors) allow to resolve
collisions. In order to save space and build time, individual hash components
are blended into a single uint64_t.
There are likely numerous ways of improving hash computation but already this
simple variant provides significant perf benefits.
**Perf testing** on the clang binary: collecting data on clang-10 and using it
to optimize clang-11 (with ~1 year of commits in between). Next, we compare
- //stale_clang// (clang-11 optimized with profile collected on clang-10 with **infer-stale-profile=0**)
- //opt_clang// (clang-11 optimized with profile collected on clang-11)
- //infer_clang// (clang-11 optimized with profile collected on clang-10 with **infer-stale-profile=1**)
`LTO-only` mode:
//stale_clang// vs //opt_clang//: task-clock [delta(%): 9.4252 ± 1.6582, p-value: 0.000002]
(That is, there is a ~9.5% perf regression)
//infer_clang// vs //opt_clang//: task-clock [delta(%): 2.1834 ± 1.8158, p-value: 0.040702]
(That is, the regression is reduced to ~2%)
Related BOLT logs:
```
BOLT-INFO: identified 2114 (18.61%) stale functions responsible for 30.96% samples
BOLT-INFO: inferred profile for 2101 (18.52% of all profiled) functions responsible for 30.95% samples
```
`LTO+AutoFDO` mode:
//stale_clang// vs //opt_clang//: task-clock [delta(%): 19.1293 ± 1.4131, p-value: 0.000002]
//infer_clang// vs //opt_clang//: task-clock [delta(%): 7.4364 ± 1.3343, p-value: 0.000002]
Related BOLT logs:
```
BOLT-INFO: identified 5452 (50.27%) stale functions responsible for 85.34% samples
BOLT-INFO: inferred profile for 5442 (50.23% of all profiled) functions responsible for 85.33% samples
```
Reviewed By: Amir
Differential Revision: https://reviews.llvm.org/D146661
BOLT often has to deal with profiles collected on binaries built from several
revisions behind release. As a result, a certain percentage of functions is
considered stale and not optimized. This diff adds an ability to match profile
to functions that are not 100% binary identical, which increases the
optimization coverage and boosts the performance of applications.
The algorithm consists of two phases: matching and inference:
- At the matching phase, we try to "guess" as many block and jump counts from
the stale profile as possible. To this end, the content of each basic block
is hashed and stored in the (yaml) profile. When BOLT optimizes a binary,
it computes block hashes and identifies the corresponding entries in the
stale profile. It yields a partial profile for every CFG in the binary.
- At the inference phase, we employ a network flow-based algorithm (profi) to
reconstruct "realistic" block and jump counts from the partial profile
generated at the first stage. In practice, we don't always produce proper
profile data but the majority (e.g., >90%) of CFGs get the correct counts.
This is a first part of the change; the next stacked diff extends the block hashing
and provides perf evaluation numbers.
Reviewed By: maksfb
Differential Revision: https://reviews.llvm.org/D144500
