When the input to ptest_first is a vector concat and the mask is all active,
performPTestFirstCombine returns a ptest_first using the first operand
of the concat, looking through any reinterpret casts.
This allows optimizePTestInstr to later remove the ptest when the first
operand is a flag setting instruction such as whilelo.
This patch adds direct code-gen support for a faster MOD intrinsic for
REAL types. Flang has maintained and keeps maintaining a high-precision
implementation of the MOD intrinsic as part of the Fortran runtime. With
the -ffast-real-mod flag, users can opt to avoid calling into the
Fortran runtime, but instead trigger code-gen that produces faster code
by avoiding the runtime call, at the expense of potentially risking bit
cancelation by having the compiler use the MOD formula a specified by
ISO Fortran.
Also removes the command line option to control this feature.
There seem to be mainly two kinds of test changes:
- Some operands of addition instructions are swapped; that is to be expected
since PTRADD is not commutative.
- Improvements in code generation, probably because the legacy lowering enabled
some transformations that were sometimes harmful.
For SWDEV-516125.
`lookupRuntimeDefinition` assumed that a process would handle only one
TU. This is not true for unit tests, for instance. Multiple snippets of
code get parsed, and their AST are unloaded each time.
Since the cache relies on pointers as keys, if the same address happens
to be reused between runs, the cache would return a stale pointer,
potentially causing a segmentation fault. This is not that unlikely if
the snippets are similar, which would trigger similar allocation
patterns.
CPP-4889
Follow up on #161347 to allow scalar fir.box/class reconstruction (at
least required for polymorphic types).
The assert in genDimInfoFromBox was rejecting scalars while there is no
functional reason for that (only assumed-rank are an issue there).
MemorySanitizer currently does a lot of pointer arithmetic using
ptrtoint+add+inttoptr instead of using getelementptr. As far as I can
tell, there is no need to use this pattern -- msan is not trying to
synthesize pointers with different provenance here. The pointers in
question stay within one object (like the TLS parameter area).
I suspect that this is just a leftover from pre-opaque-pointer types
where this was a natural way to perform offset arithmetic. Nowadays we
should just emit a getelementptr i8, aka ptradd.
### Summary
Stabilize ASan wchar tests across Darwin and Android. NFC: test-only.
Follow-up to PR #160493 (adds wchar interceptors/tests).
### Motivation
- Darwin: The top frame often resolves to `libclang_rt.asan_*` rather
than a source file, so strict checks that include file/line can fail.
See Chromium issue
[448631142](https://g-issues.chromium.org/issues/448631142).
- Android: The “ERROR:” header can go to logcat instead of stderr, so
FileCheck may not see it; stdout/stderr reordering also makes pre-crash
markers racy. See Android Buildbot
[186/12821](https://lab.llvm.org/buildbot/#/builders/186/builds/12821).
### Changes
- Android:
- Force reports to stderr via `%env_asan_opts=log_to_stderr=1`, avoiding
the “ERROR:” header going to logcat.
- Print the pre-crash “Good so far.” to stderr and `fflush(stderr)` to
avoid stdout/stderr reordering.
- Darwin:
- Relax the stack-frame check to only require the function name
(`wcscpy/wcsncpy/wcscat/wcsncat`) to tolerate `libclang_rt.asan_*`
frames.
- Common:
- Reuse FileCheck var `[[ADDR]]` instead of redefining.
- Make wide string literals `const wchar_t*` to silence
`-Wwritable-strings`.
### Risk
- NFC: test-only; no change to runtime behavior.
### References
- Follow-up to PR #160493.
- Chromium: [448631142](https://g-issues.chromium.org/issues/448631142)
(Darwin failures).
- Android Buildbot:
[186/12821](https://lab.llvm.org/buildbot/#/builders/186/builds/12821).
Signed-off-by: Yixuan Cao <caoyixuan2019@email.szu.edu.cn>
Both Usman Nadeem and I have constantly contributed to the
DFAJumpThreading pass so far. To push DFAJumpThreading forwards and make
it enabled by default, I volunteer myself and Usman Nadeem as
DFAJumpThreading maintainers.
Previously if we had a subregister extract reading from a
full copy, the no-subregister incoming copy would overwrite
the DefSubReg index of the folding context.
There's one ugly rvv regression, but it's a downstream
issue of this; an unnecessary same class reg-to-reg full copy
was avoided.
This matches what we do for regular i8 extload due to the lack of c.lb
in Zbc.
This only affects global isel because SelectionDAG won't create an
anyext i8 atomic_load today.
R_AARCH64_TLSDESC_CALL is a relocation emitted as a hint for a linker to
replace `blr r` instruction with nop. BOLT does not currently require
any special handling for it.
Note that previously existing extraction of the relocated value was
incorrect.
It's unnecessary to build the whole symtable, and on top of everything,
un-optimal to do so for every function. All we really need is the
instrumented PGO name - considering also LTO-ness - and then we can
compute the function name.
Fast strlen implementations (naive wide-reads, SIMD-based, and
x86_64/aarch64-optimized versions) all may perform
technically-out-of-bound reads, which leads to reports under ASan,
HWASan (on ARM machines), and also TSan (which also has the capability
to detect heap out-of-bound reads). So, we need to explicitly disable
instrumentation in all three cases.
Tragically, Clang didn't support `[[gnu::no_sanitize]]` syntax until
recently, and since we're supporting both GCC and Clang, we have to
revert to `__attribute__` syntax.
Refactored IR2Vec vocabulary and introduced IR (semantics) agnostic `VocabStorage`
- `Vocabulary` *has-a* `VocabStorage`
- `Vocabulary` deals with LLVM IR specific entities. This would help in efficient reuse of parts of the logic for MIR.
- Storage uses a section-based approach instead of a flat vector, improving organization and access patterns.
Previously, we only used `objcopy`, which is not available for some
build configurations. With this patch, we not only try to use `objcopy`,
but also try to use `llvm-objcopy` if available.
This is a follow-up of https://github.com/llvm/llvm-project/pull/156383.
Comparison predicates (equal, not equal, greater than, etc.) provide important semantic information about program behavior. Previously, IR2Vec only captured that a comparison was happening but not what kind of comparison it was. This PR extends the IR2Vec vocabulary to include comparison predicates (ICmp and FCmp) as part of the embedding space.
Following are the changes:
1. Expand the vocabulary slot layout to include predicate entries after opcodes, types, and operands
2. Add methods to handle predicate embedding lookups and conversions
3. Update the embedder implementations to include predicate information when processing CmpInst instructions
4. Update test files to include the new predicate entries in the vocabulary
(Tracking issues: #141817, #141833)
This test failed during testing on the RISC-V target because we couldn't
strip the main label from the binary. main is dynamically linked when
the -fPIC flag is enabled. The RISC-V ABI requires that executables
support loading at arbitrary addresses to enable shared libraries and
secure loading (ASLR). In PIC mode, function addresses cannot be
hardcoded in the code. Instead, code is generated to load addresses from
the GOT/PLT tables, which are initialized by the dynamic loader. The
reference to main thus ends up in .dynsym and is dynamically bound. We
cannot strip main or any other dynamically linked functions because
these functions are referenced indirectly via dynamic linking tables
(.plt and .got). Removing these symbols would break the dynamic linking
mechanism needed to resolve function addresses at runtime, causing the
executable to fail to correctly call them.
There's a pattern throughout LLVM of cl::opts being exported. That in
itself is probably a bit unfortunate, but what's especially bad about it
is that a lot of those symbols are in the global namespace. Move them
into the llvm namespace.
While doing this, I noticed some other variables in the global namespace
and moved them as well.
Build on Clang-BOLT infrastructure to collect sample profile for CSSPGO.
Add CSSPGO.cmake and BOLT-CSSPGO.cmake to automate CSSPGO/+BOLT
Clang builds.
Note that `CLANG_PGO_TRAINING_DATA_SOURCE_DIR` is required as built-in
training set is inadequate for collecting sampled profile.
Hardware compatibility: CSSPGO requires synchronized (0-skid) call
and branch stacks, which is only available with Intel PEBS (Sandy
Bridge+),
AMD Zen3 with BRS, Zen4 with LBRv2+LBR_PMC_FREEZE, and Zen5 with LBRv2.
This patch adds support for Intel `br_inst_retired.near_taken:uppp`
event.
Test Plan:
Added BOLT-CSSPGO.cmake with same use as BOLT-PGO.cmake,
e.g. for bootstrapped ThinLTO+CSSPGO+BOLT, with CSSPGO profile collected
from LLVM build, and BOLT profile collected from Hello World
(instrumentation):
```
cmake -B clang-csspgo-bolt -S /path/to/llvm-project/llvm \
-DLLVM_ENABLE_LLD=ON -DBOOTSTRAP_LLVM_ENABLE_LLD=ON \
-DBOOTSTRAP_BOOTSTRAP_LLVM_ENABLE_LLD=ON \
-DPGO_INSTRUMENT_LTO=Thin \
-DBOOTSTRAP_CLANG_PGO_TRAINING_DATA_SOURCE_DIR=/path/to/llvm-project/llvm \
-GNinja -C /path/to/llvm-project/clang/cmake/caches/BOLT-CSSPGO.cmake
ninja stage2-clang-bolt
...
warning: Sample PGO is estimated to optimize better with 19.5x more samples. Please consider increasing sampling rate or profiling for longer duration to get more samples.
...
[2800/2801] Optimizing Clang with BOLT
BOLT-INFO: 8189 out of 106942 functions in the binary (7.7%) have non-empty execution profile
13776393 : taken branches (-42.1%)
```
Performance testing with Clang:
- Setup: Clang-BOLT testing harness
9f2b46f67a
- CSSPGO training: building LLVM,
- InstrPGO training: building Hello World,
- BOLT training: building Hello World, instrumentation,
- benchmark: building small LLVM tool (not),
- 2S Intel SKX Xeon 6138 with 40C/80T and 256GB RAM, using 20C/40T for
build,
- Results, wall time, lower is better
- Baseline (bootstrapped build): 10.36s,
- InstrPGO + ThinLTO: 9.34s,
- CSSPGO + ThinLTO: 8.85s.
- BOLT results, for reference:
- Baseline: 9.09s,
- InstrPGO + ThinLTO: 9.09s,
- CSSPGO + ThinLTO: 8.58s.
---------
Co-authored-by: Matthias Braun <matze@braunis.de>
These includes are not used by ClangSYCLLinker.cpp directly.
Explicitly include FormatVariadic.h for formatv declaration, which was implicitly included by removed headers.
The `llvm::sys::fs::make_absolute(const Twine &, SmallVectorImpl<char>
&)` functions doesn't perform any FS access - it only modifies the
second parameter via path/string operations. This function should live
in the `llvm::sys::path` namespace for consistency and for making it
easier to spot function calls that perform IO.
The PR #160935 incorrectly replaced `llvm::sys::fs::getUniqueID()` with
`llvm::vfs::FileSystem::exists()` in a condition. That's incorrect,
since the first function returns `std::error_code` that evaluates to
`true` when there is an error (file doesn't exist), while the new code
does the opposite. This PR fixes that issue by inverting the
conditional.
Co-authored-by: ronlieb <ron.lieberman@amd.com>
