This patch refactors the trap reason demangling logic in
`lldb_private::VerboseTrapFrameRecognizer::RecognizeFrame` into a new
public function `clang::CodeGen::DemangleTrapReasonInDebugInfo`.
There are two reasons for doing this:
1. In a future patch the logic for demangling needs to be used somewhere
else in LLDB and thus the logic needs refactoring to avoid duplicating
code.
2. The logic for demangling shouldn't really be in LLDB anyway because
it's a Clang implementation detail and thus the logic really belongs
inside Clang, not LLDB.
Unit tests have been added for the new function that demonstrate how to
use the new API.
The function names recognized by VerboseTrapFrameRecognizer are
identical to before. However, this patch isn't NFC because:
* The `lldbTarget` library now links against `clangCodeGen` which it
didn't previously.
* The LLDB logging output is a little different now. The previous code
tried to log failures for an invalid regex pattern and for the
`Regex::match` API not returning the correct number of matches. These
failure conditions are unreachable via unit testing so they have been
made assertions failures inside the `DemangleTrapReasonInDebugInfo`
implementation instead of trying to log them in LLDB.
rdar://163230807
As a follow-up to PR#165841, this change addresses `prof_md` metadata
loss in AtomicExpandPass when lowering `atomicrmw xchg` to a
Load-Linked/Store-Exclusive (LL/SC) loop.
This path is distinct from the LSE path addressed previously:
PR #165841 (and its tests) used `-mtriple=aarch64-linux-gnu`, which
targets a modern **ARMv8.1+** architecture. This architecture supports
**Large System Extensions (LSE)**, allowing `atomicrmw` to be lowered
directly to a more efficient hardware instruction.
This PR (and its tests) uses `-mtriple=aarch64--` or
`-mtriple=armv8-linux-gnueabihf`. This indicates an `ARMv8.0 or lower
architecture that does not support LSE`. On these targets, the pass must
fall back to synthesizing a manual LL/SC loop using the `ldaxr/stxr`
instruction pair.
Similar to previous issue, the new conditional branch was failin to
inherit the `prof_md` metadata. Theis PR correctly fix the branch
weights to the newly created branch within the LL/SC loop, ensuring
profile information is preserved.
Co-authored-by: Jin Huang <jingold@google.com>
In both ScheduleDAGInstrs and MachineScheduler, we call `BufferSize = 0`
as _reserved_ and `BufferSize = 1` as _unbuffered_. This convention is
stem from the fact that we set `SUnit::hasReservedResource` to true when
any of the SUnit's consumed resources has BufferSize equal to zero; set
`SUnit::isUnbuffered` to true when any of its consumed resources has
BufferSize equal to one.
However, `SchedBoundary::isUnbufferedGroup` doesn't really follow this
convention: it returns true when the resource in question is a
`ProcResGroup` and its BufferSize equals to **zero** rather than one.
This could be really confusing for the reader. This patch renames this
function to `isReservedGroup` in aligned with the convention mentioned
above.
NFC.
Implements chown and getgid per the POSIX specification and adds
corresponding unit tests. getgid is added as it is required by the chown
unit tests. This PR will address #165785
Co-authored-by: shubh@DOE <shubhp@mbm3a24.local>
The splitMustacheString function previously used a loop of
StringRef::split and StringRef::trim. This was inefficient as
it scanned each segment of the accessor string multiple times.
This change introduces a custom splitAndTrim function that
performs both operations in a single pass over the string,
reducing redundant work and improving performance, most notably
in the number of CPU cycles executed.
| Metric | Baseline | Optimized | Change |
| --- | --- | --- | --- |
| Time (ms) | 35\.57 | 35\.36 | \-0.59% |
| Cycles | 34\.91M | 34\.26M | \-1.86% |
| Instructions | 85\.54M | 85\.24M | \-0.35% |
| Branch Misses | 111\.9K | 112\.2K | \+0.27% |
| Cache Misses | 242\.1K | 239\.9K | \-0.91% |
In [1], the symbol __bpf_trap (macro BPF_TRAP) is removed if it is not
used in the code. In the discussion in [1], it is found that the branch
"if (Op.isSymbol())" is actually always false. Remove it to avoid
confusion.
[1] https://github.com/llvm/llvm-project/pull/166003
This doesn't improve performance (except with optimizations disabled),
since the compiler is able to fold our current implementation. However,
it does significantly reduce the amount of code the compiler has to sift
through, reducing compile times a bit.
The optimization records YAML files generated by Clang's LTO pipeline
are named "\*.opt.ld.yaml" rather than "\*.opt.yaml". This patch adds
that pattern into the search list of `find_opt_files` as well.
Otherwise we might end up with undefined register uses. Copying implicit
uses can cause problems where a register is both defined and used in the
same LDP, so I have not tried to add them here.
Fixes#164230
Reverts llvm/llvm-project#157330
The original revision introduces a bug in `isGuaranteedCollapsible`. The
`memref<3x3x1x96xf32, strided<[288, 96, 96, 1], offset: 864>>` is no
longer collapsable with the change. The revision reverts the change to
bring back correct behavior. `stride` should be computed as `96` like
the old behavior in the failed iteration.
92a1eb3712/mlir/lib/Dialect/MemRef/IR/MemRefOps.cpp (L2597-L2605)
This fixes register information handling for the `preserve_most` and
`preserve_all` calling conventions on Windows ARM64.
The root issue was cascading `if` statements whose behavior depended on
their order. This patch makes the minimal, tactical change needed for
Swift’s two calling conventions, unblocking current work. A broader
refactor to remove the ordering dependency is still desired and will
follow in a subsequent PR.
The WMMA MI(s) are missing the isConvergent flag. This causes incorrect
behavior in passes like machine-sink, where WMMA instructions get sunk
into divergent branches.
This patch fixes the issue by setting the isConvergent flag to 1 in the
VOP3PInstructions.td file.
The splitMustacheString function was saving StringRefs that
were already backed by an arena-allocated string. This was
unnecessary work. This change removes the redundant
Ctx.Saver.save() call.
This optimization provides a small but measurable performance
improvement on top of the single-pass tokenizer, most notably
reducing branch misses.
Metric | Baseline | Optimized | Change
-------------- | -------- | --------- | -------
Time (ms) | 35.77 | 35.57 | -0.56%
Cycles | 35.16M | 34.91M | -0.71%
Instructions | 85.77M | 85.54M | -0.27%
Branch Misses | 113.9K | 111.9K | -1.76%
Cache Misses | 237.7K | 242.1K | +1.85%
Fixes#164018.
The problem is that we're unable to do an implicit conversion sequence
on a template deduced argument, so the current vector templates can't
reconcile `vector<int, 4>` with `vector<bool, Sz>`. This PR separates
the vector templates into size-specific ones, getting rid of the `Sz`
deduction and allowing for the implicit conversion to be done.
This pass rewrites certain xegpu `CreateNd` and `LoadNd` operations that
feeds into `vector.transpose` to more optimal form to improve
performance. Specifically, low precision (bitwidth < 32) `LoadNd` ops
that feeds into transpose ops are rewritten to i32 loads with a valid
transpose layout such that later passes can use the load with transpose
HW feature to accelerate such load ops.
**Update:**
Pass is renamed to `OptimizeBlockLoads ` because later we plan to add
the array length optimization into this pass as well. This will break
down a larger load (like `32x32xf16`) into more DPAS-favorable array
length loads (`32x16xf16` with array length = 2). Both these
optmizations require rewriting `CreateNd` and `LoadNd` and it makes
sense to have a common pass for both.
Follow-up to bcb3d2f5122276ed9969fe2b2ef4428652800377.
Even though 32-bit win/asan is not well supported, we shouldn't drop it
without some discussion at least. Also, we probably shouldn't drop the
other sanitizers that are gated by COMPILER_RT_BUILD_SANITIZERS.
The tests no longer pass after switching to the runtimes build however
(I believe that build mode runs more of the tests?), so disable them.
`_XPLATFORM_SOURCE` needs to be defined to improve source code
compatibility (e.g. for `O_CLOEXEC`). The define `_UNIX03_THREADS` can
be removed, because it is automatically set by `_XOPEN_SOURCE=600`.
See the documentation of feature test macros:
https://www.ibm.com/docs/en/zos/3.1.0?topic=files-feature-test-macros
Tested on z/OS 3.1 with the Open XL C/C++ 2.2 compiler.
Currently, the jump table entry contains labels only. For example, the
following is one example:
BPF.JT.0.0:
.quad LBB0_1
.quad LBB0_2
.size BPF.JT.0.0, 16
Since the jump table entry contains a label, the relocation is necessary
so linker can resolve the label value. The relocation looks like below:
Relocation section '.rel.jumptables' at offset 0x160 contains 2 entries:
Offset Info Type Symbol's Value Symbol's Name
0000000000000000 0000000200000002 R_BPF_64_ABS64 0000000000000000 .text
0000000000000008 0000000200000002 R_BPF_64_ABS64 0000000000000000 .text
You can see that the symbol value is 0 which makes .rel.jumptables not
very useful.
Instead of having the label itself in the jump table entry, use the
difference of label and the section begin symbol. This can avoid the
relocation and the eventual jumptable entries in object file remains the
same as before.
Hex dump of section '.jumptables':
0x00000000 68000000 00000000 78000000 00000000 h.......x.......
There are no tests that specifically stop/start at
x86-partial-reduction, so no test cases have been updated for this
patch.
Reviewers: phoebewang, paperchalice, RKSimon, arsenm
Reviewed By: arsenm, RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/166048
Ensure that the nested native build uses the same python interpreter as
the main build, in case the python that CMake detects first is not the
python that the user has specified explicitly.
For example, if the person building LLVM wants to use a different python
interpreter to build (eg, testing the build with `python3.14` when
`python3` is a link to `python3.8`, or the default python doesn't have
development headers available) then they could add
`-DPYTHON_EXECUTABLE=python3.14` when invoking CMake. This should be
forwarded to the native CMake build to ensure that the same python is
used.
Original fix by Anuj Mittal <anuj.mittal@intel.com>.
From
https://github.com/llvm/llvm-project/pull/148554#issuecomment-3083261858
- this adds an option for API tests to be run with the both PDB readers
on Windows. As there are a lot of failures with PDB, this is an opt-in
per test.
To get PDB, `-g -gcodeview` has to be used on Clang. `-gcodeview` alone
isn't enough, because it won't cause clang to pass `-debug` to the
linker.
#149498 tracks the (currently) failing tests.
Introduced in OpenMP 6.0, the device UID shall be a unique identifier of
a device on a given system. (Not necessarily a UUID.) Since it is not
guaranteed that the (U)UIDs defined by the device vendor libraries, such
as HSA, do not overlap with those of other vendors, the device UIDs in
offload are always combined with the offload plugin name. In case the
vendor library does not specify any device UID for a given device, we
fall back to the offload-internal device ID.
The device UID can be retrieved using the `llvm-offload-device-info`
tool.
Updated the conditions for generating the cvt.ftz.f32.bf16 instruction
to include sm90 and isa7.8, so that ftz is only generated when it is
supported.
---------
Co-authored-by: Justin Fargnoli <jfargnoli@nvidia.com>
MachineFunctionSplitter was missing a skipFunction() check, causing it
to incorrectly split functions that should be skipped (e.g., functions
with optnone attribute).
This patch adds an early skipFunction() check in runOnMachineFunction()
to ensure these functions are never split, regardless of profile data
availability or other splitting conditions.
When running UTC on SLU/guards.ll (without LLVM changes), there are a number of changes in the UTC-generated checks. Submitting those first to simplify the diff of PR #164271, as most of the changes in the latter were actually these.
Update the verifier to first check if the number of incoming values
matches the number of predecessors, before using
incoming_values_and_blocks. We unfortunately need also check here, as
this may be called before verifyPhiRecipes runs.
Also update the verifier unit tests, to actually fail for the expected
recipes.
This avoids AArch64 legality rules depending on libcall
availability.
ARM, AArch64, and X86 all had custom lowering of fsincos which
all were just to emit calls to sincos_stret / sincosf_stret. This
messes with the cost heuristics around legality, because really
it's an expand/libcall cost and not a favorable custom.
This is a bit ugly, because we're emitting code trying to match the
C ABI lowered IR type for the aggregate return type. This now also
gives an easy way to lift the unhandled x86_32 darwin case, since
ARM already handled the return as sret case.
Avoid weird lambda returning function pointer and sink the libcall
logic to where the operation is handled. This allows chaining the
libcall logic to try sincos_stret and fallback to sincos. The resulting
cost seems too low.
Eventually this should be generated by tablegen for all functions.
For now add a manually implementation for sincos_stret, which I
have an immediate use for. This will allow pulling repeated code
across targets into shared call sequence code.
Also add sqrt just to make sure we can handle adding return attributes
on the declaration.
