The -Wunsafe-buffer-usage analysis currently warns when a const sized
array is safely accessed, with an index that is bound by the remainder
operator. The false alarm is now suppressed.
Example:
int circular_buffer(int array[10], uint idx) {
return array[idx %10]; // Safe operation
}
rdar://148443453
---------
Co-authored-by: MalavikaSamak <malavika2@apple.com>
There was a problem introduced today where sometimes the CIR for
functions with no arguments would be generated with phantom arguments.
This was causing intermittent test failures.
The problem appears to have been that we were using two different
Profile implementations to generate the FoldingSetNodeID for
CIRGenFunctionInfo so occaissionally when we tried to look for a
pre-existing entry for a function with no arguments it would incorrectly
match a CIRGenFunctionInfo entry that had arguments.
To prevent this from happening again, I rewrote one of the two Profile
functions to call the other.
Split from #133161. This patch fills in the implementation for a number
of the MustacheHTMLGenerator methods. Many of these APIs are just
stubbed out, and will have their implementation filled in by later
patches.
Co-authored-by: Peter Chou <peter.chou@mail.utoronto.ca>
This reapplies 33684ac9be with appropriate
requires on tests.
COFF requires that a function exists with the same name as a comdat. Not
having this key function results in `LLVM ERROR: Associative COMDAT
symbol '...' does not exist.` CFI by default will attempt to
canonicalize a function by appending `.cfi` to its name which allows
external references to refer to the new canonical alias, but it does not
change the comdat name. We can change the comdat name since symbol and
comdat resolution occurs before LTO, so we already know which symbols
are prevailing.
## description
`tensor.concat` requires operands and the result to match on all
dimensions except the concatenation dimension. If one operand is already
static in those dimensions, the other operands and result type may
safely be refined to that same static shape. This PR adds
canonicalization patterns to refine `tensor.concat` types and propagate
static shapes to other canonicalization patterns through casts.
## example
```mlir
%2 = tensor.concat dim(0) %0, %1: (tensor<?x12xi32>, tensor<?x?xi32>) ->tensor<?x12xi32>
```
becomes:
```mlir
%cast = tensor.cast %1 : tensor<?x?xi32> to tensor<?x12xi32>
%2 = tensor.concat dim(0) %0, %cast : (tensor<?x12xi32>,
tensor<?x12xi32>) -> tensor<?x12xi32>
```
---------
Co-authored-by: Ian Wood <ianwood2024@u.northwestern.edu>
These functions don't have a _time64 variant, so we can't use time_t
directly (since our time_t is a uint64_t). The workaround is to use
longs when doing the syscall and write back when necessary.
COFF requires that a function exists with the same name as a comdat. Not
having this key function results in `LLVM ERROR: Associative COMDAT
symbol '...' does not exist.` CFI by default will attempt to
canonicalize a function by appending `.cfi` to its name which allows
external references to refer to the new canonical alias, but it does not
change the comdat name. We can change the comdat name since symbol and
comdat resolution occurs before LTO, so we already know which symbols
are prevailing.
This fixes issues when compiling SliceMatchers.h separately. In
particular, the missing include triggered a "member access into
incomplete type" error:
```
third_party/llvm/llvm-project/mlir/include/mlir/Query/Matcher/SliceMatchers.h:91:30: error: member access into incomplete type 'Operation'
91 | for (auto operand : subOp->getOperands()) {
| ^
third_party/llvm/llvm-project/mlir/include/mlir/Analysis/SliceAnalysis.h:20:7: note: forward declaration of 'mlir::Operation'
20 | class Operation;
| ^
```
Similar to VPWidenIntOrFpInductionRecipe, VPWidenPointerInductionRecipe
only has a single incoming value and no backedge value.
Currently NFC as there are no problematic uses at the moment.
This patch fixes the following error in case of LLVM_ENABLE_RPMALLOC=ON
and clang or other non MSVC compiler:
```
CMake Error at cmake/modules/AddLLVM.cmake:644 (add_library):
Target "LLVMSupport" links to item " -INCLUDE:malloc" which has leading or
trailing whitespace. This is now an error according to policy CMP0004.
Call Stack (most recent call first):
cmake/modules/AddLLVM.cmake:944 (llvm_add_library)
cmake/modules/AddLLVM.cmake:917 (add_llvm_library)
lib/Support/CMakeLists.txt:139 (add_llvm_component_library)
```
Note `delayload_flags` is empty before this line is case of non MSVC
compiler.
The behavior is the same with CMake 3.25.3 and 4.0.2.
When vectorizing loops with early exits that is nested within another
one, one of the loop exits may be outside both loops, so setting adding
it to the parent loop is incorrect. Also use the original parent loop
for exit blocks.
Omit spills when va_start is not present, like we do on other targets.
In most situations this is just an optimization, but for thunks in
arm64ec, we need this because it's illegal to reference varargs in a
thunk.
Fix the bug that prevented omitting vararg spills from working properly:
we need to avoid interfering with musttail handling. (This is not a
Windows-only issue, but it mostly affects Windows because musttail
thunks are mostly used for the Microsoft C++ ABI.)
On arm64ec, don't set x4 and x5 for musttail thunks; forward them from
the caller.
Fixes#139856 .
This transform creates target specific instructions which must have
legal types. We were checking this for fixed vectors, but not scalable
vectors. This caused a crash with <vscale x 1 x i1> which isn't legal
for Zve32x.
Given a larger-than-legal shuffle we will split into multiple sub-parts.
This adds a check to the computed costs of sub-shuffles so that repeated
sequences are not accounted for multiple times. This especially reduces
the cost of broadcasts/splats.
This function can be used to retrieve the number of bits that can be used
for arithmetic in a given address space (i.e. the range of the address
space). For most in-tree targets this should not make any difference
but differentiating between the size of a pointer in bits and the address
range is extremely important e.g. for CHERI-enabled targets, where pointers
carry additional metadata such as bounds and permissions and only a subset
of the pointer bits is used as the address.
The address size is defined to be the same as the index size.
We considered adding a separate property since targets exist where indexing
and address range actually use different sizes (AMDGPU fat pointers with
160 representation, 48 bit address and 32 bit index), but for the purposes
of LLVM semantics, differentiating them does not add much value and it
introduces a lot of complexity in ensure the correct bits are used. See
the reasoning by @nikic on https://discourse.llvm.org/t/clarifiying-the-semantics-of-ptrtoint/83987/38https://discourse.llvm.org/t/clarifiying-the-semantics-of-ptrtoint/83987/49
Originally uploaded as https://reviews.llvm.org/D135158
Reviewed By: davidchisnall, krzysz00
Pull Request: https://github.com/llvm/llvm-project/pull/139347
Currently, `Posix/sanitizer_set_report_path_test.cpp` contains the
following check: `// CHECK: ERROR: Can't create directory:
{{.*}}Posix/Output/sanitizer_set_report_path_test.cpp.tmp`. This makes
an assumption that the test file resides in `Posix/Output`, however when
testing on a remote device, an alternative temporary directory path is
used. This patch instead checks that the path in the error message
matches the requested path dynamically.
Need to set the insertion point for (non-schedulable) vector node after
the last instruction in the node to avoid def-use breakage. But it also
causes miscompilation with gather/buildvector operands of the phi nodes,
used in the same phi only in the block.
These nodes supposed to be inserted at the end of the block and after
changing the insertion point for the non-schedulable vec block, it also
may break def-use dependencies. Need to prevector such nodes, to emit
them as early as possible, so the vectorized nodes are inserted before
these nodes.
Fixes#139728
Reviewers: hiraditya, HanKuanChen, RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/139917
- Add `DagInit::get` overloads that do not need ValName to be specified.
- Fix some calls to either not create temporary arrays for DAG args or
use the std::pair<> overload.
For large binaries gsymutil ends up using too much memory. This diff
adds DIE tree cleanup per compile unit to reduce memory usage.
P. S. Not sure about formatting. Maybe it hasn't been run in a while, or
I have misconfigured something.
`$ git clang-format HEAD~1
clang-format did not modify any files
$ clang-format --version
clang-format version 21.0.0git
(git@github.com:peremyach/llvm-project.git
8d945c8357e1bd9872a34f92620d4916bfd27482)
`
Co-authored-by: Arslan Khabutdinov <akhabutdinov@fb.com>
Stack clash protection code was missing from
RISCVFrameLowering::eliminateCallFramePseudoInstr, calling allocateStack
fixes it.
This patch also fixes the tests in stack-probing-dynamic.ll that should
be testing the stack allocation before a function call.
This PR adds support for the `FlattenCFG` transformation on `switch`
statements. It also introduces the `SwitchFlatOp`, which is necessary
for subsequent lowering to LLVM.
This fixes a warning about an unhandled Decl that was recently
introduced. While adding the new decl, I noticed that one of the others
that was in the "NYI" group belonged in the "unreachable" group, so I
moved it also.
This change replaces the simplified call that we were previously using
to convert the function type provided by a global declaration to the CIR
function type. We now go through 'arrangeGlobalDeclaration' which builds
the function type in a more complicated manner. This change has no
observable differences for the currently upstreamed CIR support, but it
is necessary to prepare for C++ member function calls, which require the
extra handling.
