Clang `-Wpass-failed=transform-warning` fails MLIR build with "loop not unrolled"

[nikic]

When building with clang 20.1.8:

[197/673] Building CXX object tools/ml....MLIRVectorToGPU.dir/VectorToGPU.cpp.o
FAILED: tools/mlir/lib/Conversion/VectorToGPU/CMakeFiles/obj.MLIRVectorToGPU.dir/VectorToGPU.cpp.o 
CCACHE_CPP2=yes CCACHE_HASHDIR=yes /usr/bin/ccache /usr/lib64/ccache/clang++ -DGTEST_HAS_RTTI=0 -D_DEBUG -D_GLIBCXX_ASSERTIONS -D_GLIBCXX_USE_CXX11_ABI=1 -D_GNU_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS -I/home/npopov/repos/llvm-project/build/tools/mlir/lib/Conversion/VectorToGPU -I/home/npopov/repos/llvm-project/mlir/lib/Conversion/VectorToGPU -I/home/npopov/repos/llvm-project/build/tools/mlir/include -I/home/npopov/repos/llvm-project/mlir/include -I/home/npopov/repos/llvm-project/build/include -I/home/npopov/repos/llvm-project/llvm/include -fPIC -fno-semantic-interposition -fvisibility-inlines-hidden -Werror -Werror=date-time -Werror=unguarded-availability-new -Wall -Wextra -Wno-unused-parameter -Wwrite-strings -Wcast-qual -Wmissing-field-initializers -pedantic -Wno-long-long -Wc++98-compat-extra-semi -Wimplicit-fallthrough -Wcovered-switch-default -Wno-noexcept-type -Wnon-virtual-dtor -Wdelete-non-virtual-dtor -Wsuggest-override -Wstring-conversion -Wmisleading-indentation -Wctad-maybe-unsupported -fdiagnostics-color -ffunction-sections -fdata-sections -Wundef -Werror=mismatched-tags -Werror=global-constructors -O3 -DNDEBUG -std=c++17  -fno-exceptions -funwind-tables -fno-rtti -UNDEBUG -MD -MT tools/mlir/lib/Conversion/VectorToGPU/CMakeFiles/obj.MLIRVectorToGPU.dir/VectorToGPU.cpp.o -MF tools/mlir/lib/Conversion/VectorToGPU/CMakeFiles/obj.MLIRVectorToGPU.dir/VectorToGPU.cpp.o.d -o tools/mlir/lib/Conversion/VectorToGPU/CMakeFiles/obj.MLIRVectorToGPU.dir/VectorToGPU.cpp.o -c /home/npopov/repos/llvm-project/mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp
In file included from /home/npopov/repos/llvm-project/mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp:13:
In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/Conversion/VectorToGPU/VectorToGPU.h:12:
In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/PatternMatch.h:12:
In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/Builders.h:12:
In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/OpDefinition.h:22:
In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/Dialect.h:16:
In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/DialectRegistry.h:16:
In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/MLIRContext.h:13:
In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/Support/TypeID.h:20:
In file included from /home/npopov/repos/llvm-project/llvm/include/llvm/ADT/STLExtras.h:23:
/home/npopov/repos/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:45:14: error: loop not unrolled: the optimizer was unable to perform the requested transformation; the transformation might be disabled or specified as part of an unsupported transformation ordering [-Werror,-Wpass-failed=transform-warning]
   45 |   static Ret callback_fn(intptr_t callable, Params ...params) {
      |              ^

From the error message, I can't figure out where this unrolling request is actually coming from...

[llvmbot]

@llvm/issue-subscribers-mlir

Author: Nikita Popov (nikic)

When building with clang 20.1.8: ``` [197/673] Building CXX object tools/ml....MLIRVectorToGPU.dir/VectorToGPU.cpp.o FAILED: tools/mlir/lib/Conversion/VectorToGPU/CMakeFiles/obj.MLIRVectorToGPU.dir/VectorToGPU.cpp.o CCACHE_CPP2=yes CCACHE_HASHDIR=yes /usr/bin/ccache /usr/lib64/ccache/clang++ -DGTEST_HAS_RTTI=0 -D_DEBUG -D_GLIBCXX_ASSERTIONS -D_GLIBCXX_USE_CXX11_ABI=1 -D_GNU_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS -I/home/npopov/repos/llvm-project/build/tools/mlir/lib/Conversion/VectorToGPU -I/home/npopov/repos/llvm-project/mlir/lib/Conversion/VectorToGPU -I/home/npopov/repos/llvm-project/build/tools/mlir/include -I/home/npopov/repos/llvm-project/mlir/include -I/home/npopov/repos/llvm-project/build/include -I/home/npopov/repos/llvm-project/llvm/include -fPIC -fno-semantic-interposition -fvisibility-inlines-hidden -Werror -Werror=date-time -Werror=unguarded-availability-new -Wall -Wextra -Wno-unused-parameter -Wwrite-strings -Wcast-qual -Wmissing-field-initializers -pedantic -Wno-long-long -Wc++98-compat-extra-semi -Wimplicit-fallthrough -Wcovered-switch-default -Wno-noexcept-type -Wnon-virtual-dtor -Wdelete-non-virtual-dtor -Wsuggest-override -Wstring-conversion -Wmisleading-indentation -Wctad-maybe-unsupported -fdiagnostics-color -ffunction-sections -fdata-sections -Wundef -Werror=mismatched-tags -Werror=global-constructors -O3 -DNDEBUG -std=c++17 -fno-exceptions -funwind-tables -fno-rtti -UNDEBUG -MD -MT tools/mlir/lib/Conversion/VectorToGPU/CMakeFiles/obj.MLIRVectorToGPU.dir/VectorToGPU.cpp.o -MF tools/mlir/lib/Conversion/VectorToGPU/CMakeFiles/obj.MLIRVectorToGPU.dir/VectorToGPU.cpp.o.d -o tools/mlir/lib/Conversion/VectorToGPU/CMakeFiles/obj.MLIRVectorToGPU.dir/VectorToGPU.cpp.o -c /home/npopov/repos/llvm-project/mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp In file included from /home/npopov/repos/llvm-project/mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp:13: In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/Conversion/VectorToGPU/VectorToGPU.h:12: In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/PatternMatch.h:12: In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/Builders.h:12: In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/OpDefinition.h:22: In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/Dialect.h:16: In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/DialectRegistry.h:16: In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/IR/MLIRContext.h:13: In file included from /home/npopov/repos/llvm-project/mlir/include/mlir/Support/TypeID.h:20: In file included from /home/npopov/repos/llvm-project/llvm/include/llvm/ADT/STLExtras.h:23: /home/npopov/repos/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:45:14: error: loop not unrolled: the optimizer was unable to perform the requested transformation; the transformation might be disabled or specified as part of an unsupported transformation ordering [-Werror,-Wpass-failed=transform-warning] 45 | static Ret callback_fn(intptr_t callable, Params ...params) { | ^ ``` From the error message, I can't figure out where this unrolling request is actually coming from...

[joker-eph]

Can you preprocess the file? Maybe there is a loop unroll annotation in the source?
This error hints at a:

#pragma clang loop unroll(full)

that can't be honored if I understand correctly?

[nuudlman]

I believe it comes from std::find_if:

[3649/4399] Building CXX object tools/mlir/lib/Conversion/VectorToGPU/CMakeFiles/obj.MLIRVectorToGPU.dir/VectorToGPU.cpp.o
In file included from /home/isaac/Work/llvm-project/src/mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp:13:
In file included from /home/isaac/Work/llvm-project/src/mlir/include/mlir/Conversion/VectorToGPU/VectorToGPU.h:12:
In file included from /home/isaac/Work/llvm-project/src/mlir/include/mlir/IR/PatternMatch.h:12:
In file included from /home/isaac/Work/llvm-project/src/mlir/include/mlir/IR/Builders.h:12:
In file included from /home/isaac/Work/llvm-project/src/mlir/include/mlir/IR/OpDefinition.h:22:
In file included from /home/isaac/Work/llvm-project/src/mlir/include/mlir/IR/Dialect.h:16:
In file included from /home/isaac/Work/llvm-project/src/mlir/include/mlir/IR/DialectRegistry.h:16:
In file included from /home/isaac/Work/llvm-project/src/mlir/include/mlir/IR/MLIRContext.h:12:
In file included from /home/isaac/Work/llvm-project/src/mlir/include/mlir/Support/LLVM.h:23:
In file included from /home/isaac/Work/llvm-project/src/llvm/include/llvm/Support/Casting.h:20:
In file included from /usr/bin/../lib64/gcc/x86_64-pc-linux-gnu/15.2.1/../../../../include/c++/15.2.1/memory:71:
In file included from /usr/bin/../lib64/gcc/x86_64-pc-linux-gnu/15.2.1/../../../../include/c++/15.2.1/bits/stl_uninitialized.h:63:
/usr/bin/../lib64/gcc/x86_64-pc-linux-gnu/15.2.1/../../../../include/c++/15.2.1/bits/stl_algobase.h:2095:7: warning: loop not unrolled: the optimizer was unable to perform the requested transformation; the transformation might be disabled or specified as part of an unsupported transformation ordering [-Wpass-failed=transform-warning]
 2095 |       while (__first != __last && !__pred(__first))
      |       ^
1 warning generated.

The relevant part of stl_algobase.h:

  // Implementation of std::find_if, also used in std::remove_if and others.
  template<typename _Iterator, typename _Predicate>
    _GLIBCXX20_CONSTEXPR
    inline _Iterator
    __find_if(_Iterator __first, _Iterator __last, _Predicate __pred)
    {
#pragma GCC unroll 4
      while (__first != __last && !__pred(__first))
	++__first;
      return __first;
    }

[nuudlman]

This seems to be have been introduced with GCC 15 as the relevant patch (https://gcc.gnu.org/pipermail/libstdc++/2024-August/059366.html) landed after GCC 14 was branched.

For reference here is GCC 14 (which doesn't have it):
https://github.com/gcc-mirror/gcc/blob/c9cd41fba9ebd288c4f101e4b99da934bcb96a11/libstdc%2B%2B-v3/include/bits/stl_algobase.h#L2083C1-L2093C6

And here is GCC 15 (which has it):
https://github.com/gcc-mirror/gcc/blob/1b306039ac49f8ad91ca71d3de3150a3c9fa792a/libstdc%2B%2B-v3/include/bits/stl_algobase.h#L2088C1-L2098C6

[nikic]

I've put up #160472 to just disable this warning. It's not like we care about missed transforms, and we have no control over this pragma.

Possibly some code could be rewritten to not fire the warning, but the diagnostics are pretty terrible so I have no idea what the actual code causing this is.

[joker-eph]

Conveying here the relevant part of my comment on the PR:

• Why isn't this warning be opt-in? Seems like this warning itself is conceptually flawed somehow, at least not at the usual bar of default warnings in clang. I don't remember having to disable clang warnings permanently, do we have precedent for this?
• Why is it a warning an not an "optimization remark"?
• maybe it should be able to maybe not hit on system headers (which I don't see how do implement in a robust way), or just not warn on #pragma unroll if it isn't robust?

Tagging clang on this issue, seems like more relevant to clang than MLIR.

[nikic]

I think the basic idea here is that if the code author added an explicit optimization pragma (for unrolling, vectorization, etc) and it was ignored, they likely want to know about it. This is different from general optimization remarks, in that an explicit optimization request has been made.

Of course, this doesn't hold up if system headers include such pragmas. If it's technically possible to exclude those, then that would certainly be an improvement.

[joker-eph]

Of course, this doesn't hold up if system headers include such pragmas. If it's technically possible to exclude those, then that would certainly be an improvement.

Actually I was thinking it would be hard to filter in a robust way because I was thinking about filtering in LLVM, but we could do so reliably in Clang at the LLVM IR emission time: we just need to be able to express in LLVM IR differently a "pragma request with diagnostic" and "pragma request without diagnostic" for this.

[nikic]

For the record, the MLIR code that triggers this issue is:

llvm-project/mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp

Lines 369 to 375 in 2195fe7

	if (llvm::any_of(dependentOps, [useNvGpu](Operation *op) {
	if (!supportsMMaMatrixType(op, useNvGpu)) {
	LDBG() << "cannot convert op: " << *op;
	return true;
	}
	return false;
	}))