Port Classic Flang to LLVM 19.1.0 #193
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
bryanpkc
force-pushed
the
release_19x
branch
2 times, most recently
from
615c1fb to
4d1a0c9
Compare
This commit cherry-picks 427c0e8 from the release_14x branch, which is a combination of the following legacy patches: Commit 82b38d2: [ClassicFlang] Port release_90 changes from flang-compiler/llvm Cherry-picked commit 2085211cfcca70411dc63f0d08763facc8a02090 by Eric Schweitz, resolved merge conflicts, fixed build failures (e.g. adapted CGDebugInfo.cpp to the new API), and fixed the DIGlobalVariable unit tests, which have been broken since commit edfad65eebdf045b050f37380b6b61d673513982. Commit 885dd87e5fdc: [DebugInfo] Removal of DIFortranArrayType and DIFortranSubrange These extensions are no more required after merge of below PR. Commit 5c9b2e0867d5: Modification to incorporate comments from @bryanpkc. Commit 822de2c: [AsmPrinter] Fix redundant names and types A bug was introduced in 82b38d2 while cherry-picking a DIGlobalVariable-related patch. Commit 45a70a8: Port driver changes from release/11x Cherry-picked c51f89679135f84675f492d560ec5535c2000cfe by Varun Jayathirtha from release_90, and resolved merge conflicts. To avoid conflicts with the new Flang, lib/Driver/ToolChains/Flang.{cpp,h} have been renamed to ClassicFlang.{cpp,h}, and the ENABLE_CLASSIC_FLANG macro is introduced to select which incarnation of Flang to build. The macro is set by running CMake with -DLLVM_ENABLE_CLASSIC_FLANG. After merge with LLVM 11: Move flang options to the end of the definitions list. Commit a9a8036: Port Classic Flang to LLVM 13 File changes to TargetLibraryInfo and DebugLocEntry to adapt to the code from release/13.x and make it work. Comment out the changes due a segmentation fault, code need to be reviewed properly once all commits are in Commit fe989b7: Fix -fveclib=PGMATH The use of #ifdef in include/clang/Driver/Options.td was incorrect and unsupported. As a result -fveclib=PGMATH was silently ignored, and in LLVM 12, it causes the invocation to fail. This patch unguards the option so that it is parsed correctly, but lets the FLANG_LLVM_EXTENSIONS macro continue to toggle the feature. Commit 7c224ae: Fix use of classic Flang as preprocessor Commit 8403c83: Merge FLANG_LLVM_EXTENSIONS macro into ENABLE_CLASSIC_FLANG Commit 486741e: Fix test failures when in classic Flang mode Add a new lit feature tag "classic_flang" to select which tests can or cannot be run when the driver is built for classic Flang. Handle LLVM_ENABLE_CLASSIC_FLANG in llvm/cmake/modules/HandleLLVMOptions.cmake instead of clang/CMakeLists.txt so that macro works in both clang and llvm. Commit a10f592: Port Classic Flang to LLVM 13 LLVM port from release_12x to release_13x, changes done in order to make project able to be build. Commit d385321 (partial): Change to Options.td in order to add the correct invocation for ffixed_line_length_VALUE. This commit also includes the following legacy patch: Commit 5da2c11: Add DebugInfo tests Commit c379c8d: [ClassicFlang][Driver] Correct the LLVM version passed by the Driver
Make TableGen respect the ENABLE_CLASSIC_FLANG macro and better separate the driver options that are mutually exclusive. When the macro is defined, do not build LLVM Flang at all.
…lang-compiler#92) * Support for -gdwarf-5 option in Flang driver. Summary: FLANG driver doesnt pass -gdwarf-4/5 to flang1 in form of xbits, while it passes for -gdwarf-2/3 -gdwarf-2 => -x 120 0x200 -gdwarf-3 => -x 120 0x4000 Due to this -gdwarf-5 is never honored and default option -gdwarf-4 is taken. # flang -gdwarf-5 test.f90 # llvm-dwarfdump a.out | grep version 0x00000000: Compile Unit: length = 0x0000008e version = 0x0004 Now 0x1000000/0x2000000 will be passed for -gdwarf-4/5 -gdwarf-4 => -x 120 0x1000000 -gdwarf-5 => -x 120 0x2000000 Testing: - GNU gdb fortran testsuite - check-llvm - check-debuginfo * Flang doenst choose correct dwarf version when multiple -g/-gdwarfN mentioned Summary: When multiple -g/-gdwarfN options are passed together at compile time, flang chooses the least one. Clang/gfortran etc choose the last one. -gdwarf-5 -gdwarf-3 => flang chooses 5 while clang/gfortran choose 3 -gdwarf-5 -g => flang choses the default while clang/gfortran choose 5 Testing: - check-llvm - check-debuginfo * Default dwarf version should be 4 for -g with flang Currently flang dumps dwarf version 2 for -g and 4 for absence of -g ------------------------- $ flang my.f90 $ llvm-dwarfdump a.out | grep version 0x00000000: Compile Unit: length = 0x0000003d version = 0x0004 abbr_offset = 0x0000 addr_size = 0x08 (next unit at 0x00000041) $ flang -g my.f90 $ llvm-dwarfdump a.out | grep version 0x00000000: Compile Unit: length = 0x00000047 version = 0x0002 abbr_offset = 0x0000 addr_size = 0x08 (next unit at 0x0000004b) ------------------------- It should be 4 for -g as it is the case with clang.
This commit is motivated by reducing the merge burden by shrinking the diff between llvm upstream and classic-flang-llvm-project. Outside of Flang, Fortran code is fed through the Compile phase, and the appropriate tooling is picked up through ToolChain::SelectTool. Classic Flang introduced a FortranFrontend, but these days this seems unnecessary. Fortran can go through the same machinery as everything else. * Use the Preprocess phase to preprocess Fortran code. This phase is always combined with the Compile phase. * Use the Compile phase to lower Fortran code to LLVM IR, and use the Backend phase to compile and optimize the IR. These phases are never combined. * Remove FortranFrontendJobClass. * Remove FortranFrontend tool (instead it's just the Flang tool, which in Classic Flang mode is Classic Flang). * Update tests which inspect the output of the Classic Flang tooling, and ensures that the driver does the right thing for various types of inputs. Based on a patch from Peter Waller <peter.waller@arm.com>.
Signed-off-by: Itay Bookstein <itay.bookstein@nextsilicon.com>
* -Wsuggest-override for ClassicFlangMacroBuilder::defineMacro Signed-off-by: Itay Bookstein <itay.bookstein@nextsilicon.com>
-emit-flang-llvm instructs flang to stop after flang2 and dump the LLVM IR. Can be useful for debugging and also would be a useful option for testing flang output more accurately. Signed-off-by: Richard Barton <richard.barton@arm.com>
Add %flang as a tool substitution available in lit tests. This apes the way %clang is defined and adds a $FLANG override in a similar vein. To avoid this being dead code, add a single test to check the flang driver is reporting the correct phases when running under various phase control options. Signed-off-by: Richard Barton <richard.barton@arm.com>
This patch partially reverts the following commit from an early version of Classic Flang LLVM: flang-compiler/llvm@6866909 When the LoopVectorize.cpp change in that commit is ported to LLVM 14, it causes a failure in the following test: llvm/test/Transforms/LoopVectorize/AArch64/smallest-and-widest-types.ll Reverting the change allows all tests to run cleanly.
Port commit 64bf2a6 to LLVM 16 and resolve merge conflicts. The original commit message follows: Pass LLVM target features (-mattr strings) to flang to embed in generated .ll files. For normal compilation this won't make much difference as the attributes are passed to clang after flang2 and can be applied there but this is crucial to enable LTO with flang as clang will not apply the attributes when building the flang2 output. libLTO will need to read these out of the object files to apply them. Signed-off-by: Richard Barton <richard.barton@arm.com>
Use existing clang support for uniquifying target feature lists to pass a unique list through to flang.
Signed-off-by: Paul Osmialowski <pawel.osmialowski@arm.com>
Version screen has been erroneously reporting as flang-new since LLVM 12
These CMake variables are used to set python variables in the lit config. As such, they need to bet set with a valid CMake boolean value that is also a valid python boolean value (1, True, etc.) and not one that is not (ON, NO, etc.) to work. This is fragile. Call the LLVM cmake helper function on these two downstream CMake
Signed-off-by: Richard Barton <richard.barton@arm.com>
Fix a bug where flang -E -save-temps input.F90 would preprocess the input file twice. Add a new test for preprocessor behaviour that would expose the bug and updated fortran-phases test for new behaviour. Also added -emit-flang-llvm to that test for completeness. Signed-off-by: Richard Barton <richard.barton@arm.com>
…tency out) Signed-off-by: Paul Osmialowski <pawel.osmialowski@arm.com>
1) All fma and contraction behvaviour follows from the -ffp-contract= flag setting. 2) At optimization level 0, -ffp-contract=fast flag will not be honoured. 3) At all other levels, -ffp-contract=fast will be the default behaviour and the flag is honoured.
In some workloads, users get "-lompstub" by running "flang -v" and link it explicitly, which may cause failure when compiling programs with OpenMP (flang -fopenmp test.f90 -lompstub). There is no runtime call in ompstub.c generated when "-fopenmp" is not added to compile the program. When "-fopenmp" is added to compile the program, the runtime in llvm/openmp(libomp.so) will be used. If users use some runtime library routines such as "omp_get_thread_num" and compiles it without the option "-fopenmp", both gfortran and ifort report errors of "undefined reference to ...". After remove "-lompstub" in the driver, classic-flang reports the error, too. So, it's safe to remove the "-lompstub" in the driver.
In LLVM 15, the Fortran and OpenMP runtime libraries are added to the linker command line using common methods (addFortranRuntime* and addOpenMPRuntime*). This commit adds Classic Flang awareness to addFortranRuntime*, so that Classic Flang doesn't attempt to link with LLVM Flang libraries. Re-using the same methods as Clang and LLVM Flang also helps reduce downstream delta. A Classic Flang test is added to ensure that the linker command is constructed correctly. LLVM 19 porting note: This commit has been amended to add back the -fno-fortran-main option that was deleted in upstream commit 8d53866.
Delete some snippets that had been commented out since commit a10f592 in release_13x.
Otherwise Classic Flang's CMakeLists.txt will be unable to find llvm-lit and cannot generate test-related goals, e.g. `check-all`.
The orignal PR was created by Isuruf flang-compiler#65
The -save-temps option has been causing the driver to run flang1/flang2 on the input file twice, and the fortran-preprocessor.f90 test was unable to detect the error, because it had used -c instead of -x in a couple of RUN commands. This patch fixes the test and the driver to make sure that the frontend is not invoked multiple times with -save-temps.
bryanpkc
force-pushed
the
release_19x
branch
2 times, most recently
from
c7e51fe to
31c269c
Compare
Some header files have been moved to new directories in LLVM 17. StringExtras.h is no longer included transitively since D155178, and now needs to be included explicitly by clang/Frontend/Utils.h.
Upstream commit 555b572 had reverted D118493, which let Clang add a toolchain-specific -rpath option to the linker command for the OpenMP library. This patch fixes up a Classic Flang driver test which is impacted by the revert.
The IR in clang/test/CodeGen/libpgmath-logfun-aarch64.ll can now be vectorized by a factor of two. The IR still does not call the unsupported libpgmath functions, as per the intention of the test.
-Mnomain and -fno-fortran-main were ineffective unless they were both used at the same time. This patch fixes the bug, formats needFortranMain and needFortranLibs, and adds some test cases for the logic.
Upstream commit 01797da had added a new VABIPrefix field to the VecDesc structure; modify the downstream libpgmath mappings accordingly.
...because the runners do not have enough threads.
Classic Flang accepts some Clang options that LLVM Flang (F18) does not. The Visibility of such options are modified to include "FlangOption".
...after upstream commit 13e977d changed how target feature dependencies are handled.
This patch cherry-picks a number of Visibility changes from the tip of trunk upstream, to make selected Clang options visible to Classic Flang. It also makes target-specific codegen options (-mxxxx) visible to Flang by default. Note that accepting the option does not imply support; Classic Flang may simply ignore certain options for backward compatibility.
Through this commit Mbackslash will be the default option. By default, escape characters are treated as regular characters unless Mnobackslash is specified. There is also the issue of treating both Mbackslash and Mnobackslash options independently. This is addressed in this patch and the testcases in flang repository will test this feature.
…lang-compiler#185) PGMATH has AVX512 runtime functions and can be executed only when the application is compiled in avx512 mode. The VecFuncs.def has no information about the TargetOptions and avx512 functions are selected even in avx2 mode. This issue is fixed by creating separate table for AVX512 functions and using them only when avx512 mode is specified.
LLVM 19 had changed how target feature flags are generated. In particular, the prerequisite flags "+sha2" and "+sha3" to "+sve2-sha3" are now explicit.
pawosm-arm
approved these changes
Oct 7, 2024
|
The AArch64 build is failing due to CMake not finding |
What is the use of Zlib in LLVM? There's CMake option for not using it (and a separate option for not using ZSTD too) |
Yes, I know. The problem here is that it isn't the LLVM build that breaks (because the zlib libraries and headers do exist), but the Flang build. We have never had to specify |
|
I have fixed the AArch64 Docker container which was using an old version of CMake, causing the problem. See fa9105e. |
|
@shivaramaarao Could you review this PR? |
shivaramaarao
approved these changes
Oct 14, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This pull request cherry-picks the Classic Flang patches from the
release_18xbranch, resolving a number of minor conflicts:clang/include/clang/Driver/Options.tdthat had been backported torelease_18xwere reverted, or reconciled with the new base branch.-fno-fortran-mainin llvm/llvm-project@8d53866. To resolve a merge conflict and to keep the code building without error, the option had been restored (but now guarded with theENABLE_CLASSIC_FLANGmacro).StringRef::equalshave been replaced withStringRef::operator==,StringRef::endswithwithStringRef::ends_with, andStringRef::startswithwithStringRef::starts_with. See llvm/llvm-project@3fa409f and llvm/llvm-project@4ec9a66.Driver::InstalledDirhad been replaced withDriver::Dir. See llvm/llvm-project@2b5cd8b.#undefdirectives inllvm/include/llvm/Analysis/VecFuncs.defhad been moved intollvm/lib/Analysis/TargetLibraryInfo.cppin llvm/llvm-project@430729d.Workflow-related commits have already been pushed to the
release_19xbranch.