Restructure llvm-patches into per-version directories - #1414
Merged
Conversation
Replace the flat llvm-patches/{llvm,spirv-translator} layout, where every
patch was applied to every LLVM version through ~115 lines of skip logic
in configure_llvm.sh, with self-contained per-version directories:
llvm-patches/llvm-{20,21,22}/{llvm,spirv-translator}/NNNN-*.patch
Patches apply in numeric order to pristine release/<V>.x and
llvm_release_<V>0 trees with no gating and no cross-patch offsets.
Verified per version: the resulting tree is identical to what the old
pipeline produced, except for three intended deltas:
- Dropped 0005-fix-archive-data-layout.patch: verified no-op on every
supported version; upstream IRMover already inherits the data layout
(IRMover.cpp, present since llvmorg-17).
- Dropped spirv-translator fp_fast_mode patch: lit-test-only change,
re-inserting a flag removed upstream; zero product-code effect.
- Added preserve-device-debug-info (was untracked 0007): upstream
llvm#210504, first ships in LLVM 24.
Old 0001-Allow-up-to-v1.2 + 0004-only-necessary-exts are merged into a
single spirv-version-and-extensions patch per version, retiring the
llvm22-specific combined variant. llvm-21-macos.patch (applied by
nothing, conflicted with configure_llvm.sh output) and the LLVM 17/18
translator variant are deleted. LLVM 17/18/19 support is dropped; 17/18
were unbuildable (macos patch and phi-coalesce patch fail to apply).
Each patch header carries Applies-to and Upstream-Status lines.
Patch application is now: apply everything in
llvm-patches/llvm-${VERSION}/{llvm,spirv-translator}/*.patch in numeric
order, hard-failing on the first error. All per-patch version gating is
gone (412 -> 349 lines).
--version accepts 20, 21, 22, or latest. latest clones the maintained
CHIP-SPV/llvm-project branch chipStar-llvm-23 (single-branch) plus
translator llvm_release_230 and applies no patches; it is labeled
experimental. Everything else (retry, arg parsing, binutils handling,
variant/link-type, -N and --configure-only flows) is unchanged.
presubmit.yml matrix pins move 19 -> 20 since the script now rejects 19.
PoCL 7.1 supports LLVM 20 and chipStar-llvm-20 cache-key branches exist.
README.md: prerequisites and toolchain instructions now name LLVM 20/21/22 (+ experimental latest via the maintained chipStar-llvm-23 branch); manual steps clone upstream release branches and apply llvm-patches/llvm-<V>/. MacOS-Issues.md, docs/MacOS.md: llvm-21-macos.patch is gone; macOS support is applied automatically by configure_llvm.sh from the per-version dirs. Stale LLVM 19 references in live instructions updated; historical issue write-ups left as-is.
Sync the preserve-device-debug-info patch in all three version dirs with
the final merged form of llvm#210504: its second commit (f44bf2b4ce5c)
also enables SPV_INTEL_optnone when -g is given. clang emits the optnone
attribute at -O0; without the extension the translator drops it, the
device compiler optimizes the kernel anyway, and gdb-oneapi reports every
local as <optimized out> - the exact symptom of chipStar#1004. Verified
each series still applies cleanly to pristine release/{20,21,22}.x.
LLVMCheck.cmake allowed clang 14+ with CLANG_VERSION_LESS_{15,16,17}
feature flags. With LLVM 17/18/19 toolchain support dropped, raise the
floor to 20 and remove the flag ladder. The remaining
CLANG_VERSION_LESS_* consumers (CMakeLists.txt:366,699,
bitcode/CMakeLists.txt:33) evaluate correctly with the flags never set.
docker.md: DockerfileFull section now states reality - it is stale and broken (builds LLVM 16, which is no longer supported, and passes --only-necessary-spirv-exts, an option configure_llvm.sh no longer has) with rework notes for the supported 20/21/22 set. DockerfileLatest section: stale LLVM 15 reference corrected to the actual LLVM 22 native setup. DockerfileBase: patch-dir layout comment updated to the per-version llvm-patches/llvm-<V>/ scheme.
It built LLVM 16 with a configure_llvm.sh option that no longer exists, was referenced by nothing, and could not have worked for years. Deleted along with its docker.md section.
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Replaces the flat llvm-patches layout and the ~115 lines of per-patch version gating in configure_llvm.sh with self-contained directories llvm-patches/llvm-{20,21,22} applied in numeric order (script: 412 to 349 lines). LLVM 17/18/19 support is dropped (17/18 could not build), the no-op archive-data-layout and fp_fast_mode patches are deleted, the debug info patch now matches the merged form of llvm/llvm-project#210504 including SPV_INTEL_optnone, and --version latest clones the maintained chipStar-llvm-23 branch. Every series applies cleanly to pristine release branches, and all three versions were built on Aurora with gdb-oneapi -g debugging validated per #1350.