-
Notifications
You must be signed in to change notification settings - Fork 327
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Support building against llvm15 #1035
Conversation
1) Drop 2 passes no longer supported by old PM 2) Update domTree passes to their wrapper passes 3) Update getAddress() api 4) Disable opaque pointer mode 5) Give warning when build with llvm15 All the changes are put inside llvm version check guard so that it builds with both llvm14 and llvm15
Removing Edit: Turns out I can't add reviewers. @gmarkall please add yourself and other reviewers who might be interested in this. |
ffi/core.cpp
Outdated
LLVMPY_GetGlobalContext() { | ||
auto context = LLVMGetGlobalContext(); | ||
#if LLVM_VERSION_MAJOR > 14 | ||
LLVMContextSetOpaquePointers(context, false); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Note to other reviewers - this does result in opaque pointers being disabled for the global context multiple times, but this operation is idempotent.
@@ -2,6 +2,11 @@ | |||
#define LLVMPY_CORE_H_ | |||
|
|||
#include "llvm-c/Core.h" | |||
|
|||
// Needed for macros that control version-specific behaviour - included here so | |||
// that they are available in all ffi translation units |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
In particular, this defines the LLVM_VERSION_MAJOR
macro referenced in various places in this PR.
@@ -653,7 +654,7 @@ def break_up_asm(self, asm): | |||
def test_rv32d_ilp32(self): | |||
self.check_riscv_target() | |||
llmod = self.fpadd_ll_module() | |||
target = self.riscv_target_machine(features="+f,+d") | |||
target = self.riscv_target_machine(features="+f,+d", abiname="ilp32") |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LLVM 14 defaults to the soft-float ABI, LLVM 15 defaults to the hard-float ABI. We explicitly specify the soft-float ABI for consistency with the expected output. Note that hard float and hard double ABIs are tested in the subsequent two tests below.
self.assertIn(patch, range(10)) | ||
valid = (14, 15) | ||
self.assertIn(major, valid) | ||
self.assertIn(patch, range(8)) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
(Max patch versions are 14.0.6 and 15.0.7)
This includes changes: - Adding the Windows LLVM 15 config to the Azure setup. - Converting the commands in the Azure pipeline file to a call to the setup_conda_environment.cmd script (it's easier to edit a script rather than individual lines in the Azure file). - Installation of LLVM 15 from conda-forge. - A workaround / supporting change for the conda-forge solution to conda-forge/llvmdev-feedstock#175.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The code changes here have my approval - I've already done a couple of rounds of review internally on this PR, and I made a few edits for a couple of left over changes that needed removing (adjustments to the Makefile, and a test that was accidentally skipped).
I added the CI config so we can test with LLVM 15 from conda-forge on each platform - someone else might want to review those changes.
I ran the Numba test suite with this PR locally, and didn't encounter any issues - the two passes that aren't available with the legacy pass manager anymore (arg promotion and loop unswitch) don't seem to be used by Numba.
As I see it, this PR provides us a way to give experimental support for LLVM 15 once merged, with no further blockers. Following that, I think we can:
- Build an LLVM 15 llvmdev package and move testing to LLVM 15 exclusively, with the aim of officially moving to 15.
- Move to the new pass manager, and re-enable the passes that have to be disabled with LLVM 15 at present.
- Switch to opaque pointers (at least in the binding layer)
- Move to LLVM 16+
Some discussion of the ordering / dependencies of these is warranted though - perhaps at the next office hours.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for putting this up!
ffi/passmanagers.cpp
Outdated
@@ -161,8 +161,13 @@ LLVMPY_AddCallGraphDOTPrinterPass(LLVMPassManagerRef PM) { | |||
|
|||
API_EXPORT(void) | |||
LLVMPY_AddDotDomPrinterPass(LLVMPassManagerRef PM, bool showBody) { | |||
#if LLVM_VERSION_MAJOR <= 14 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Suggest being consistent with the #ifs, either all "<= 14" or all "< 15"
@@ -333,12 +345,14 @@ LLVMPY_AddLoopUnrollAndJamPass(LLVMPassManagerRef PM) { | |||
LLVMAddLoopUnrollAndJamPass(PM); | |||
} | |||
|
|||
#if LLVM_VERSION_MAJOR < 15 | |||
API_EXPORT(void) | |||
LLVMPY_AddLoopUnswitchPass(LLVMPassManagerRef PM, bool optimizeForSize, |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
For reference the loop unswitch pass has been renamed to SimpleLoopUnswitchPass
in the new pass manager (https://reviews.llvm.org/D124376). Naturally to enable this we need to move to NPM so it makes sense that this is disabled for now but it can come back with NPM.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actually createSimpleLoopUnswitchLegacyPass
still exists in LLVM-15 so it can be dropped in here without issue.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
API_EXPORT(void)
LLVMPY_AddLoopUnswitchPass(LLVMPassManagerRef PM, bool optimizeForSize,
bool hasBranchDivergence) {
#if LLVM_VERSION_MAJOR < 15
unwrap(PM)->add(
createLoopUnswitchPass(optimizeForSize, hasBranchDivergence));
#else
unwrap(PM)->add(
createSimpleLoopUnswitchLegacyPass(optimizeForSize));
#endif
}
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for finding this, I have updated the code.
@@ -244,10 +254,12 @@ LLVMPY_AddAlwaysInlinerPass(LLVMPassManagerRef PM, bool insertLifetime) { | |||
unwrap(PM)->add(llvm::createAlwaysInlinerLegacyPass(insertLifetime)); | |||
} | |||
|
|||
#if LLVM_VERSION_MAJOR < 15 | |||
API_EXPORT(void) | |||
LLVMPY_AddArgPromotionPass(LLVMPassManagerRef PM, unsigned int maxElements) { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ditto with ArgumentPromotionPass
in NPM as below
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Did you mean legacy version of this pass is available as well? I couldn't find it.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sorry I meant the new pass version, I wrote both of these comments together then saw the legacy pass still existed.
On that note, the NPM setup slots in differently (using pipeline tuning options) so we can’t replace these in-situ. No needed for this change though.
…on checking condtions
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Re-approving following latest changes.
LGTM |
Added "Pending Buildfarm" to make sure this is all OK on all other platforms. |
BFID: |
|
Also passed buildfarm test against numba main ( |
Add support for llvmlite to be built against llvm15 while keeping support for llvm14 as well.
All the changes are put inside llvm version check guard so that it builds with both llvm14 and llvm15