-
Notifications
You must be signed in to change notification settings - Fork 103
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
[CIR][LowerToLLVM] Lowered LLVM code for pointer arithmetic should have inbounds #1191
base: main
Are you sure you want to change the base?
Conversation
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.
Hello! Thank you for your interest and contribution. Here are some suggestions on the overall direction:
- This change might not be suitable as an upstream modification to the LLVM dialect. If you believe it benefits all LLVM users, consider submitting this patch to the
llvm-project
monorepo. - Alternatively, investigate what might be missing in CIR's responsibilities, comparing with original Clang CodeGen. For instance, in the original case described in issue CIR generated LLVM code for pointer arithmetic misses inbounds #952, the
cir.ptr_stride
operation generates a GEP when lowering to LLVM. It might be more appropriate to address the issue there, perhaps by introducing common helpers for creating GEPs. (I don't have full context of the issue, for your reference only 😉) - Please add a test for each part of your changes. Ideally, development should be driven by your test cases.
Thanks for your time working on this! The current changes are not related to ClangIR and it's not an appropriate way to resolve #952 . Given the following input code: void foo(int *iptr) { iptr + 2; } The CIR generated for the above code would be something similar to: // ...
%1 = cir.const 2 : i32
%2 = cir.ptr_stride(%0 : !cir.ptr<i32>, %1 : i32), !cir.ptr<i32>
// ... The generated CIR is further lowered to the following LLVM dialect code in %0 = llvm.getelementptr %1[%2] : (!llvm.ptr, i32) -> !llvm.ptr, i32 Apparently the root cause is that |
Alright, I'll work on |
@liusy58 The The C++ standard says that if the result of pointer arithmetic is out of bounds, the behavior is undefined. So I believe for |
Thank you. Let me check it. |
Thanks @Lancern and @seven-mile for the great review and clarifications. @liusy58 welcome to the ClangIR project! |
12458ba
to
9c7e362
Compare
@Lancern Hi, I have update the code and could you please review 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.
Thanks for working on this! The CI shows you have 13 failed tests, please resolve them and it should be good to go!
cec82cc
to
d9b7c41
Compare
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 the update and bearing all the comments! Adding inbounds
unconditionally might be considered not quite right.
I believe inbounds
of GEP is about low-level pointer arithmetic rather than memory model in the language. The keyword controls the overflow behaviour concisely (ref), which leads to a common pattern in OG CodeGen:
clangir/clang/lib/CodeGen/CGExprScalar.cpp
Lines 4090 to 4095 in eacaabb
if (CGF.getLangOpts().isSignedOverflowDefined()) | |
return CGF.Builder.CreateGEP(elemTy, pointer, index, "add.ptr"); | |
return CGF.EmitCheckedInBoundsGEP( | |
elemTy, pointer, index, isSigned, isSubtraction, op.E->getExprLoc(), | |
"add.ptr"); |
Additionally, we'd better be careful to apply language conformance: some options are designed to control the conformance or provide some extensions. The code above indicates an instance: -fwrapv
controlling SOB. We should take care of them to keep the frontend functional ; )
There might be other considerations for a specific case in OG CodeGen. Usually the reliability comes from the correspondence of skeleton between the old and new codes. Given the fact that we have no choice but migrate these logic to LowerToLLVM, we should be especially cautious. IMHO this fix is not necessarily finished in one single patch.
For the next step, I think we can discuss what changes should this patch include. A good start is to just consider #952. If it's suitable, pack more changes in your following patches, and so on. It's your first-time contribution after all, no need to hurry 😉
d9b7c41
to
4e05834
Compare
…tic should have inbounds.
17d944f
to
0362c64
Compare
Hi, @seven-mile , I have updated the commit, please review it. Thanks! |
ptrStrideOp, resultTy, elementTy, adaptor.getBase(), index); | ||
rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>(ptrStrideOp, resultTy, | ||
elementTy, adaptor.getBase(), | ||
index, /*inbounds=*/true); |
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.
It's still unconditional. We cannot accept miscompilation. Please make sure we emit the same LLVM IR as e.g. this godbolt, or optionally lead it to an assersion failure.
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.
ok, I will work on it later.
#886 is a related potential area for follow-up work here if you're interested :) |
Fix issue in #952.