[ML Inliner] Fix inconsistancy between CallGraph and FunctionLevels #122830
Conversation
When building our internal code, one of the function annotated with `co_await` (c++20) will cause inconsistancy: The function exists in CG but not FunctionLevels. This patch fixes it by using FunctionLevels only. Signed-off-by: Peter Rong <[email protected]>
DataCorrupted
requested review from
boomanaiden154,
ellishg,
kyulee-com and
mtrofin
llvmbot
added
mlgo
llvm:analysis
|
@llvm/pr-subscribers-llvm-analysis Author: Peter Rong (DataCorrupted) ChangesWhen building our internal code, one of the function annotated with This patch fixes it by using FunctionLevels only. Full diff: https://github.com/llvm/llvm-project/pull/122830.diff 1 Files Affected:
diff --git a/llvm/lib/Analysis/MLInlineAdvisor.cpp b/llvm/lib/Analysis/MLInlineAdvisor.cpp
index 2db58d1c2578bf..68d8f794d23897 100644
--- a/llvm/lib/Analysis/MLInlineAdvisor.cpp
+++ b/llvm/lib/Analysis/MLInlineAdvisor.cpp
@@ -189,7 +189,9 @@ MLInlineAdvisor::MLInlineAdvisor(
}
unsigned MLInlineAdvisor::getInitialFunctionLevel(const Function &F) const {
- return CG.lookup(F) ? FunctionLevels.at(CG.lookup(F)) : 0;
+ return FunctionLevels.find(CG.lookup(F)) != FunctionLevels.end()
+ ? FunctionLevels.at(CG.lookup(F))
+ : 0;
}
void MLInlineAdvisor::onPassEntry(LazyCallGraph::SCC *CurSCC) {
|
|
@llvm/pr-subscribers-mlgo Author: Peter Rong (DataCorrupted) ChangesWhen building our internal code, one of the function annotated with This patch fixes it by using FunctionLevels only. Full diff: https://github.com/llvm/llvm-project/pull/122830.diff 1 Files Affected:
diff --git a/llvm/lib/Analysis/MLInlineAdvisor.cpp b/llvm/lib/Analysis/MLInlineAdvisor.cpp
index 2db58d1c2578bf..68d8f794d23897 100644
--- a/llvm/lib/Analysis/MLInlineAdvisor.cpp
+++ b/llvm/lib/Analysis/MLInlineAdvisor.cpp
@@ -189,7 +189,9 @@ MLInlineAdvisor::MLInlineAdvisor(
}
unsigned MLInlineAdvisor::getInitialFunctionLevel(const Function &F) const {
- return CG.lookup(F) ? FunctionLevels.at(CG.lookup(F)) : 0;
+ return FunctionLevels.find(CG.lookup(F)) != FunctionLevels.end()
+ ? FunctionLevels.at(CG.lookup(F))
+ : 0;
}
void MLInlineAdvisor::onPassEntry(LazyCallGraph::SCC *CurSCC) {
|
I'm happy to, but I've spent a few hours trying to create one with no luck. I don't have enough understanding of co_await nor CG to understand the root cause of the inconsistency. It appears that there will be a cleanup function that exists in CG but not FunctionLevels. Any hints? |
Co-authored-by: Ellis Hoag <[email protected]>
Have you tried |
|
I put the task as low priority and after three month the file that used to trigger the error compiles fine now. Maybe it's because we moved from LLVM 17 -> 19, maybe the file structure changed. Either way I'm closing this and won't reopen until I can spot similar bugs again. |
4 participants
When building our internal code, one of the function annotated with
co_await(c++20) will cause inconsistancy: The function exists in CG but not FunctionLevels.This patch fixes it by using FunctionLevels only.