[BOLT][AArch64] Compensate for missing code markers #151060
Conversation
Code written in assembly can have missing code markers. In BOLT, we can compensate by recognizing that a function entry point should start a code sequence. Seen such code in lua jit library.
maksfb
requested review from
aaupov,
ayermolo,
paschalis-mpeis,
rafaelauler and
yota9
as code owners
|
@llvm/pr-subscribers-bolt Author: Maksim Panchenko (maksfb) ChangesCode written in assembly can have missing code markers. In BOLT, we can compensate by recognizing that a function entry point should start a code sequence. Seen such code in lua jit library. Full diff: https://github.com/llvm/llvm-project/pull/151060.diff 2 Files Affected:
diff --git a/bolt/lib/Rewrite/RewriteInstance.cpp b/bolt/lib/Rewrite/RewriteInstance.cpp
index 9f243a1366928..fe4a23cc01382 100644
--- a/bolt/lib/Rewrite/RewriteInstance.cpp
+++ b/bolt/lib/Rewrite/RewriteInstance.cpp
@@ -896,6 +896,20 @@ void RewriteInstance::discoverFileObjects() {
continue;
MarkerSymType MarkerType = BC->getMarkerType(SymInfo.Symbol);
+
+ // Treat ST_Function as code.
+ Expected<object::SymbolRef::Type> TypeOrError = SymInfo.Symbol.getType();
+ consumeError(TypeOrError.takeError());
+ if (TypeOrError && *TypeOrError == SymbolRef::ST_Function) {
+ if (IsData) {
+ Expected<StringRef> NameOrError = SymInfo.Symbol.getName();
+ consumeError(NameOrError.takeError());
+ BC->errs() << "BOLT-WARNING: function symbol " << *NameOrError
+ << " lacks code marker\n";
+ }
+ MarkerType = MarkerSymType::CODE;
+ }
+
if (MarkerType != MarkerSymType::NONE) {
SortedMarkerSymbols.push_back(MarkerSym{SymInfo.Address, MarkerType});
LastAddr = SymInfo.Address;
diff --git a/bolt/test/AArch64/missing-code-marker.s b/bolt/test/AArch64/missing-code-marker.s
new file mode 100644
index 0000000000000..591c9abd34c23
--- /dev/null
+++ b/bolt/test/AArch64/missing-code-marker.s
@@ -0,0 +1,26 @@
+## Check that llvm-bolt is able to recover a missing code marker.
+
+# RUN: %clang %cflags %s -o %t.exe -nostdlib -fuse-ld=lld -Wl,-q
+# RUN: llvm-bolt %t.exe -o %t.bolt 2>&1 | FileCheck %s
+
+# CHECK: BOLT-WARNING: function symbol foo lacks code marker
+
+.text
+.balign 4
+
+.word 0
+
+## Function foo starts immediately after a data object and does not have
+## a matching "$x" symbol to indicate the start of code.
+.global foo
+.type foo, %function
+foo:
+ .word 0xd65f03c0
+.size foo, .-foo
+
+.global _start
+.type _start, %function
+_start:
+ bl foo
+ ret
+.size _start, .-_start
|
There was a problem hiding this comment.
Thanks Maksim. We've encountered this issue a few of times with some JITs.
It looks good, and I'm happy with the warning emitted.
BTW, before applying your patch, I though BOLT would present an error when encountering code that was missing a marker, but that isn't the case with the added test?
AFAIU, the error is ultimately on the source code side (the JIT's codegen). So I am wondering whether colleagues would prefer to guard this behaviour behind a flag as a way of encouraging fixes to those projects. I'll follow-up if that's the case.
|
Thanks for reviews.
I haven't seen the error regarding the missing marker. In my case, the error came later from JITLink after a chain of events caused by treating code as data.
Lua JIT has its own assembler and most likely it's the cause of the missing marker. On the BOLT side, if we refactor the code that prints user messages, we can have something like |
4 participants
Code written in assembly can have missing code markers. In BOLT, we can compensate by recognizing that a function entry point should start a code sequence.
Seen such code in lua jit library.