[scudo] Move ALWAYS_INLINE up in tsd_shared's ScopedTSD ctor

[fabio-d]

Fix for performance regression introduced by #80061 that slowed
down Fuchsia's MallocFree microbenchmark by 3.5 - 8%

[llvmbot]

@llvm/pr-subscribers-compiler-rt-sanitizer

Author: Fabio D'Urso (fabio-d)

Changes

This PR should fix the performance regression that was introduced by #80061

---

Full diff: https://github.com/llvm/llvm-project/pull/81982.diff

1 Files Affected:

• (modified) compiler-rt/lib/scudo/standalone/tsd_shared.h (+2-2)

diff --git a/compiler-rt/lib/scudo/standalone/tsd_shared.h b/compiler-rt/lib/scudo/standalone/tsd_shared.h
index 0ac2be3d4c294d..db5f6ec2094538 100644
--- a/compiler-rt/lib/scudo/standalone/tsd_shared.h
+++ b/compiler-rt/lib/scudo/standalone/tsd_shared.h
@@ -29,7 +29,7 @@ struct TSDRegistrySharedT {
   using ThisT = TSDRegistrySharedT<Allocator, TSDsArraySize, DefaultTSDCount>;
 
   struct ScopedTSD {
-    ScopedTSD(ThisT &TSDRegistry) {
+    ALWAYS_INLINE ScopedTSD(ThisT &TSDRegistry) {
       CurrentTSD = TSDRegistry.getTSDAndLock();
       DCHECK_NE(CurrentTSD, nullptr);
     }
@@ -133,7 +133,7 @@ struct TSDRegistrySharedT {
   }
 
 private:
-  ALWAYS_INLINE TSD<Allocator> *getTSDAndLock() NO_THREAD_SAFETY_ANALYSIS {
+  TSD<Allocator> *getTSDAndLock() NO_THREAD_SAFETY_ANALYSIS {
     TSD<Allocator> *TSD = getCurrentTSD();
     DCHECK(TSD);
     // Try to lock the currently associated context.

[ChiaHungDuan]

Can you add more details about the performance regression in the description? Besides, please do the related change to the tsd_exclusive as well. Thanks

[ChiaHungDuan]

It's better to mention that's the micro benchmarks on Fuchsia so that we will know more context when look back.

BTW, let's do the change for tsd_exclusive as well

[ChiaHungDuan]

It's better to mention that's the micro benchmarks on Fuchsia so that we will know more context when look back.

BTW, let's do the change for tsd_exclusive as well

Sorry my bad. You have updated the tsd_exclusive.

[fabio-d]

(for future reference, the issue was https://g-issues.fuchsia.dev/issues/323911094)