519831 - VMPI_lock functions should be inlined

Status: VERIFIED FIXED

(Tamarin Graveyard :: Garbage Collection (mmGC), defect)

Description

[Steven Johnson]

In older versions of Tamarin, the spinlocks for MMGC were effectively inlined. When VMPI was introduced, the locks became out-of-line. For a few benchmarks that make super-heavy use of mmfx allocation (e.g., RegExp torture tests), the extra cost of the out-of-line function calls becomes significant relative to the cost of the lock itself.

Comment 1

[Steven Johnson]

Attachment: Patch

Move the VMPI_lockXXX() functions into the appropriate header files and force inlining. Nuke the SpinLockXXX.cpp files. 

Note 1: Mac implementations now use the documented OSSpinLock calls rather than the apparently-undocumented-but-happen-to-work _spin_lock calls.

Note 2: Removed the ifdef DEBUG owner stuff for simplicity of implementation (mainly because VMPI doesn't have access to an assert macro at that point of inclusion).

Comment 2

[Steven Johnson]

Attachment: Patch #2

Additive: also force GCSpinLock to inline.

Comment 3

[Steven Johnson]

pushed both as changeset:   2657:22355a34aba3

Comment 4

[Chris Peyer]

Resolved fixed engineering / work item that has been pushed.  Setting status to verified.