771683 - Mark some functions MOZ_ALWAYS_INILINE

Status: RESOLVED FIXED

(Core :: XPConnect, defect)

Description

[Rafael Ávila de Espíndola (:espindola) (not reading bugmail)]

Attachment: use MOZ_ALWAYS_INLINE

Comparing gcc 4.2 performance with clang I found out that forcing the inlining of these functions improves clang's performance in dromaeo. Looking at it a bit, the reasons seems to be in addition to the call overhead:

* rv becomes a register. These two functions take a pointer to rv an the callers have it on the stack. This allows the compiler to simplify things like

if (foo)
  *rv = bar;

to a simple phi and delete the BB once they are inlined.

* ~XPCLazyCallContext has:

        if (mCcxToDestroy)
            mCcxToDestroy->~XPCCallContext();

By inlining the compiler realizes that nothing is setting mCcxToDestroy and removes the check an the call. These also seems to help other compiler that are more conservative that gcc 4.2. Talos is still running, but the comparison is at bit.ly/Ps8qxz

Comment 1

[Rafael Ávila de Espíndola (:espindola) (not reading bugmail)]

Forgot to add: compiling the same preprocessed dom_quickstubs.cpp (so no clang only features like final) with and without this patch, with gcc 4.2 and clang 159509 the sizes are:

-rw-r--r--  1 espindola  staff  1197428  6 Jul 18:23 dom_quickstubs-clang-fast.o
-rw-r--r--  1 espindola  staff  1156736  6 Jul 18:23 dom_quickstubs-clang-master.o
-rw-r--r--  1 espindola  staff  1316792  6 Jul 18:22 dom_quickstubs-gcc-fast.o
-rw-r--r--  1 espindola  staff  1316832  6 Jul 18:22 dom_quickstubs-gcc-master.o

Note how the size with gcc 4.2 decreases with this patch. I guess it was already inlining every anyway and does it earlier with this patch
.

Comment 2

[(no longer active)]

Comment on attachment 639827 [details] [diff] [review]
use MOZ_ALWAYS_INLINE

This looks very sane to me, but I would also like us to have Boris' take on this as well.

Comment 3

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 639827 [details] [diff] [review]
use MOZ_ALWAYS_INLINE

r=me.  I can totally see how not inlining these would have a big effect!

We should take a good look at our new DOM bindings and what clang actually generates for them codewise...

Comment 4

[Rafael Ávila de Espíndola (:espindola) (not reading bugmail)]

https://tbpl.mozilla.org/?tree=Mozilla-Inbound&rev=51e02ddd7482

Comment 5

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/51e02ddd7482

Comment 6

[Bobby Holley (:bholley)]

FWIW, I can definitely attest to these functions being hot. See bug 622301 comment 15.

Comment 7

[(no longer active)]

Rafael, it would be really interesting if you could do another try push to compare the perf numbers with this fix being in.  Thanks!

Comment 8

[Rafael Ávila de Espíndola (:espindola) (not reading bugmail)]

I did:

http://bit.ly/MgFlAl

traverse.html is fixed, I am looking at modify.html. It seems to be in isalloc_validate.