On Tue, Mar 31, 2020 at 12:46:28AM -0700, Christoph Hellwig wrote:
> On Sat, Mar 28, 2020 at 08:51:56AM -0700, Matthew Wilcox wrote:
> > By splitting iomap_apply into __iomap_apply and an inline iomap_apply,
> > we convert the call to 'actor' into a direct function call. I haven't
> > done any performance measurements, but given the current costs of indirect
> > function calls, this would seem worthwhile to me?
>
> Hmm. Given that emount of compiler stupidity we are dealing with did
> you at least look at the assembly output to see if this actually removes
> the indirect call? I wouldn't be quite sure.
If it does get inlined, then the compiler does it:
b9d: e8 ae fe ff ff callq a50 <iomap_page_mkwrite_actor>
If not, then the compiler emits a function in each file called iomap_apply
which makes an indirect call. So s/inline/__always_inline/ in the original
patch.
before:
text data bss dec hex filename
5314 4648 0 9962 26ea fs/iomap/trace.o
1050 72 0 1122 462 fs/iomap/apply.o
17316 636 224 18176 4700 fs/iomap/buffered-io.o
4773 76 0 4849 12f1 fs/iomap/direct-io.o
1335 28 0 1363 553 fs/iomap/fiemap.o
1928 28 0 1956 7a4 fs/iomap/seek.o
1394 8 0 1402 57a fs/iomap/swapfile.o
after:
text data bss dec hex filename
5314 4648 0 9962 26ea fs/iomap/trace.o
722 72 0 794 31a fs/iomap/apply.o
18784 636 224 19644 4cbc fs/iomap/buffered-io.o
5169 76 0 5245 147d fs/iomap/direct-io.o
2093 28 0 2121 849 fs/iomap/fiemap.o
2467 28 0 2495 9bf fs/iomap/seek.o
1664 8 0 1672 688 fs/iomap/swapfile.o
33110 to 36213 bytes of text. So not free. Worthwhile?
