635966 - elfhack fails with llvm LTO

Status: RESOLVED FIXED

(Firefox Build System :: General, defect)

Description

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

Attachment: build/unix/elfhack/test.o

A build with both LTO and elfhack fails at

/home/espindola/tmpfs/objdir-lto-clang/build/unix/elfhack/elfhack -b test.so
test.so: No gain. Aborting

Comment 1

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

Attachment: build/unix/elfhack/test.so

Comment 2

[Mike Hommey [:glandium]]

Turns out it's the same problem as the other bug

Comment 3

[Mike Hommey [:glandium]]

I have to retract, it looks like it's different. If I remove the relro section, it still says No gain.

Comment 4

[Mike Hommey [:glandium]]

Ah, this is an interesting case where test.so is too optimized and doesn't exhibit the patterns that are expected to show up from the source... In the present case, there are no relocations...

Comment 5

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

Attachment: Use -O0 and pass LDFLAGS

Comment 6

[Mike Hommey [:glandium]]

Attachment: Avoid elfhack failure with llvm

There are two different things mixed in this patch, but they are two one-liners:
- the rule to build the dummy binary works for gcc but not for clang because of missing flags
- the test library is over-optimized by clang/gold, and ends up not containing most of the code from test.c, making the testcase not contain what we expect it to. Forcing to build at the lowest optimization level avoids surprises.

Comment 7

[Mike Hommey [:glandium]]

Comment on attachment 514296 [details] [diff] [review]
Avoid elfhack failure with llvm

See comment 6. No risk.

Comment 8

[Mike Shaver (:shaver -- probably not reading bugmail closely)]

Comment on attachment 514296 [details] [diff] [review]
Avoid elfhack failure with llvm

if this causes so much as a blip on the tinderbox radar, it comes out until after 4.

Comment 9

[Mike Hommey [:glandium]]

http://hg.mozilla.org/mozilla-central/rev/494662b968c2