perf annotate gets confused by thumb addresses

[Dr. David Alan Gilbert]

Using linux-linaro-tools-2.6.35-1008 (either the package or a version of the perf tools I built from the linux-linaro-next tree at commit 2140f30a928b2fc72af85257b15725ddc4144201 )

I've rebuilt a libc with a thumb2 routine (memset) and perf annotate (either the command or the newt gui version) is starting to disassemble it at an odd (i.e. +1 address); I don't think it understands the use of odd addresses to denote thumb functions:

------------------------------------------------
 Percent | Source code & Disassembly of neon-memset.so
------------------------------------------------
         :
         :
         :
         : Disassembly of section .text:
         :
         : 0005e5b1 <memset+0x1>:
    0.14 : 5e5b1: f246 10b1 movw r0, #25009 ; 0x61b1
    0.00 : 5e5b5: 1ff0 subs r0, r6, #7

The 5e5b1 is the right address as per readelf:
 13022: 0005e5b1 0 FUNC GLOBAL DEFAULT 12 memset

maybe the right answer is a fix up in dso__load_sym in util/symbol.c

but I just fixed up the addresses passed to objdump in hist_entry__annotate which seems to work.

Dave

[Dr. David Alan Gilbert (davidgil-uk)]

Fix is in upstream 2.6.38-rc3, in the Ubuntu built linux-tools-2.6.38-4 and should land in a linaro-linux-tools just as soon as a 2.6.38 lands

[Mounir Bsaibes (mounir-bsaibes)]

I have set the prioirty to medium since the fix is committed.

[Dr. David Alan Gilbert (davidgil-uk)]

This is fixed in linux-linaro-tools-2.6.38-1000