[Bug ld/18414] TOC optimization bug (ignoring data deps on addis/ld -> nop/ld opt)

[hfinkel at anl dot gov]

https://sourceware.org/bugzilla/show_bug.cgi?id=18414

--- Comment #3 from Hal Finkel <hfinkel at anl dot gov> ---
(In reply to Hal Finkel from comment #2)
> (In reply to Alan Modra from comment #1)
> > True, ld will do this.  However, the optimization is relying on gcc to
> > generate code that can be optimized this way..  So I'm assuming the example
> > code was from llvm or assembly.
> 
> Yes, it was LLVM.
> 
> > 
> > gcc ensures that the optimization is valid by adding a dependency on r2 to
> > the low part of the address calculation.
> 
> Indeed, I'll need to do the same in LLVM. However, it would be nice if the
> bug could be fixed in binutils as well. Are these kinds of restrictions
> documented anywhere?

Also, just noticing that you tagged this as WONTFIX, why not? You need to scan
the instructions to find the second in the pair anyhow; why not break if you
find an instruction that writes to r2?

-- 
You are receiving this mail because:
You are on the CC list for the bug.