Re: About the effect of flag_gcse_las option

[Steven Bosscher <stevenb dot gcc at gmail dot com>]

On Fri, Nov 2, 2012 at 11:01 AM, Bin.Cheng wrote:
> Hi Steven,
> You mentioned the flag_gcse_las may be very useful for RISC machines in
> thread : http://gcc.gnu.org/ml/gcc-patches/2012-10/msg00031.html
>
> I took some time to understand the code and think it should be workable with
> hoist. However, I measured code size for CSiBE with "-fgcse-las" enabled but
> found no benefit at all.
>
> Do you have some background information and the effect about this flag?

The "las" in flag_gcse_las is for "load after store". With -fgcse-las,
RTL-PRE tries to eliminate loads from locations for which the last
stored value is known. This used to happen a lot (and probably still
does) on POWER when converting floats to ints and vice versa, because
float-to-int and int-to-float conversions have to go through memory on
POWER and this target detail isn't exposed in GIMPLE.

In general, -fgcse-las can help eliminate memory operations that
aren't exposed in GIMPLE but are visible in RTL before register
allocation. This also applies e.g. to variables pushed to the stack by
expand, or loads from constant pools if CONSTANT_POOL_ADDRESS_P could
for some reason not be set on the load address.

Finally, -fgcse-las eliminates partial-partial cases that tree-ssa-pre
only catches if flag_tree_partial_pre is set (RTL PRE is LCM, which
always eliminates the partial-partial cases). You wouldn't see this in
code size tests because RTL PRE is not enabled at -Os.

So the pass is helpful if you have target idiosyncrasies that are not
visible in GIMPLE (most RISC targets have one or more such cases), and
you're optimizing for code speed.

For code hoisting, I don't expect -fgcse-las to do much. Cases of
load-after-store that survive to RTL are probably eliminated by
pass_cse already.

Ciao!
Steven