Re: [PATCH 1/X, i386, PR54232] Enable EBX for x86 in 32bits PIC code

[Jeff Law <law at redhat dot com>]

On 10/14/14 07:00, Jakub Jelinek wrote:
> For the first two, I think (and said it before already) that the current
> model of emitting set_got from a target hook during RA can't work, as there
> can be calls in the prologue, and the prologue is inserted before the
> set_got in that case.  I really think the RA should in that case just tell
> the backend whether and in which register it wants to have the PIC register
> loaded upon start of the function, and it should be emit prologue pass
> that should arrange for that.
That works for me -- I've been encouraging Intel to push emitting the 
PIC setup further and further back in the RTL pipeline.  Their early 
patches had it very early in the RTL pipeline and naturally there was 
fallout/bleedout in various places in the optimizers.

I don't see much value in emitting the PIC setup prior to allocation, 
all I see is problems.


> As for the code quality, either some RA improvements are needed, or
> postreload must be able to fix it up, or hardreg propagation (though,
> cprop_hardreg is forward propagation rather than backwards, right?).
> Better before prologue is emitted though, because that will save/restore
> the badly chosen hard reg too.
RA improvements are the way to go -- however, my understanding is that 
overall the code is better now than it was before Intel's changes, so I 
don't consider the performance side as a blocker for this code.

The biggest performance issue identified so far is rematerialization. 
The initial patch Intel sent to me was totally unacceptable as it just 
hacked off optimizers rather than digging into the guts of why IRA/LRA 
was unable to sanely rematerialize the PIC register value.

jeff