RFA: another patch to fix PR61360
Tue Sep 23 06:07:00 GMT 2014
On Tue, Sep 23, 2014 at 3:26 AM, Vladimir Makarov <email@example.com> wrote:
> The previous patch to solve PR61360 fixed the problem in IRA (it was
> easier for me to do as I know the code well)
> Although imo it was an ok fix, Richard expressed concerns with the patch
> and the practice to have different enable attribute values depending on the
> current pass.
> I don't understand why "x,m" alternative is better to "x,r" and "x,r"
> should be disabled. Even if the path from general regs to sse regs is slow
> (usually such slow path is implemented internally by micro-architecture
> through cache). "x,r" alternative results in only smaller insns (including
> number of insns) with probably the same time for the movement. So "x,r"
> should be at least no slower, insn cache should have more locality, and less
> overhead for decoding/translating insns.
> Here I propose another solution avoiding to have different enable
> attribute values.
> The patch was successfully bootstrapped on x86/x86-64 and tested with and
> without -march=amdfam10 (actually the patch results in 2 less failures when
> -march=amdfam10 were used).
> Uros, is i386.md change ok for the trunk?
I don't think so. This would be a regression, since 4.8 (and later
versions until Richard's patch) were able to handle this functionality
just fine. Please also note that there are a couple of other patterns
with the same problem, that is using ("nonimmediate_operand" "m")
Please see PR 60704, comment 7 . If LRA is able to fixup
("nonimmediate_operand" "m") to a memory from a register, then other
RTL infrastructure should also be updated for this functionality. IMO,
recog should be fixed/enhanced, so pseudo registers will also satisfy
("nonimmediate_operand" "m") constraint before LRA.
More information about the Gcc-patches