Patchwork [mips] Patch for mips generic scheduler

login
register
mail settings
Submitter Steve Ellcey
Date May 20, 2013, 10:19 p.m.
Message ID <9714dd4b-6beb-4007-9bf4-c9ee80ab5ad6@BAMAIL02.ba.imgtec.org>
Download mbox | patch
Permalink /patch/245124/
State New
Headers show

Comments

Steve Ellcey - May 20, 2013, 10:19 p.m.
While looking at some matrix code I noticed that the generic mips scheduler
was putting out a bunch of integer madd instructions in a row and that
I got better performance (on 74kc and proAptiv) when they were spread out.

I was wondering what folks thought of this change to specify that the
integer madd instruction uses the alu and the imuldiv functional units?

Tested on 74kc and proAptive with no regressions and better performance.

OK to checkin?

Steve Ellcey
sellcey@imgtec.com



2013-05-20  Steve Ellcey  <sellcey@imgtec.com>

	* config/mips/generic.md (generic_imul): Remove imadd.
	(generic_imadd): New.
Richard Sandiford - May 21, 2013, 6:55 p.m.
"Steve Ellcey " <sellcey@imgtec.com> writes:
> While looking at some matrix code I noticed that the generic mips scheduler
> was putting out a bunch of integer madd instructions in a row and that
> I got better performance (on 74kc and proAptiv) when they were spread out.
>
> I was wondering what folks thought of this change to specify that the
> integer madd instruction uses the alu and the imuldiv functional units?
>
> Tested on 74kc and proAptive with no regressions and better performance.

Hmm, but generic.md is very much legacy and shouldn't be used for
vaguely modern cores.  Even something like -mips32 is supposed to avoid
the generic scheduler; it should use the 4k scheduler instead.
What options were you using to trigger it?

generic.md is actually inherited by some of the very old, pre-DFA schedulers,
including those that were probably used with the old PMC cores.  I'm not
even sure how much generic.md was ever supposed to be a good "all-round"
scheduling description, akin to the modern -mtune=generic on x86 targets.
I'm a bit reluctant to change it at this stage.

It might be worth having a new "generic" scheduler that's supposed to be
a good compromise for modern cores though.  Or, more simply, we could just
change:

MIPS_CPU ("mips32", PROCESSOR_4KC, 32, PTF_AVOID_BRANCHLIKELY)
MIPS_CPU ("mips32r2", PROCESSOR_M4K, 33, PTF_AVOID_BRANCHLIKELY)
MIPS_CPU ("mips64", PROCESSOR_5KC, 64, PTF_AVOID_BRANCHLIKELY)
/* ??? For now just tune the generic MIPS64r2 for 5KC as well.   */
MIPS_CPU ("mips64r2", PROCESSOR_5KC, 65, PTF_AVOID_BRANCHLIKELY)

to tune for other processors instead, if you don't think 4kc, etc. are
representative enough.

Richard
Steve Ellcey - May 21, 2013, 8:47 p.m.
On Tue, 2013-05-21 at 19:55 +0100, Richard Sandiford wrote:

> Hmm, but generic.md is very much legacy and shouldn't be used for
> vaguely modern cores.  Even something like -mips32 is supposed to avoid
> the generic scheduler; it should use the 4k scheduler instead.
> What options were you using to trigger it?

I am just doing a default compilation of my mips-mti-linux-gnu target.
This defaults to mips32r2.  I just tried adding -mtune=mips32r2 and that
made no difference, -mtune=4k also makes no difference, but -mtune=4kc
does affect the output.

> It might be worth having a new "generic" scheduler that's supposed to be
> a good compromise for modern cores though.  Or, more simply, we could just
> change:
> 
> MIPS_CPU ("mips32", PROCESSOR_4KC, 32, PTF_AVOID_BRANCHLIKELY)
> MIPS_CPU ("mips32r2", PROCESSOR_M4K, 33, PTF_AVOID_BRANCHLIKELY)
> MIPS_CPU ("mips64", PROCESSOR_5KC, 64, PTF_AVOID_BRANCHLIKELY)
> /* ??? For now just tune the generic MIPS64r2 for 5KC as well.   */
> MIPS_CPU ("mips64r2", PROCESSOR_5KC, 65, PTF_AVOID_BRANCHLIKELY)
> 
> to tune for other processors instead, if you don't think 4kc, etc. are
> representative enough.

Hm, I think the problem may be that mips32r2 defaults to PROCESSOR_M4K
and mips32 defaults to PROCESSOR_4KC.  I don't see any special scheduler
for m4k.  Is there supposed to be a scheduler for m4k?

Steve Ellcey
Richard Sandiford - May 22, 2013, 6:30 a.m.
Steve Ellcey <sellcey@imgtec.com> writes:
>> It might be worth having a new "generic" scheduler that's supposed to be
>> a good compromise for modern cores though.  Or, more simply, we could just
>> change:
>> 
>> MIPS_CPU ("mips32", PROCESSOR_4KC, 32, PTF_AVOID_BRANCHLIKELY)
>> MIPS_CPU ("mips32r2", PROCESSOR_M4K, 33, PTF_AVOID_BRANCHLIKELY)
>> MIPS_CPU ("mips64", PROCESSOR_5KC, 64, PTF_AVOID_BRANCHLIKELY)
>> /* ??? For now just tune the generic MIPS64r2 for 5KC as well.   */
>> MIPS_CPU ("mips64r2", PROCESSOR_5KC, 65, PTF_AVOID_BRANCHLIKELY)
>> 
>> to tune for other processors instead, if you don't think 4kc, etc. are
>> representative enough.
>
> Hm, I think the problem may be that mips32r2 defaults to PROCESSOR_M4K
> and mips32 defaults to PROCESSOR_4KC.  I don't see any special scheduler
> for m4k.  Is there supposed to be a scheduler for m4k?

Oops -- only if someone submitted one :-)  So we should definitely
change the mips32r2 entry.  I'd suggest one of PROCESSOR_24KF* or
PROCESSOR_74KF*, so that we get the FPU scheduling, but I don't know
which would be more representative of the general case.  TUNE_74*
has quite a lot of special code associated with it, whereas TUNE_24*
sets TUNE_MACC_CHAINS, which might overemphasise the use of MADD
for 74k targets when -mimadd is used.

But any choice is going to be a compromise.  A patch to do either
is preapproved.

Thanks,
Richard

Patch

diff --git a/gcc/config/mips/generic.md b/gcc/config/mips/generic.md
index 6051630..f187e0b 100644
--- a/gcc/config/mips/generic.md
+++ b/gcc/config/mips/generic.md
@@ -47,9 +47,13 @@ 
   "imuldiv*3")
 
 (define_insn_reservation "generic_imul" 17
-  (eq_attr "type" "imul,imul3,imadd")
+  (eq_attr "type" "imul,imul3")
   "imuldiv*17")
 
+(define_insn_reservation "generic_imadd" 17
+  (eq_attr "type" "imadd")
+  "(alu+imuldiv)*17")
+
 (define_insn_reservation "generic_idiv" 38
   (eq_attr "type" "idiv")
   "imuldiv*38")