* [PATCH][MIPS] Don't split shifts by default for MIPS16.
@ 2016-05-24 15:49 Robert Suchanek
2016-06-09 15:46 ` Maciej W. Rozycki
0 siblings, 1 reply; 2+ messages in thread
From: Robert Suchanek @ 2016-05-24 15:49 UTC (permalink / raw)
To: Catherine_Moore, Matthew Fortune; +Cc: gcc-patches
Hi,
The following changes the default behaviour of shift splitting
for MIPS16 e.g. the shifts will be split only when used with
undocumented -mno-debugd option that is now switched on by default.
This appears to enable better optimization in certain cases, and hence,
giving slightly better performance.
Ok to apply?
Regards,
Robert
gcc/
* config/mips/mips.md (<optab><mode>3): Don't split shifts
when used with -mdebugd.
* config/mips/mips.opt (mdebugd): Init to 1 by default.
---
gcc/config/mips/mips.md | 1 +
gcc/config/mips/mips.opt | 2 +-
2 files changed, 2 insertions(+), 1 deletion(-)
diff --git a/gcc/config/mips/mips.md b/gcc/config/mips/mips.md
index 22f4f0b..01d7edd 100644
--- a/gcc/config/mips/mips.md
+++ b/gcc/config/mips/mips.md
@@ -5602,6 +5602,7 @@ (define_expand "<optab><mode>3"
be careful not to allocate a new register if we've reached the
reload pass. */
if (TARGET_MIPS16
+ && !TARGET_DEBUG_D_MODE
&& optimize
&& CONST_INT_P (operands[2])
&& INTVAL (operands[2]) > 8
diff --git a/gcc/config/mips/mips.opt b/gcc/config/mips/mips.opt
index 53feb23..b6c839d 100644
--- a/gcc/config/mips/mips.opt
+++ b/gcc/config/mips/mips.opt
@@ -127,7 +127,7 @@ mdebug
Target Var(TARGET_DEBUG_MODE) Undocumented
mdebugd
-Target Var(TARGET_DEBUG_D_MODE) Undocumented
+Target Var(TARGET_DEBUG_D_MODE) Undocumented Init(1)
meb
Target Report RejectNegative Mask(BIG_ENDIAN)
--
2.8.2.396.g5fe494
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH][MIPS] Don't split shifts by default for MIPS16.
2016-05-24 15:49 [PATCH][MIPS] Don't split shifts by default for MIPS16 Robert Suchanek
@ 2016-06-09 15:46 ` Maciej W. Rozycki
0 siblings, 0 replies; 2+ messages in thread
From: Maciej W. Rozycki @ 2016-06-09 15:46 UTC (permalink / raw)
To: Robert Suchanek; +Cc: Catherine_Moore, Matthew Fortune, gcc-patches
On Tue, 24 May 2016, Robert Suchanek wrote:
> The following changes the default behaviour of shift splitting
> for MIPS16 e.g. the shifts will be split only when used with
> undocumented -mno-debugd option that is now switched on by default.
>
> This appears to enable better optimization in certain cases, and hence,
> giving slightly better performance.
Thank you for your contribution, however I have some issues with your
proposal.
First of all since TARGET_DEBUG_D_MODE is used across several places I
think this really should be split into two separate pieces, independently
reviewed.
Then we have two parts:
1. The change to the expander is probably all but obviously correct as it
complements a similar one applied to the corresponding splitters eons
ago (back in 2001), although I wonder why we need to have both the
splitters and instructions split manually in the expander in the first
place. Can you please investigatie it?
2. The other change is far from obvious since it flips the splitting
default, which has been there since forever (or the addition of MIPS16
support back in 1998). So I think it needs a justification more
elaborate than just "certain cases", and perhaps a proper command-line
option defined rather than using an obscure debugging hook. It may
also make sense to set the default dynamically from the CPU tuning
selected -- if the "certain cases" change from CPU to CPU, that is.
Can you please look into it?
Maciej
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2016-06-09 15:46 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-05-24 15:49 [PATCH][MIPS] Don't split shifts by default for MIPS16 Robert Suchanek
2016-06-09 15:46 ` Maciej W. Rozycki
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).