Message ID | AM5PR0802MB26107011663DF719F014F84183DC0@AM5PR0802MB2610.eurprd08.prod.outlook.com |
---|---|
State | New |
Headers | show |
On 27/06/17 16:38, Wilco Dijkstra wrote: > > > ping > > > From: Wilco Dijkstra > Sent: 10 November 2016 17:19 > To: GCC Patches > Cc: nd > Subject: [PATCH][ARM] Improve max_insns_skipped logic > > Improve the logic when setting max_insns_skipped. Limit the maximum > size of IT > to MAX_INSN_PER_IT_BLOCK as otherwise multiple IT instructions are needed, > increasing codesize. Given 4 works well for Thumb-2, use the same > limit for ARM > for consistency. > > ChangeLog: > 2016-11-04 Wilco Dijkstra <wdijkstr@arm.com> > > * config/arm/arm.c (arm_option_params_internal): Improve > setting of > max_insns_skipped. > -- > > diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c > index > f046854e9665d54911616fc1c60fee407188f7d6..29e8d1d07d918fbb2a627a653510dfc8587ee01a > 100644 > --- a/gcc/config/arm/arm.c > +++ b/gcc/config/arm/arm.c > @@ -2901,20 +2901,12 @@ arm_option_params_internal (void) > targetm.max_anchor_offset = TARGET_MAX_ANCHOR_OFFSET; > } > > - if (optimize_size) > - { > - /* If optimizing for size, bump the number of instructions that we > - are prepared to conditionally execute (even on a StrongARM). */ > - max_insns_skipped = 6; > + /* Increase the number of conditional instructions with -Os. */ > + max_insns_skipped = optimize_size ? 4 : > current_tune->max_insns_skipped; > > - /* For THUMB2, we limit the conditional sequence to one IT > block. */ > - if (TARGET_THUMB2) > - max_insns_skipped = arm_restrict_it ? 1 : 4; > - } > - else > - /* When -mrestrict-it is in use tone down the if-conversion. */ > - max_insns_skipped = (TARGET_THUMB2 && arm_restrict_it) > - ? 1 : current_tune->max_insns_skipped; > + /* For THUMB2, we limit the conditional sequence to one IT block. */ > + if (TARGET_THUMB2) > + max_insns_skipped = MIN (max_insns_skipped, MAX_INSN_PER_IT_BLOCK); I like the simplifications in the selection logic here :) However, changing the value for ARM from 6 to 4 looks a bit arbitrary to me. There's probably a reason why default values for ARM and Thumb-2 are different (maybe not a good one) and I'd rather not change it without some code size data measurements. So I'd rather not let that hold this cleanup patch though, so this is ok (assuming a normal bootstrap and testing cycle) without changing the 6 to a 4 and you can propose a change to 4 as a separate patch that can be discussed on its own. Thanks, Kyrill > } > > /* True if -mflip-thumb should next add an attribute for the default >
Kyrill Tkachov wrote: > I like the simplifications in the selection logic here :) > However, changing the value for ARM from 6 to 4 looks a bit arbitrary to me. > There's probably a reason why default values for ARM and Thumb-2 are > different > (maybe not a good one) and I'd rather not change it without some code > size data measurements. To quote myself from the thread: Long conditional sequences are slow on modern cores - the value 6 for max_insns_skipped is a few decades out of date as it was meant for ARM2! Even with -Os the performance loss for larger values is not worth the small codesize gain (there are many better options to reduce codesize that actually improve performance at the same time). So using the same code generation heuristics for ARM and Thumb-2 is a good idea. A simple codesize comparison on CSiBE shows using 4 rather than 6 for max_insns_skipped is just 0.07% larger on ARM with -Os. So it's not obvious that increasing max_insns_skipped in -Os is a useful codesize optimization... >So I'd rather not let that hold this cleanup patch though, so this is ok > (assuming a normal bootstrap and testing cycle) without changing the 6 > to a 4 > and you can propose a change to 4 as a separate patch that can be > discussed on its own. Based on the above is that really needed? What specific problem do you expect to occur with the value 4? Wilco
On 05/09/17 11:32, Wilco Dijkstra wrote: > Kyrill Tkachov wrote: > >> I like the simplifications in the selection logic here :) >> However, changing the value for ARM from 6 to 4 looks a bit arbitrary to me. >> There's probably a reason why default values for ARM and Thumb-2 are >> different >> (maybe not a good one) and I'd rather not change it without some code >> size data measurements. > To quote myself from the thread: > > Long conditional sequences are slow on modern cores - the value 6 for > max_insns_skipped is a few decades out of date as it was meant for ARM2! > Even with -Os the performance loss for larger values is not worth the > small codesize gain (there are many better options to reduce codesize > that actually improve performance at the same time). So using the same > code generation heuristics for ARM and Thumb-2 is a good idea. > > A simple codesize comparison on CSiBE shows using 4 rather than 6 for > max_insns_skipped is just 0.07% larger on ARM with -Os. So it's not > obvious that increasing max_insns_skipped in -Os is a useful codesize > optimization... > >> So I'd rather not let that hold this cleanup patch though, so this is ok >> (assuming a normal bootstrap and testing cycle) without changing the 6 >> to a 4 >> and you can propose a change to 4 as a separate patch that can be >> discussed on its own. > Based on the above is that really needed? What specific problem do you > expect to occur with the value 4? Nothing in particular, thanks for providing some numbers. I think unifying the heuristics for ARM and Thumb2 makes sense and a 0.07% code size hit on ARM (i.e. not code-size-optimised Thumb-2) mode seems acceptable to me if it allows more performant code. So this is ok then. Thanks, Kyrill > Wilco > >
diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c index f046854e9665d54911616fc1c60fee407188f7d6..29e8d1d07d918fbb2a627a653510dfc8587ee01a 100644 --- a/gcc/config/arm/arm.c +++ b/gcc/config/arm/arm.c @@ -2901,20 +2901,12 @@ arm_option_params_internal (void) targetm.max_anchor_offset = TARGET_MAX_ANCHOR_OFFSET; } - if (optimize_size) - { - /* If optimizing for size, bump the number of instructions that we - are prepared to conditionally execute (even on a StrongARM). */ - max_insns_skipped = 6; + /* Increase the number of conditional instructions with -Os. */ + max_insns_skipped = optimize_size ? 4 : current_tune->max_insns_skipped; - /* For THUMB2, we limit the conditional sequence to one IT block. */ - if (TARGET_THUMB2) - max_insns_skipped = arm_restrict_it ? 1 : 4; - } - else - /* When -mrestrict-it is in use tone down the if-conversion. */ - max_insns_skipped = (TARGET_THUMB2 && arm_restrict_it) - ? 1 : current_tune->max_insns_skipped; + /* For THUMB2, we limit the conditional sequence to one IT block. */ + if (TARGET_THUMB2) + max_insns_skipped = MIN (max_insns_skipped, MAX_INSN_PER_IT_BLOCK); } /* True if -mflip-thumb should next add an attribute for the default
ping From: Wilco Dijkstra Sent: 10 November 2016 17:19 To: GCC Patches Cc: nd Subject: [PATCH][ARM] Improve max_insns_skipped logic Improve the logic when setting max_insns_skipped. Limit the maximum size of IT to MAX_INSN_PER_IT_BLOCK as otherwise multiple IT instructions are needed, increasing codesize. Given 4 works well for Thumb-2, use the same limit for ARM for consistency. ChangeLog: 2016-11-04 Wilco Dijkstra <wdijkstr@arm.com> * config/arm/arm.c (arm_option_params_internal): Improve setting of max_insns_skipped. --