Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On Wed, Jul 14, 2021 at 5:11 AM guojiufu via Gcc-patches wrote: > > On 2021-07-13 23:38, Segher Boessenkool wrote: > > On Mon, Jul 12, 2021 at 08:20:14AM +0200, Richard Biener wrote: > >> On Fri, 9 Jul 2021, Segher Boessenkool wrote: > >> > Almost all targets just use Pmode, but there is no such guarantee I > >> > think, and esp. some targets that do not have machine insns for this > >> > (but want to generate different code for this anyway) can do pretty much > >> > anything. > >> > > >> > Maybe using just Pmode here is good enough though? > >> > >> I think Pmode is a particularly bad choice and I'd prefer word_mode > >> if we go for any hardcoded mode. > > > > In many important cases you use a pointer as iteration variable. > > > > Is word_mode register size on most current targets? > Had a search on the implementation, word_mode is the mode on size > BITS_PER_WORD > in MODE_INTs. Actually, when targets define Pmode and BITS_PER_WORD, > these two > macros are aligned -:), and it seems most targets define both these two > macros. > > > > >> s390x for example seems to handle > >> both SImode and DImode (but names the helper gen_doloop_si64 > >> for SImode?!). > > > > Yes, so Pmode will work fine for 390. It would be nice if we could > > allow multiple modes here, certainly. Can we? > > :), for other IVs, multiple modes are allowed to add as candidates; > while only one doloop iv is added. Comparing the supporting more > doloop IVs, it seems changing the doloop iv mode is easy relatively > for me. So, the patch is trying to update doloop iv. > > > > >> But indeed it looks like somehow querying doloop_end > >> is going to be difficult since the expander doesn't have any mode, > >> so we'd have to actually try emit RTL here. > > > > Or add a well-designed target macro for this. "Which modes do we like > > for IVs", perhaps? > > In the new patch, a target hook preferred_doloop_mode is introduced. > While > this hook is only for doloop iv at this time. > Maybe we could have preferred_iv_mode if needed. In the current code, > IVs > are free to be added in different types, and the cost model is applied > to determine which IV may be better. The iv mode would be one factor for > cost. I guess IVOPTs could directly look at PROMOTE_MODE & friends to decide whether the IV increment has extra cost due to subregs/extensions and account for that. The rest should be really driven by the cost to express IV uses, I don't think putting another target hook in here will help reduce GIGO. Richard. > > BR, > Jiufu > > > > > > > Segher
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On 2021-07-13 23:38, Segher Boessenkool wrote: On Mon, Jul 12, 2021 at 08:20:14AM +0200, Richard Biener wrote: On Fri, 9 Jul 2021, Segher Boessenkool wrote: > Almost all targets just use Pmode, but there is no such guarantee I > think, and esp. some targets that do not have machine insns for this > (but want to generate different code for this anyway) can do pretty much > anything. > > Maybe using just Pmode here is good enough though? I think Pmode is a particularly bad choice and I'd prefer word_mode if we go for any hardcoded mode. In many important cases you use a pointer as iteration variable. Is word_mode register size on most current targets? Had a search on the implementation, word_mode is the mode on size BITS_PER_WORD in MODE_INTs. Actually, when targets define Pmode and BITS_PER_WORD, these two macros are aligned -:), and it seems most targets define both these two macros. s390x for example seems to handle both SImode and DImode (but names the helper gen_doloop_si64 for SImode?!). Yes, so Pmode will work fine for 390. It would be nice if we could allow multiple modes here, certainly. Can we? :), for other IVs, multiple modes are allowed to add as candidates; while only one doloop iv is added. Comparing the supporting more doloop IVs, it seems changing the doloop iv mode is easy relatively for me. So, the patch is trying to update doloop iv. But indeed it looks like somehow querying doloop_end is going to be difficult since the expander doesn't have any mode, so we'd have to actually try emit RTL here. Or add a well-designed target macro for this. "Which modes do we like for IVs", perhaps? In the new patch, a target hook preferred_doloop_mode is introduced. While this hook is only for doloop iv at this time. Maybe we could have preferred_iv_mode if needed. In the current code, IVs are free to be added in different types, and the cost model is applied to determine which IV may be better. The iv mode would be one factor for cost. BR, Jiufu Segher
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On 2021-07-13 23:51, Segher Boessenkool wrote: On Tue, Jul 13, 2021 at 10:09:25AM +0800, guojiufu wrote: >For loop looks like: > do ; > while (n-- > 0); /* while (n-- > low); */ (This whole loop as written will be optimised away, but :-) ) At -O2, the loop is optimized away. At -O1, the loop is there. .cfi_startproc addi %r3,%r3,1 .L2: addi %r9,%r3,-1 mr %r3,%r9 andi. %r9,%r9,0xff bne %cr0,.L2 The patch v2 (https://gcc.gnu.org/pipermail/gcc-patches/2021-July/574596.html) could help it to be: .cfi_startproc addi %r3,%r3,1 mtctr %r3 .L2: addi %r3,%r3,-1 bdnz .L2 There is a patch that could mitigate "-1 +1 pair" in rtl part. https://gcc.gnu.org/g:8a15faa730f99100f6f3ed12663563356ec5a2c0 Does that solve PR67288 (and its many duplicates)? Had a test, yes, the "-1 +1" issue in PR67288 was fixed by that patch. BR, Jiufu. Segher
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On Tue, Jul 13, 2021 at 10:09:25AM +0800, guojiufu wrote: > >For loop looks like: > > do ; > > while (n-- > 0); /* while (n-- > low); */ (This whole loop as written will be optimised away, but :-) ) > There is a patch that could mitigate "-1 +1 pair" in rtl part. > https://gcc.gnu.org/g:8a15faa730f99100f6f3ed12663563356ec5a2c0 Does that solve PR67288 (and its many duplicates)? Segher
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On Mon, Jul 12, 2021 at 08:20:14AM +0200, Richard Biener wrote: > On Fri, 9 Jul 2021, Segher Boessenkool wrote: > > Almost all targets just use Pmode, but there is no such guarantee I > > think, and esp. some targets that do not have machine insns for this > > (but want to generate different code for this anyway) can do pretty much > > anything. > > > > Maybe using just Pmode here is good enough though? > > I think Pmode is a particularly bad choice and I'd prefer word_mode > if we go for any hardcoded mode. In many important cases you use a pointer as iteration variable. Is word_mode register size on most current targets? > s390x for example seems to handle > both SImode and DImode (but names the helper gen_doloop_si64 > for SImode?!). Yes, so Pmode will work fine for 390. It would be nice if we could allow multiple modes here, certainly. Can we? > But indeed it looks like somehow querying doloop_end > is going to be difficult since the expander doesn't have any mode, > so we'd have to actually try emit RTL here. Or add a well-designed target macro for this. "Which modes do we like for IVs", perhaps? Segher
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On 2021-07-13 15:09, Richard Biener wrote: On Tue, 13 Jul 2021, guojiufu wrote: On 2021-07-12 23:53, guojiufu via Gcc-patches wrote: > On 2021-07-12 22:46, Richard Biener wrote: >> On Mon, 12 Jul 2021, guojiufu wrote: >> >>> On 2021-07-12 18:02, Richard Biener wrote: >>> > On Mon, 12 Jul 2021, guojiufu wrote: >>> > >>> >> On 2021-07-12 16:57, Richard Biener wrote: >>> >> > On Mon, 12 Jul 2021, guojiufu wrote: >>> >> > >>> >> >> On 2021-07-12 14:20, Richard Biener wrote: >>> >> >> > On Fri, 9 Jul 2021, Segher Boessenkool wrote: >>> >> >> > >>> >> >> >> On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: >>> >> >> >> > I wonder if there's a way to query the target what modes the >>> >> >> >> > doloop >>> >> >> >> > pattern can handle (not being too familiar with the doloop >>> >> >> >> > code). >>> >> >> >> >>> >> >> >> You can look what modes are allowed for operand 0 of doloop_end, >>> >> >> >> perhaps? Although that is a define_expand, not a define_insn, so >>> >> >> >> it >>> >> >> >> is >>> >> >> >> hard to introspect. >>> >> >> >> >>> >> >> >> > Why do you need to do any checks besides the new type being >>> >> >> >> > able to >>> >> >> >> > represent all IV values? The original doloop IV will never >>> >> >> >> > wrap >>> >> >> >> > (OTOH if niter is U*_MAX then we compute niter + 1 which will >>> >> >> >> > become >>> >> >> >> > zero ... I suppose the doloop might still do the correct thing >>> >> >> >> > here >>> >> >> >> > but it also still will with a IV with larger type). >>> >> >> >>> >> >> The issue comes from U*_MAX (original short MAX), as you said: on >>> >> >> which >>> >> >> niter + 1 becomes zero. And because the step for doloop is -1; >>> >> >> then, on >>> >> >> larger type 'zero - 1' will be a very large number on larger type >>> >> >> (e.g. 0xff...ff); but on the original short type 'zero - 1' is a >>> >> >> small >>> >> >> value >>> >> >> (e.g. "0xff"). >>> >> > >>> >> > But for the larger type the small type MAX + 1 fits and does not >>> >> > yield >>> >> > zero so it should still work exactly as before, no? Of course you >>> >> > have to compute the + 1 in the larger type. >>> >> > >>> >> You are right, if compute the "+ 1" in the larger type it is ok, as >>> >> below >>> >> code: >>> >> ``` >>> >>/* Use type in word size may fast. */ >>> >> if (TYPE_PRECISION (ntype) < BITS_PER_WORD) >>> >> { >>> >> ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); >>> >> niter = fold_convert (ntype, niter); >>> >> } >>> >> >>> >> tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), >>> >> build_int_cst (ntype, 1)); >>> >> >>> >> >>> >> add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, >>> >> NULL, >>> >> true); >>> >> ``` >>> >> The issue of this is, this code generates more stmt for doloop.xxx: >>> >> _12 = (unsigned int) xx(D); >>> >> _10 = _12 + 4294967295; >>> >> _24 = (long unsigned int) _10; >>> >> doloop.6_8 = _24 + 1; >>> >> >>> >> if use previous patch, "+ 1" on original type, then the stmts will >>> >> looks >>> >> like: >>> >> _12 = (unsigned int) xx(D); >>> >> doloop.6_8 = (long unsigned int) _12; >>> >> >>> >> This is the reason for checking >>> >>wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype))) >>> > >>> > But this then only works when there's an upper bound on the number >>> > of iterations. Note you should not use TYPE_MAX_VALUE here but >>> > you can instead use >>> > >>> > wi::ltu_p (niter_desc->max, wi::to_widest (wi::max_value >>> > (TYPE_PRECISION (ntype), TYPE_SIGN (ntype; >>> >>> Ok, Thanks! >>> I remember you mentioned that: >>> widest_int::from (wi::max_value (TYPE_PRECISION (ntype), TYPE_SIGN >>> (ntype)), >>> TYPE_SIGN (ntype)) >>> would be better than >>> wi::to_widest (TYPE_MAX_VALUE (ntype)). >>> >>> It seems that: >>> "TYPE_MAX_VALUE (ntype)" is "NUMERICAL_TYPE_CHECK >>> (NODE)->type_non_common.maxval" >>> which do a numerical-check and return the field of maxval. And then call >>> to >>> wi::to_widest >>> >>> The other code "widest_int::from (wi::max_value (..,..),..)" calls >>> wi::max_value >>> and widest_int::from. >>> >>> I'm wondering if wi::to_widest (TYPE_MAX_VALUE (ntype)) is cheaper? >> >> TYPE_MAX_VALUE can be "suprising", it does not necessarily match the >> underlying modes precision. At some point we've tried to eliminate >> most of its uses, not sure what the situation/position is right now. > Ok, get it, thanks. > I will use "widest_int::from (wi::max_value (..,..),..)". > >> >>> > I think the -1 above comes from number of latch iterations vs. header >>> > entries - it's a common source for this kind of issues. range analysis >>> > might be able to prove that we can still merge the two adds even with >>> > the intermediate extension. >>> Yes, as you mentioned here, it relates to number of latch iterations >>> For loop looks like : while (l < n) or for (i = 0; i < n; i++)
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On Tue, 13 Jul 2021, guojiufu wrote: > On 2021-07-12 23:53, guojiufu via Gcc-patches wrote: > > On 2021-07-12 22:46, Richard Biener wrote: > >> On Mon, 12 Jul 2021, guojiufu wrote: > >> > >>> On 2021-07-12 18:02, Richard Biener wrote: > >>> > On Mon, 12 Jul 2021, guojiufu wrote: > >>> > > >>> >> On 2021-07-12 16:57, Richard Biener wrote: > >>> >> > On Mon, 12 Jul 2021, guojiufu wrote: > >>> >> > > >>> >> >> On 2021-07-12 14:20, Richard Biener wrote: > >>> >> >> > On Fri, 9 Jul 2021, Segher Boessenkool wrote: > >>> >> >> > > >>> >> >> >> On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: > >>> >> >> >> > I wonder if there's a way to query the target what modes the > >>> >> >> >> > doloop > >>> >> >> >> > pattern can handle (not being too familiar with the doloop > >>> >> >> >> > code). > >>> >> >> >> > >>> >> >> >> You can look what modes are allowed for operand 0 of doloop_end, > >>> >> >> >> perhaps? Although that is a define_expand, not a define_insn, so > >>> >> >> >> it > >>> >> >> >> is > >>> >> >> >> hard to introspect. > >>> >> >> >> > >>> >> >> >> > Why do you need to do any checks besides the new type being > >>> >> >> >> > able to > >>> >> >> >> > represent all IV values? The original doloop IV will never > >>> >> >> >> > wrap > >>> >> >> >> > (OTOH if niter is U*_MAX then we compute niter + 1 which will > >>> >> >> >> > become > >>> >> >> >> > zero ... I suppose the doloop might still do the correct thing > >>> >> >> >> > here > >>> >> >> >> > but it also still will with a IV with larger type). > >>> >> >> > >>> >> >> The issue comes from U*_MAX (original short MAX), as you said: on > >>> >> >> which > >>> >> >> niter + 1 becomes zero. And because the step for doloop is -1; > >>> >> >> then, on > >>> >> >> larger type 'zero - 1' will be a very large number on larger type > >>> >> >> (e.g. 0xff...ff); but on the original short type 'zero - 1' is a > >>> >> >> small > >>> >> >> value > >>> >> >> (e.g. "0xff"). > >>> >> > > >>> >> > But for the larger type the small type MAX + 1 fits and does not > >>> >> > yield > >>> >> > zero so it should still work exactly as before, no? Of course you > >>> >> > have to compute the + 1 in the larger type. > >>> >> > > >>> >> You are right, if compute the "+ 1" in the larger type it is ok, as > >>> >> below > >>> >> code: > >>> >> ``` > >>> >>/* Use type in word size may fast. */ > >>> >> if (TYPE_PRECISION (ntype) < BITS_PER_WORD) > >>> >> { > >>> >> ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); > >>> >> niter = fold_convert (ntype, niter); > >>> >> } > >>> >> > >>> >> tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), > >>> >> build_int_cst (ntype, 1)); > >>> >> > >>> >> > >>> >> add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, > >>> >> NULL, > >>> >> true); > >>> >> ``` > >>> >> The issue of this is, this code generates more stmt for doloop.xxx: > >>> >> _12 = (unsigned int) xx(D); > >>> >> _10 = _12 + 4294967295; > >>> >> _24 = (long unsigned int) _10; > >>> >> doloop.6_8 = _24 + 1; > >>> >> > >>> >> if use previous patch, "+ 1" on original type, then the stmts will > >>> >> looks > >>> >> like: > >>> >> _12 = (unsigned int) xx(D); > >>> >> doloop.6_8 = (long unsigned int) _12; > >>> >> > >>> >> This is the reason for checking > >>> >>wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype))) > >>> > > >>> > But this then only works when there's an upper bound on the number > >>> > of iterations. Note you should not use TYPE_MAX_VALUE here but > >>> > you can instead use > >>> > > >>> > wi::ltu_p (niter_desc->max, wi::to_widest (wi::max_value > >>> > (TYPE_PRECISION (ntype), TYPE_SIGN (ntype; > >>> > >>> Ok, Thanks! > >>> I remember you mentioned that: > >>> widest_int::from (wi::max_value (TYPE_PRECISION (ntype), TYPE_SIGN > >>> (ntype)), > >>> TYPE_SIGN (ntype)) > >>> would be better than > >>> wi::to_widest (TYPE_MAX_VALUE (ntype)). > >>> > >>> It seems that: > >>> "TYPE_MAX_VALUE (ntype)" is "NUMERICAL_TYPE_CHECK > >>> (NODE)->type_non_common.maxval" > >>> which do a numerical-check and return the field of maxval. And then call > >>> to > >>> wi::to_widest > >>> > >>> The other code "widest_int::from (wi::max_value (..,..),..)" calls > >>> wi::max_value > >>> and widest_int::from. > >>> > >>> I'm wondering if wi::to_widest (TYPE_MAX_VALUE (ntype)) is cheaper? > >> > >> TYPE_MAX_VALUE can be "suprising", it does not necessarily match the > >> underlying modes precision. At some point we've tried to eliminate > >> most of its uses, not sure what the situation/position is right now. > > Ok, get it, thanks. > > I will use "widest_int::from (wi::max_value (..,..),..)". > > > >> > >>> > I think the -1 above comes from number of latch iterations vs. header > >>> > entries - it's a common source for this kind of issues. range analysis > >>> > might be able to prove that we can still
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On 2021-07-12 23:53, guojiufu via Gcc-patches wrote: On 2021-07-12 22:46, Richard Biener wrote: On Mon, 12 Jul 2021, guojiufu wrote: On 2021-07-12 18:02, Richard Biener wrote: > On Mon, 12 Jul 2021, guojiufu wrote: > >> On 2021-07-12 16:57, Richard Biener wrote: >> > On Mon, 12 Jul 2021, guojiufu wrote: >> > >> >> On 2021-07-12 14:20, Richard Biener wrote: >> >> > On Fri, 9 Jul 2021, Segher Boessenkool wrote: >> >> > >> >> >> On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: >> >> >> > I wonder if there's a way to query the target what modes the doloop >> >> >> > pattern can handle (not being too familiar with the doloop code). >> >> >> >> >> >> You can look what modes are allowed for operand 0 of doloop_end, >> >> >> perhaps? Although that is a define_expand, not a define_insn, so it >> >> >> is >> >> >> hard to introspect. >> >> >> >> >> >> > Why do you need to do any checks besides the new type being able to >> >> >> > represent all IV values? The original doloop IV will never wrap >> >> >> > (OTOH if niter is U*_MAX then we compute niter + 1 which will >> >> >> > become >> >> >> > zero ... I suppose the doloop might still do the correct thing here >> >> >> > but it also still will with a IV with larger type). >> >> >> >> The issue comes from U*_MAX (original short MAX), as you said: on which >> >> niter + 1 becomes zero. And because the step for doloop is -1; then, on >> >> larger type 'zero - 1' will be a very large number on larger type >> >> (e.g. 0xff...ff); but on the original short type 'zero - 1' is a small >> >> value >> >> (e.g. "0xff"). >> > >> > But for the larger type the small type MAX + 1 fits and does not yield >> > zero so it should still work exactly as before, no? Of course you >> > have to compute the + 1 in the larger type. >> > >> You are right, if compute the "+ 1" in the larger type it is ok, as below >> code: >> ``` >>/* Use type in word size may fast. */ >> if (TYPE_PRECISION (ntype) < BITS_PER_WORD) >> { >> ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); >> niter = fold_convert (ntype, niter); >> } >> >> tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), >> build_int_cst (ntype, 1)); >> >> >> add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, NULL, >> true); >> ``` >> The issue of this is, this code generates more stmt for doloop.xxx: >> _12 = (unsigned int) xx(D); >> _10 = _12 + 4294967295; >> _24 = (long unsigned int) _10; >> doloop.6_8 = _24 + 1; >> >> if use previous patch, "+ 1" on original type, then the stmts will looks >> like: >> _12 = (unsigned int) xx(D); >> doloop.6_8 = (long unsigned int) _12; >> >> This is the reason for checking >>wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype))) > > But this then only works when there's an upper bound on the number > of iterations. Note you should not use TYPE_MAX_VALUE here but > you can instead use > > wi::ltu_p (niter_desc->max, wi::to_widest (wi::max_value > (TYPE_PRECISION (ntype), TYPE_SIGN (ntype; Ok, Thanks! I remember you mentioned that: widest_int::from (wi::max_value (TYPE_PRECISION (ntype), TYPE_SIGN (ntype)), TYPE_SIGN (ntype)) would be better than wi::to_widest (TYPE_MAX_VALUE (ntype)). It seems that: "TYPE_MAX_VALUE (ntype)" is "NUMERICAL_TYPE_CHECK (NODE)->type_non_common.maxval" which do a numerical-check and return the field of maxval. And then call to wi::to_widest The other code "widest_int::from (wi::max_value (..,..),..)" calls wi::max_value and widest_int::from. I'm wondering if wi::to_widest (TYPE_MAX_VALUE (ntype)) is cheaper? TYPE_MAX_VALUE can be "suprising", it does not necessarily match the underlying modes precision. At some point we've tried to eliminate most of its uses, not sure what the situation/position is right now. Ok, get it, thanks. I will use "widest_int::from (wi::max_value (..,..),..)". > I think the -1 above comes from number of latch iterations vs. header > entries - it's a common source for this kind of issues. range analysis > might be able to prove that we can still merge the two adds even with > the intermediate extension. Yes, as you mentioned here, it relates to number of latch iterations For loop looks like : while (l < n) or for (i = 0; i < n; i++) This kind of loop, the niter is used to be 'n - 1' after transformed into 'do-while' form. For this kind of loop, the max value for the number of iteration "n - 1" would be "max_value_type(n) - 1" which is wi::ltu than max_value_type. This kind of loop is already common, and we could use wi::ltu (max, max_value_type) to check. For loop looks like: do ; while (n-- > 0); /* while (n-- > low); */ The niter_desc->max will wi::eq to max_value_type, and niter would be "n", and then doloop.xx is 'n+1'. I would see how to merge these two adds safely at this point when generating doloop iv. (maybe range info, thanks! For some
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On 2021-07-12 23:53, guojiufu via Gcc-patches wrote: On 2021-07-12 22:46, Richard Biener wrote: On Mon, 12 Jul 2021, guojiufu wrote: On 2021-07-12 18:02, Richard Biener wrote: > On Mon, 12 Jul 2021, guojiufu wrote: > >> On 2021-07-12 16:57, Richard Biener wrote: >> > On Mon, 12 Jul 2021, guojiufu wrote: >> > >> >> On 2021-07-12 14:20, Richard Biener wrote: >> >> > On Fri, 9 Jul 2021, Segher Boessenkool wrote: >> >> > >> >> >> On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: >> >> >> > I wonder if there's a way to query the target what modes the doloop >> >> >> > pattern can handle (not being too familiar with the doloop code). >> >> >> >> >> >> You can look what modes are allowed for operand 0 of doloop_end, >> >> >> perhaps? Although that is a define_expand, not a define_insn, so it >> >> >> is >> >> >> hard to introspect. >> >> >> >> >> >> > Why do you need to do any checks besides the new type being able to >> >> >> > represent all IV values? The original doloop IV will never wrap >> >> >> > (OTOH if niter is U*_MAX then we compute niter + 1 which will >> >> >> > become >> >> >> > zero ... I suppose the doloop might still do the correct thing here >> >> >> > but it also still will with a IV with larger type). >> >> >> >> The issue comes from U*_MAX (original short MAX), as you said: on which >> >> niter + 1 becomes zero. And because the step for doloop is -1; then, on >> >> larger type 'zero - 1' will be a very large number on larger type >> >> (e.g. 0xff...ff); but on the original short type 'zero - 1' is a small >> >> value >> >> (e.g. "0xff"). >> > >> > But for the larger type the small type MAX + 1 fits and does not yield >> > zero so it should still work exactly as before, no? Of course you >> > have to compute the + 1 in the larger type. >> > >> You are right, if compute the "+ 1" in the larger type it is ok, as below >> code: >> ``` >>/* Use type in word size may fast. */ >> if (TYPE_PRECISION (ntype) < BITS_PER_WORD) >> { >> ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); >> niter = fold_convert (ntype, niter); >> } >> >> tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), >> build_int_cst (ntype, 1)); >> >> >> add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, NULL, >> true); >> ``` >> The issue of this is, this code generates more stmt for doloop.xxx: >> _12 = (unsigned int) xx(D); >> _10 = _12 + 4294967295; >> _24 = (long unsigned int) _10; >> doloop.6_8 = _24 + 1; >> >> if use previous patch, "+ 1" on original type, then the stmts will looks >> like: >> _12 = (unsigned int) xx(D); >> doloop.6_8 = (long unsigned int) _12; >> >> This is the reason for checking >>wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype))) > > But this then only works when there's an upper bound on the number > of iterations. Note you should not use TYPE_MAX_VALUE here but > you can instead use > > wi::ltu_p (niter_desc->max, wi::to_widest (wi::max_value > (TYPE_PRECISION (ntype), TYPE_SIGN (ntype; Ok, Thanks! I remember you mentioned that: widest_int::from (wi::max_value (TYPE_PRECISION (ntype), TYPE_SIGN (ntype)), TYPE_SIGN (ntype)) would be better than wi::to_widest (TYPE_MAX_VALUE (ntype)). It seems that: "TYPE_MAX_VALUE (ntype)" is "NUMERICAL_TYPE_CHECK (NODE)->type_non_common.maxval" which do a numerical-check and return the field of maxval. And then call to wi::to_widest The other code "widest_int::from (wi::max_value (..,..),..)" calls wi::max_value and widest_int::from. I'm wondering if wi::to_widest (TYPE_MAX_VALUE (ntype)) is cheaper? TYPE_MAX_VALUE can be "suprising", it does not necessarily match the underlying modes precision. At some point we've tried to eliminate most of its uses, not sure what the situation/position is right now. Ok, get it, thanks. I will use "widest_int::from (wi::max_value (..,..),..)". > I think the -1 above comes from number of latch iterations vs. header > entries - it's a common source for this kind of issues. range analysis > might be able to prove that we can still merge the two adds even with > the intermediate extension. Yes, as you mentioned here, it relates to number of latch iterations For loop looks like : while (l < n) or for (i = 0; i < n; i++) This kind of loop, the niter is used to be 'n - 1' after transformed into 'do-while' form. For this kind of loop, the max value for the number of iteration "n - 1" would be "max_value_type(n) - 1" which is wi::ltu than max_value_type. This kind of loop is already common, and we could use wi::ltu (max, max_value_type) to check. For loop looks like: do ; while (n-- > 0); /* while (n-- > low); */ The niter_desc->max will wi::eq to max_value_type, and niter would be "n", and then doloop.xx is 'n+1'. I would see how to merge these two adds safely at this point when generating doloop iv. (maybe range info, thanks! > > Is this
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On 2021-07-12 22:46, Richard Biener wrote: On Mon, 12 Jul 2021, guojiufu wrote: On 2021-07-12 18:02, Richard Biener wrote: > On Mon, 12 Jul 2021, guojiufu wrote: > >> On 2021-07-12 16:57, Richard Biener wrote: >> > On Mon, 12 Jul 2021, guojiufu wrote: >> > >> >> On 2021-07-12 14:20, Richard Biener wrote: >> >> > On Fri, 9 Jul 2021, Segher Boessenkool wrote: >> >> > >> >> >> On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: >> >> >> > I wonder if there's a way to query the target what modes the doloop >> >> >> > pattern can handle (not being too familiar with the doloop code). >> >> >> >> >> >> You can look what modes are allowed for operand 0 of doloop_end, >> >> >> perhaps? Although that is a define_expand, not a define_insn, so it >> >> >> is >> >> >> hard to introspect. >> >> >> >> >> >> > Why do you need to do any checks besides the new type being able to >> >> >> > represent all IV values? The original doloop IV will never wrap >> >> >> > (OTOH if niter is U*_MAX then we compute niter + 1 which will >> >> >> > become >> >> >> > zero ... I suppose the doloop might still do the correct thing here >> >> >> > but it also still will with a IV with larger type). >> >> >> >> The issue comes from U*_MAX (original short MAX), as you said: on which >> >> niter + 1 becomes zero. And because the step for doloop is -1; then, on >> >> larger type 'zero - 1' will be a very large number on larger type >> >> (e.g. 0xff...ff); but on the original short type 'zero - 1' is a small >> >> value >> >> (e.g. "0xff"). >> > >> > But for the larger type the small type MAX + 1 fits and does not yield >> > zero so it should still work exactly as before, no? Of course you >> > have to compute the + 1 in the larger type. >> > >> You are right, if compute the "+ 1" in the larger type it is ok, as below >> code: >> ``` >>/* Use type in word size may fast. */ >> if (TYPE_PRECISION (ntype) < BITS_PER_WORD) >> { >> ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); >> niter = fold_convert (ntype, niter); >> } >> >> tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), >> build_int_cst (ntype, 1)); >> >> >> add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, NULL, >> true); >> ``` >> The issue of this is, this code generates more stmt for doloop.xxx: >> _12 = (unsigned int) xx(D); >> _10 = _12 + 4294967295; >> _24 = (long unsigned int) _10; >> doloop.6_8 = _24 + 1; >> >> if use previous patch, "+ 1" on original type, then the stmts will looks >> like: >> _12 = (unsigned int) xx(D); >> doloop.6_8 = (long unsigned int) _12; >> >> This is the reason for checking >>wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype))) > > But this then only works when there's an upper bound on the number > of iterations. Note you should not use TYPE_MAX_VALUE here but > you can instead use > > wi::ltu_p (niter_desc->max, wi::to_widest (wi::max_value > (TYPE_PRECISION (ntype), TYPE_SIGN (ntype; Ok, Thanks! I remember you mentioned that: widest_int::from (wi::max_value (TYPE_PRECISION (ntype), TYPE_SIGN (ntype)), TYPE_SIGN (ntype)) would be better than wi::to_widest (TYPE_MAX_VALUE (ntype)). It seems that: "TYPE_MAX_VALUE (ntype)" is "NUMERICAL_TYPE_CHECK (NODE)->type_non_common.maxval" which do a numerical-check and return the field of maxval. And then call to wi::to_widest The other code "widest_int::from (wi::max_value (..,..),..)" calls wi::max_value and widest_int::from. I'm wondering if wi::to_widest (TYPE_MAX_VALUE (ntype)) is cheaper? TYPE_MAX_VALUE can be "suprising", it does not necessarily match the underlying modes precision. At some point we've tried to eliminate most of its uses, not sure what the situation/position is right now. Ok, get it, thanks. I will use "widest_int::from (wi::max_value (..,..),..)". > I think the -1 above comes from number of latch iterations vs. header > entries - it's a common source for this kind of issues. range analysis > might be able to prove that we can still merge the two adds even with > the intermediate extension. Yes, as you mentioned here, it relates to number of latch iterations For loop looks like : while (l < n) or for (i = 0; i < n; i++) This kind of loop, the niter is used to be 'n - 1' after transformed into 'do-while' form. For this kind of loop, the max value for the number of iteration "n - 1" would be "max_value_type(n) - 1" which is wi::ltu than max_value_type. This kind of loop is already common, and we could use wi::ltu (max, max_value_type) to check. For loop looks like: do ; while (n-- > 0); /* while (n-- > low); */ The niter_desc->max will wi::eq to max_value_type, and niter would be "n", and then doloop.xx is 'n+1'. I would see how to merge these two adds safely at this point when generating doloop iv. (maybe range info, thanks! > > Is this pre-loop extra add really offsetting the in-loop
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On Mon, 12 Jul 2021, guojiufu wrote: > On 2021-07-12 18:02, Richard Biener wrote: > > On Mon, 12 Jul 2021, guojiufu wrote: > > > >> On 2021-07-12 16:57, Richard Biener wrote: > >> > On Mon, 12 Jul 2021, guojiufu wrote: > >> > > >> >> On 2021-07-12 14:20, Richard Biener wrote: > >> >> > On Fri, 9 Jul 2021, Segher Boessenkool wrote: > >> >> > > >> >> >> On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: > >> >> >> > I wonder if there's a way to query the target what modes the doloop > >> >> >> > pattern can handle (not being too familiar with the doloop code). > >> >> >> > >> >> >> You can look what modes are allowed for operand 0 of doloop_end, > >> >> >> perhaps? Although that is a define_expand, not a define_insn, so it > >> >> >> is > >> >> >> hard to introspect. > >> >> >> > >> >> >> > Why do you need to do any checks besides the new type being able to > >> >> >> > represent all IV values? The original doloop IV will never wrap > >> >> >> > (OTOH if niter is U*_MAX then we compute niter + 1 which will > >> >> >> > become > >> >> >> > zero ... I suppose the doloop might still do the correct thing here > >> >> >> > but it also still will with a IV with larger type). > >> >> > >> >> The issue comes from U*_MAX (original short MAX), as you said: on which > >> >> niter + 1 becomes zero. And because the step for doloop is -1; then, on > >> >> larger type 'zero - 1' will be a very large number on larger type > >> >> (e.g. 0xff...ff); but on the original short type 'zero - 1' is a small > >> >> value > >> >> (e.g. "0xff"). > >> > > >> > But for the larger type the small type MAX + 1 fits and does not yield > >> > zero so it should still work exactly as before, no? Of course you > >> > have to compute the + 1 in the larger type. > >> > > >> You are right, if compute the "+ 1" in the larger type it is ok, as below > >> code: > >> ``` > >>/* Use type in word size may fast. */ > >> if (TYPE_PRECISION (ntype) < BITS_PER_WORD) > >> { > >> ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); > >> niter = fold_convert (ntype, niter); > >> } > >> > >> tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), > >> build_int_cst (ntype, 1)); > >> > >> > >> add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, NULL, > >> true); > >> ``` > >> The issue of this is, this code generates more stmt for doloop.xxx: > >> _12 = (unsigned int) xx(D); > >> _10 = _12 + 4294967295; > >> _24 = (long unsigned int) _10; > >> doloop.6_8 = _24 + 1; > >> > >> if use previous patch, "+ 1" on original type, then the stmts will looks > >> like: > >> _12 = (unsigned int) xx(D); > >> doloop.6_8 = (long unsigned int) _12; > >> > >> This is the reason for checking > >>wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype))) > > > > But this then only works when there's an upper bound on the number > > of iterations. Note you should not use TYPE_MAX_VALUE here but > > you can instead use > > > > wi::ltu_p (niter_desc->max, wi::to_widest (wi::max_value > > (TYPE_PRECISION (ntype), TYPE_SIGN (ntype; > > Ok, Thanks! > I remember you mentioned that: > widest_int::from (wi::max_value (TYPE_PRECISION (ntype), TYPE_SIGN (ntype)), > TYPE_SIGN (ntype)) > would be better than > wi::to_widest (TYPE_MAX_VALUE (ntype)). > > It seems that: > "TYPE_MAX_VALUE (ntype)" is "NUMERICAL_TYPE_CHECK > (NODE)->type_non_common.maxval" > which do a numerical-check and return the field of maxval. And then call to > wi::to_widest > > The other code "widest_int::from (wi::max_value (..,..),..)" calls > wi::max_value > and widest_int::from. > > I'm wondering if wi::to_widest (TYPE_MAX_VALUE (ntype)) is cheaper? TYPE_MAX_VALUE can be "suprising", it does not necessarily match the underlying modes precision. At some point we've tried to eliminate most of its uses, not sure what the situation/position is right now. > > I think the -1 above comes from number of latch iterations vs. header > > entries - it's a common source for this kind of issues. range analysis > > might be able to prove that we can still merge the two adds even with > > the intermediate extension. > Yes, as you mentioned here, it relates to number of latch iterations > For loop looks like : while (l < n) or for (i = 0; i < n; i++) > This kind of loop, the niter is used to be 'n - 1' after transformed > into 'do-while' form. > I would see how to merge these two adds safely at this point > when generating doloop iv. (maybe range info, thanks! > > > > > Is this pre-loop extra add really offsetting the in-loop doloop > > improvements? > I'm not catching this question too much, sorry. I guess your concern > is if the "+1" is an offset: it may not, "+1" may be just that doloop.xx > is decreasing niter until 0 (all number >0). > If misunderstand, thanks for point out. I'm questioning the argument that not being able to eliminate the +1-1 pair
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On 2021-07-12 18:02, Richard Biener wrote: On Mon, 12 Jul 2021, guojiufu wrote: On 2021-07-12 16:57, Richard Biener wrote: > On Mon, 12 Jul 2021, guojiufu wrote: > >> On 2021-07-12 14:20, Richard Biener wrote: >> > On Fri, 9 Jul 2021, Segher Boessenkool wrote: >> > >> >> On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: >> >> > I wonder if there's a way to query the target what modes the doloop >> >> > pattern can handle (not being too familiar with the doloop code). >> >> >> >> You can look what modes are allowed for operand 0 of doloop_end, >> >> perhaps? Although that is a define_expand, not a define_insn, so it is >> >> hard to introspect. >> >> >> >> > Why do you need to do any checks besides the new type being able to >> >> > represent all IV values? The original doloop IV will never wrap >> >> > (OTOH if niter is U*_MAX then we compute niter + 1 which will become >> >> > zero ... I suppose the doloop might still do the correct thing here >> >> > but it also still will with a IV with larger type). >> >> The issue comes from U*_MAX (original short MAX), as you said: on which >> niter + 1 becomes zero. And because the step for doloop is -1; then, on >> larger type 'zero - 1' will be a very large number on larger type >> (e.g. 0xff...ff); but on the original short type 'zero - 1' is a small >> value >> (e.g. "0xff"). > > But for the larger type the small type MAX + 1 fits and does not yield > zero so it should still work exactly as before, no? Of course you > have to compute the + 1 in the larger type. > You are right, if compute the "+ 1" in the larger type it is ok, as below code: ``` /* Use type in word size may fast. */ if (TYPE_PRECISION (ntype) < BITS_PER_WORD) { ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); niter = fold_convert (ntype, niter); } tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), build_int_cst (ntype, 1)); add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, NULL, true); ``` The issue of this is, this code generates more stmt for doloop.xxx: _12 = (unsigned int) xx(D); _10 = _12 + 4294967295; _24 = (long unsigned int) _10; doloop.6_8 = _24 + 1; if use previous patch, "+ 1" on original type, then the stmts will looks like: _12 = (unsigned int) xx(D); doloop.6_8 = (long unsigned int) _12; This is the reason for checking wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype))) But this then only works when there's an upper bound on the number of iterations. Note you should not use TYPE_MAX_VALUE here but you can instead use wi::ltu_p (niter_desc->max, wi::to_widest (wi::max_value (TYPE_PRECISION (ntype), TYPE_SIGN (ntype; Ok, Thanks! I remember you mentioned that: widest_int::from (wi::max_value (TYPE_PRECISION (ntype), TYPE_SIGN (ntype)), TYPE_SIGN (ntype)) would be better than wi::to_widest (TYPE_MAX_VALUE (ntype)). It seems that: "TYPE_MAX_VALUE (ntype)" is "NUMERICAL_TYPE_CHECK (NODE)->type_non_common.maxval" which do a numerical-check and return the field of maxval. And then call to wi::to_widest The other code "widest_int::from (wi::max_value (..,..),..)" calls wi::max_value and widest_int::from. I'm wondering if wi::to_widest (TYPE_MAX_VALUE (ntype)) is cheaper? I think the -1 above comes from number of latch iterations vs. header entries - it's a common source for this kind of issues. range analysis might be able to prove that we can still merge the two adds even with the intermediate extension. Yes, as you mentioned here, it relates to number of latch iterations For loop looks like : while (l < n) or for (i = 0; i < n; i++) This kind of loop, the niter is used to be 'n - 1' after transformed into 'do-while' form. I would see how to merge these two adds safely at this point when generating doloop iv. (maybe range info, thanks! Is this pre-loop extra add really offsetting the in-loop doloop improvements? I'm not catching this question too much, sorry. I guess your concern is if the "+1" is an offset: it may not, "+1" may be just that doloop.xx is decreasing niter until 0 (all number >0). If misunderstand, thanks for point out. >> >> >> >> doloop_valid_p guarantees it is simple and doesn't wrap. >> >> >> >> > I'd have expected sth like >> >> > >> >> >ntype = lang_hooks.types.type_for_mode (word_mode, TYPE_UNSIGNED >> >> > (ntype)); >> >> > >> >> > thus the decision made using a mode - which is also why I wonder >> >> > if there's a way to query the target for this. As you say, >> >> > it _may_ be fast, so better check (somehow). >> >> >> I was also thinking of using hooks like type_for_size/type_for_mode. >> /* Use type in word size may fast. */ >> if (TYPE_PRECISION (ntype) < BITS_PER_WORD >> && Wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE >> (ntype >> { >> ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1);
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On Mon, 12 Jul 2021, guojiufu wrote: > On 2021-07-12 16:57, Richard Biener wrote: > > On Mon, 12 Jul 2021, guojiufu wrote: > > > >> On 2021-07-12 14:20, Richard Biener wrote: > >> > On Fri, 9 Jul 2021, Segher Boessenkool wrote: > >> > > >> >> On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: > >> >> > I wonder if there's a way to query the target what modes the doloop > >> >> > pattern can handle (not being too familiar with the doloop code). > >> >> > >> >> You can look what modes are allowed for operand 0 of doloop_end, > >> >> perhaps? Although that is a define_expand, not a define_insn, so it is > >> >> hard to introspect. > >> >> > >> >> > Why do you need to do any checks besides the new type being able to > >> >> > represent all IV values? The original doloop IV will never wrap > >> >> > (OTOH if niter is U*_MAX then we compute niter + 1 which will become > >> >> > zero ... I suppose the doloop might still do the correct thing here > >> >> > but it also still will with a IV with larger type). > >> > >> The issue comes from U*_MAX (original short MAX), as you said: on which > >> niter + 1 becomes zero. And because the step for doloop is -1; then, on > >> larger type 'zero - 1' will be a very large number on larger type > >> (e.g. 0xff...ff); but on the original short type 'zero - 1' is a small > >> value > >> (e.g. "0xff"). > > > > But for the larger type the small type MAX + 1 fits and does not yield > > zero so it should still work exactly as before, no? Of course you > > have to compute the + 1 in the larger type. > > > You are right, if compute the "+ 1" in the larger type it is ok, as below > code: > ``` >/* Use type in word size may fast. */ > if (TYPE_PRECISION (ntype) < BITS_PER_WORD) > { > ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); > niter = fold_convert (ntype, niter); > } > > tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), > build_int_cst (ntype, 1)); > > > add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, NULL, > true); > ``` > The issue of this is, this code generates more stmt for doloop.xxx: > _12 = (unsigned int) xx(D); > _10 = _12 + 4294967295; > _24 = (long unsigned int) _10; > doloop.6_8 = _24 + 1; > > if use previous patch, "+ 1" on original type, then the stmts will looks like: > _12 = (unsigned int) xx(D); > doloop.6_8 = (long unsigned int) _12; > > This is the reason for checking >wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype))) But this then only works when there's an upper bound on the number of iterations. Note you should not use TYPE_MAX_VALUE here but you can instead use wi::ltu_p (niter_desc->max, wi::to_widest (wi::max_value (TYPE_PRECISION (ntype), TYPE_SIGN (ntype; I think the -1 above comes from number of latch iterations vs. header entries - it's a common source for this kind of issues. range analysis might be able to prove that we can still merge the two adds even with the intermediate extension. Is this pre-loop extra add really offsetting the in-loop doloop improvements? > >> >> > >> >> doloop_valid_p guarantees it is simple and doesn't wrap. > >> >> > >> >> > I'd have expected sth like > >> >> > > >> >> >ntype = lang_hooks.types.type_for_mode (word_mode, TYPE_UNSIGNED > >> >> > (ntype)); > >> >> > > >> >> > thus the decision made using a mode - which is also why I wonder > >> >> > if there's a way to query the target for this. As you say, > >> >> > it _may_ be fast, so better check (somehow). > >> > >> > >> I was also thinking of using hooks like type_for_size/type_for_mode. > >> /* Use type in word size may fast. */ > >> if (TYPE_PRECISION (ntype) < BITS_PER_WORD > >> && Wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE > >> (ntype > >> { > >> ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); > >> base = fold_convert (ntype, base); > >> } > >> > >> As you pointed out, this does not query the mode from targets. > >> As Segher pointed out "doloop_end" checks unsupported mode, while it seems > >> not easy to use it in tree-ssa-loop-ivopts.c. > >> For implementations of doloop_end, tartgets like rs6000/aarch64/ia64 > >> requires > >> Pmode/DImode; while there are other targets that work on other 'mode' (e.g. > >> SI). > >> > >> > >> In doloop_optimize, there is code: > >> > >> ``` > >> mode = desc->mode; > >> . > >> doloop_reg = gen_reg_rtx (mode); > >> rtx_insn *doloop_seq = targetm.gen_doloop_end (doloop_reg, > >> start_label); > >> > >> word_mode_size = GET_MODE_PRECISION (word_mode); > >> word_mode_max = (HOST_WIDE_INT_1U << (word_mode_size - 1) << 1) - > >> 1; > >> if (! doloop_seq > >> && mode != word_mode > >> /* Before trying mode different from the one in that # of > >> iterations is > >>computed, we must be sure that the
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On 2021-07-12 16:57, Richard Biener wrote: On Mon, 12 Jul 2021, guojiufu wrote: On 2021-07-12 14:20, Richard Biener wrote: > On Fri, 9 Jul 2021, Segher Boessenkool wrote: > >> On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: >> > I wonder if there's a way to query the target what modes the doloop >> > pattern can handle (not being too familiar with the doloop code). >> >> You can look what modes are allowed for operand 0 of doloop_end, >> perhaps? Although that is a define_expand, not a define_insn, so it is >> hard to introspect. >> >> > Why do you need to do any checks besides the new type being able to >> > represent all IV values? The original doloop IV will never wrap >> > (OTOH if niter is U*_MAX then we compute niter + 1 which will become >> > zero ... I suppose the doloop might still do the correct thing here >> > but it also still will with a IV with larger type). The issue comes from U*_MAX (original short MAX), as you said: on which niter + 1 becomes zero. And because the step for doloop is -1; then, on larger type 'zero - 1' will be a very large number on larger type (e.g. 0xff...ff); but on the original short type 'zero - 1' is a small value (e.g. "0xff"). But for the larger type the small type MAX + 1 fits and does not yield zero so it should still work exactly as before, no? Of course you have to compute the + 1 in the larger type. You are right, if compute the "+ 1" in the larger type it is ok, as below code: ``` /* Use type in word size may fast. */ if (TYPE_PRECISION (ntype) < BITS_PER_WORD) { ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); niter = fold_convert (ntype, niter); } tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), build_int_cst (ntype, 1)); add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, NULL, true); ``` The issue of this is, this code generates more stmt for doloop.xxx: _12 = (unsigned int) xx(D); _10 = _12 + 4294967295; _24 = (long unsigned int) _10; doloop.6_8 = _24 + 1; if use previous patch, "+ 1" on original type, then the stmts will looks like: _12 = (unsigned int) xx(D); doloop.6_8 = (long unsigned int) _12; This is the reason for checking wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype))) >> >> doloop_valid_p guarantees it is simple and doesn't wrap. >> >> > I'd have expected sth like >> > >> >ntype = lang_hooks.types.type_for_mode (word_mode, TYPE_UNSIGNED >> > (ntype)); >> > >> > thus the decision made using a mode - which is also why I wonder >> > if there's a way to query the target for this. As you say, >> > it _may_ be fast, so better check (somehow). I was also thinking of using hooks like type_for_size/type_for_mode. /* Use type in word size may fast. */ if (TYPE_PRECISION (ntype) < BITS_PER_WORD && Wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype { ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); base = fold_convert (ntype, base); } As you pointed out, this does not query the mode from targets. As Segher pointed out "doloop_end" checks unsupported mode, while it seems not easy to use it in tree-ssa-loop-ivopts.c. For implementations of doloop_end, tartgets like rs6000/aarch64/ia64 requires Pmode/DImode; while there are other targets that work on other 'mode' (e.g. SI). In doloop_optimize, there is code: ``` mode = desc->mode; . doloop_reg = gen_reg_rtx (mode); rtx_insn *doloop_seq = targetm.gen_doloop_end (doloop_reg, start_label); word_mode_size = GET_MODE_PRECISION (word_mode); word_mode_max = (HOST_WIDE_INT_1U << (word_mode_size - 1) << 1) - 1; if (! doloop_seq && mode != word_mode /* Before trying mode different from the one in that # of iterations is computed, we must be sure that the number of iterations fits into the new mode. */ && (word_mode_size >= GET_MODE_PRECISION (mode) || wi::leu_p (iterations_max, word_mode_max))) { if (word_mode_size > GET_MODE_PRECISION (mode)) count = simplify_gen_unary (ZERO_EXTEND, word_mode, count, mode); else count = lowpart_subreg (word_mode, count, mode); PUT_MODE (doloop_reg, word_mode); doloop_seq = targetm.gen_doloop_end (doloop_reg, start_label); } if (! doloop_seq) { if (dump_file) fprintf (dump_file, "Doloop: Target unwilling to use doloop pattern!\n"); return false; } ``` The above code first tries the mode of niter_desc by call targetm.gen_doloop_end to see if the target can generate doloop insns, if fail, then try to use 'word_mode' against gen_doloop_end. >> >> Almost all targets just use Pmode, but there is no such guarantee I >> think, and esp. some targets that do not have machine insns for this >>
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On Mon, 12 Jul 2021, guojiufu wrote: > On 2021-07-12 14:20, Richard Biener wrote: > > On Fri, 9 Jul 2021, Segher Boessenkool wrote: > > > >> On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: > >> > I wonder if there's a way to query the target what modes the doloop > >> > pattern can handle (not being too familiar with the doloop code). > >> > >> You can look what modes are allowed for operand 0 of doloop_end, > >> perhaps? Although that is a define_expand, not a define_insn, so it is > >> hard to introspect. > >> > >> > Why do you need to do any checks besides the new type being able to > >> > represent all IV values? The original doloop IV will never wrap > >> > (OTOH if niter is U*_MAX then we compute niter + 1 which will become > >> > zero ... I suppose the doloop might still do the correct thing here > >> > but it also still will with a IV with larger type). > > The issue comes from U*_MAX (original short MAX), as you said: on which > niter + 1 becomes zero. And because the step for doloop is -1; then, on > larger type 'zero - 1' will be a very large number on larger type > (e.g. 0xff...ff); but on the original short type 'zero - 1' is a small value > (e.g. "0xff"). But for the larger type the small type MAX + 1 fits and does not yield zero so it should still work exactly as before, no? Of course you have to compute the + 1 in the larger type. > >> > >> doloop_valid_p guarantees it is simple and doesn't wrap. > >> > >> > I'd have expected sth like > >> > > >> >ntype = lang_hooks.types.type_for_mode (word_mode, TYPE_UNSIGNED > >> > (ntype)); > >> > > >> > thus the decision made using a mode - which is also why I wonder > >> > if there's a way to query the target for this. As you say, > >> > it _may_ be fast, so better check (somehow). > > > I was also thinking of using hooks like type_for_size/type_for_mode. > /* Use type in word size may fast. */ > if (TYPE_PRECISION (ntype) < BITS_PER_WORD > && Wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE > (ntype > { > ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); > base = fold_convert (ntype, base); > } > > As you pointed out, this does not query the mode from targets. > As Segher pointed out "doloop_end" checks unsupported mode, while it seems > not easy to use it in tree-ssa-loop-ivopts.c. > For implementations of doloop_end, tartgets like rs6000/aarch64/ia64 requires > Pmode/DImode; while there are other targets that work on other 'mode' (e.g. > SI). > > > In doloop_optimize, there is code: > > ``` > mode = desc->mode; > . > doloop_reg = gen_reg_rtx (mode); > rtx_insn *doloop_seq = targetm.gen_doloop_end (doloop_reg, > start_label); > > word_mode_size = GET_MODE_PRECISION (word_mode); > word_mode_max = (HOST_WIDE_INT_1U << (word_mode_size - 1) << 1) - 1; > if (! doloop_seq > && mode != word_mode > /* Before trying mode different from the one in that # of > iterations is >computed, we must be sure that the number of iterations fits into >the new mode. */ > && (word_mode_size >= GET_MODE_PRECISION (mode) > || wi::leu_p (iterations_max, word_mode_max))) > { > if (word_mode_size > GET_MODE_PRECISION (mode)) > count = simplify_gen_unary (ZERO_EXTEND, word_mode, count, mode); > else > count = lowpart_subreg (word_mode, count, mode); > PUT_MODE (doloop_reg, word_mode); > doloop_seq = targetm.gen_doloop_end (doloop_reg, start_label); > } > if (! doloop_seq) > { > if (dump_file) > fprintf (dump_file, >"Doloop: Target unwilling to use doloop pattern!\n"); > return false; > } > ``` > The above code first tries the mode of niter_desc by call > targetm.gen_doloop_end > to see if the target can generate doloop insns, if fail, then try to use > 'word_mode' against gen_doloop_end. > > > >> > >> Almost all targets just use Pmode, but there is no such guarantee I > >> think, and esp. some targets that do not have machine insns for this > >> (but want to generate different code for this anyway) can do pretty much > >> anything. > >> > >> Maybe using just Pmode here is good enough though? > > > > I think Pmode is a particularly bad choice and I'd prefer word_mode > > if we go for any hardcoded mode. s390x for example seems to handle > > both SImode and DImode (but names the helper gen_doloop_si64 > > for SImode?!). But indeed it looks like somehow querying doloop_end > > is going to be difficult since the expander doesn't have any mode, > > so we'd have to actually try emit RTL here. > > Instead of using hardcode mode, maybe we could add a hook for targets to > return > the preferred mode. That's a possiblity of course. Like the following which just shows the default implementation then (pass in current mode, return a more preferred mode or the
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On 2021-07-12 14:20, Richard Biener wrote: On Fri, 9 Jul 2021, Segher Boessenkool wrote: On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: > I wonder if there's a way to query the target what modes the doloop > pattern can handle (not being too familiar with the doloop code). You can look what modes are allowed for operand 0 of doloop_end, perhaps? Although that is a define_expand, not a define_insn, so it is hard to introspect. > Why do you need to do any checks besides the new type being able to > represent all IV values? The original doloop IV will never wrap > (OTOH if niter is U*_MAX then we compute niter + 1 which will become > zero ... I suppose the doloop might still do the correct thing here > but it also still will with a IV with larger type). The issue comes from U*_MAX (original short MAX), as you said: on which niter + 1 becomes zero. And because the step for doloop is -1; then, on larger type 'zero - 1' will be a very large number on larger type (e.g. 0xff...ff); but on the original short type 'zero - 1' is a small value (e.g. "0xff"). doloop_valid_p guarantees it is simple and doesn't wrap. > I'd have expected sth like > >ntype = lang_hooks.types.type_for_mode (word_mode, TYPE_UNSIGNED > (ntype)); > > thus the decision made using a mode - which is also why I wonder > if there's a way to query the target for this. As you say, > it _may_ be fast, so better check (somehow). I was also thinking of using hooks like type_for_size/type_for_mode. /* Use type in word size may fast. */ if (TYPE_PRECISION (ntype) < BITS_PER_WORD && Wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype { ntype = lang_hooks.types.type_for_size (BITS_PER_WORD, 1); base = fold_convert (ntype, base); } As you pointed out, this does not query the mode from targets. As Segher pointed out "doloop_end" checks unsupported mode, while it seems not easy to use it in tree-ssa-loop-ivopts.c. For implementations of doloop_end, tartgets like rs6000/aarch64/ia64 requires Pmode/DImode; while there are other targets that work on other 'mode' (e.g. SI). In doloop_optimize, there is code: ``` mode = desc->mode; . doloop_reg = gen_reg_rtx (mode); rtx_insn *doloop_seq = targetm.gen_doloop_end (doloop_reg, start_label); word_mode_size = GET_MODE_PRECISION (word_mode); word_mode_max = (HOST_WIDE_INT_1U << (word_mode_size - 1) << 1) - 1; if (! doloop_seq && mode != word_mode /* Before trying mode different from the one in that # of iterations is computed, we must be sure that the number of iterations fits into the new mode. */ && (word_mode_size >= GET_MODE_PRECISION (mode) || wi::leu_p (iterations_max, word_mode_max))) { if (word_mode_size > GET_MODE_PRECISION (mode)) count = simplify_gen_unary (ZERO_EXTEND, word_mode, count, mode); else count = lowpart_subreg (word_mode, count, mode); PUT_MODE (doloop_reg, word_mode); doloop_seq = targetm.gen_doloop_end (doloop_reg, start_label); } if (! doloop_seq) { if (dump_file) fprintf (dump_file, "Doloop: Target unwilling to use doloop pattern!\n"); return false; } ``` The above code first tries the mode of niter_desc by call targetm.gen_doloop_end to see if the target can generate doloop insns, if fail, then try to use 'word_mode' against gen_doloop_end. Almost all targets just use Pmode, but there is no such guarantee I think, and esp. some targets that do not have machine insns for this (but want to generate different code for this anyway) can do pretty much anything. Maybe using just Pmode here is good enough though? I think Pmode is a particularly bad choice and I'd prefer word_mode if we go for any hardcoded mode. s390x for example seems to handle both SImode and DImode (but names the helper gen_doloop_si64 for SImode?!). But indeed it looks like somehow querying doloop_end is going to be difficult since the expander doesn't have any mode, so we'd have to actually try emit RTL here. Instead of using hardcode mode, maybe we could add a hook for targets to return the preferred mode. Thanks for those valuable comments! Jiufu Guo Richard.
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On Fri, 9 Jul 2021, Segher Boessenkool wrote: > On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: > > I wonder if there's a way to query the target what modes the doloop > > pattern can handle (not being too familiar with the doloop code). > > You can look what modes are allowed for operand 0 of doloop_end, > perhaps? Although that is a define_expand, not a define_insn, so it is > hard to introspect. > > > Why do you need to do any checks besides the new type being able to > > represent all IV values? The original doloop IV will never wrap > > (OTOH if niter is U*_MAX then we compute niter + 1 which will become > > zero ... I suppose the doloop might still do the correct thing here > > but it also still will with a IV with larger type). > > doloop_valid_p guarantees it is simple and doesn't wrap. > > > I'd have expected sth like > > > >ntype = lang_hooks.types.type_for_mode (word_mode, TYPE_UNSIGNED > > (ntype)); > > > > thus the decision made using a mode - which is also why I wonder > > if there's a way to query the target for this. As you say, > > it _may_ be fast, so better check (somehow). > > Almost all targets just use Pmode, but there is no such guarantee I > think, and esp. some targets that do not have machine insns for this > (but want to generate different code for this anyway) can do pretty much > anything. > > Maybe using just Pmode here is good enough though? I think Pmode is a particularly bad choice and I'd prefer word_mode if we go for any hardcoded mode. s390x for example seems to handle both SImode and DImode (but names the helper gen_doloop_si64 for SImode?!). But indeed it looks like somehow querying doloop_end is going to be difficult since the expander doesn't have any mode, so we'd have to actually try emit RTL here. Richard.
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On Fri, Jul 09, 2021 at 08:43:59AM +0200, Richard Biener wrote: > I wonder if there's a way to query the target what modes the doloop > pattern can handle (not being too familiar with the doloop code). You can look what modes are allowed for operand 0 of doloop_end, perhaps? Although that is a define_expand, not a define_insn, so it is hard to introspect. > Why do you need to do any checks besides the new type being able to > represent all IV values? The original doloop IV will never wrap > (OTOH if niter is U*_MAX then we compute niter + 1 which will become > zero ... I suppose the doloop might still do the correct thing here > but it also still will with a IV with larger type). doloop_valid_p guarantees it is simple and doesn't wrap. > I'd have expected sth like > >ntype = lang_hooks.types.type_for_mode (word_mode, TYPE_UNSIGNED > (ntype)); > > thus the decision made using a mode - which is also why I wonder > if there's a way to query the target for this. As you say, > it _may_ be fast, so better check (somehow). Almost all targets just use Pmode, but there is no such guarantee I think, and esp. some targets that do not have machine insns for this (but want to generate different code for this anyway) can do pretty much anything. Maybe using just Pmode here is good enough though? Segher
Re: [PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
On Fri, 9 Jul 2021, Jiufu Guo wrote: > Currently, doloop.xx variable is using the type as niter which may shorter > than word size. For some cases, it may be better to use word size type. > For example, on some 64bit system, to access 32bit niter, subreg maybe used. > Then using 64bit type would not need to use subreg if the value can be > present in both 32bit and 64bit. > > This patch updates doloop iv to BIT_PER_WORD size if it is fine. > > Bootstrap and regtest pass on powerpc64le and x86, is this ok for trunk? > > BR. > Jiufu > > gcc/ChangeLog: > > 2021-07-08 Jiufu Guo > > PR target/61837 > * tree-ssa-loop-ivopts.c (add_iv_candidate_for_doloop): > Update iv on BITS_PER_WORD for niter. > > gcc/testsuite/ChangeLog: > > 2021-07-08 Jiufu Guo > > PR target/61837 > * gcc.target/powerpc/pr61837.c: New test. > > --- > gcc/testsuite/gcc.target/powerpc/pr61837.c | 16 > gcc/tree-ssa-loop-ivopts.c | 10 ++ > 2 files changed, 26 insertions(+) > create mode 100644 gcc/testsuite/gcc.target/powerpc/pr61837.c > > diff --git a/gcc/testsuite/gcc.target/powerpc/pr61837.c > b/gcc/testsuite/gcc.target/powerpc/pr61837.c > new file mode 100644 > index 000..dc44eb9cb41 > --- /dev/null > +++ b/gcc/testsuite/gcc.target/powerpc/pr61837.c > @@ -0,0 +1,16 @@ > +/* { dg-do compile } */ > +/* { dg-options "-O2" } */ > +void foo(int *p1, long *p2, int s) > +{ > + int n, v, i; > + > + v = 0; > + for (n = 0; n <= 100; n++) { > + for (i = 0; i < s; i++) > +if (p2[i] == n) > + p1[i] = v; > + v += 88; > + } > +} > + > +/* { dg-final { scan-assembler-not {\mrldicl\M} } } */ > diff --git a/gcc/tree-ssa-loop-ivopts.c b/gcc/tree-ssa-loop-ivopts.c > index 12a8a49a307..c3c2f97918d 100644 > --- a/gcc/tree-ssa-loop-ivopts.c > +++ b/gcc/tree-ssa-loop-ivopts.c > @@ -5690,6 +5690,16 @@ add_iv_candidate_for_doloop (struct ivopts_data *data) > >tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), > build_int_cst (ntype, 1)); > + > + /* Use type in word size may fast. */ > + if (TYPE_PRECISION (ntype) < BITS_PER_WORD > + && TYPE_PRECISION (long_unsigned_type_node) == BITS_PER_WORD > + && wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype I wonder if there's a way to query the target what modes the doloop pattern can handle (not being too familiar with the doloop code). Why do you need to do any checks besides the new type being able to represent all IV values? The original doloop IV will never wrap (OTOH if niter is U*_MAX then we compute niter + 1 which will become zero ... I suppose the doloop might still do the correct thing here but it also still will with a IV with larger type). I'd have expected sth like ntype = lang_hooks.types.type_for_mode (word_mode, TYPE_UNSIGNED (ntype)); thus the decision made using a mode - which is also why I wonder if there's a way to query the target for this. As you say, it _may_ be fast, so better check (somehow). > +{ > + ntype = long_unsigned_type_node; > + base = fold_convert (ntype, base); > +} > + >add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, NULL, > true); > } > > -- Richard Biener SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany; GF: Felix Imendörffer; HRB 36809 (AG Nuernberg)
[PATCH] Check type size for doloop iv on BITS_PER_WORD [PR61837]
Currently, doloop.xx variable is using the type as niter which may shorter than word size. For some cases, it may be better to use word size type. For example, on some 64bit system, to access 32bit niter, subreg maybe used. Then using 64bit type would not need to use subreg if the value can be present in both 32bit and 64bit. This patch updates doloop iv to BIT_PER_WORD size if it is fine. Bootstrap and regtest pass on powerpc64le and x86, is this ok for trunk? BR. Jiufu gcc/ChangeLog: 2021-07-08 Jiufu Guo PR target/61837 * tree-ssa-loop-ivopts.c (add_iv_candidate_for_doloop): Update iv on BITS_PER_WORD for niter. gcc/testsuite/ChangeLog: 2021-07-08 Jiufu Guo PR target/61837 * gcc.target/powerpc/pr61837.c: New test. --- gcc/testsuite/gcc.target/powerpc/pr61837.c | 16 gcc/tree-ssa-loop-ivopts.c | 10 ++ 2 files changed, 26 insertions(+) create mode 100644 gcc/testsuite/gcc.target/powerpc/pr61837.c diff --git a/gcc/testsuite/gcc.target/powerpc/pr61837.c b/gcc/testsuite/gcc.target/powerpc/pr61837.c new file mode 100644 index 000..dc44eb9cb41 --- /dev/null +++ b/gcc/testsuite/gcc.target/powerpc/pr61837.c @@ -0,0 +1,16 @@ +/* { dg-do compile } */ +/* { dg-options "-O2" } */ +void foo(int *p1, long *p2, int s) +{ + int n, v, i; + + v = 0; + for (n = 0; n <= 100; n++) { + for (i = 0; i < s; i++) +if (p2[i] == n) + p1[i] = v; + v += 88; + } +} + +/* { dg-final { scan-assembler-not {\mrldicl\M} } } */ diff --git a/gcc/tree-ssa-loop-ivopts.c b/gcc/tree-ssa-loop-ivopts.c index 12a8a49a307..c3c2f97918d 100644 --- a/gcc/tree-ssa-loop-ivopts.c +++ b/gcc/tree-ssa-loop-ivopts.c @@ -5690,6 +5690,16 @@ add_iv_candidate_for_doloop (struct ivopts_data *data) tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), build_int_cst (ntype, 1)); + + /* Use type in word size may fast. */ + if (TYPE_PRECISION (ntype) < BITS_PER_WORD + && TYPE_PRECISION (long_unsigned_type_node) == BITS_PER_WORD + && wi::ltu_p (niter_desc->max, wi::to_widest (TYPE_MAX_VALUE (ntype +{ + ntype = long_unsigned_type_node; + base = fold_convert (ntype, base); +} + add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, NULL, true); } -- 2.17.1