Re: [PATCH] PR tree-optimization/108359 - Utilize op1 == op2 when invoking range-ops folding.
On Mon, Jan 16, 2023 at 08:19:25AM +0100, Richard Biener wrote: > + // If op1 and op2 are equivalences, then we don't need a complete cross > + // product, just pairs of matching elements. > + if (relation_equiv_p (rel) && lh == rh && num_lh <= 16) > +{ > + int_range_max tmp; > + r.set_undefined (); > + for (unsigned x = 0; x < num_lh; ++x) > + { > + wide_int lh_lb = lh.lower_bound (x); > + wide_int lh_ub = lh.upper_bound (x); > + wi_fold_in_parts_equiv (tmp, type, lh_lb, lh_ub); > > and that does > > + widest_int lh_range = wi::sub (widest_int::from (lh_ub, TYPE_SIGN (type)), > +widest_int::from (lh_lb, TYPE_SIGN (type))); > + // if there are 1 to 8 values in the LH range, split them up. > + r.set_undefined (); > + if (lh_range >= 0 && lh_range <= 7) > +{ > + for (unsigned x = 0; x <= lh_range; x++) > > which in total limits the number of sub-ranges in the output but in an > odd way. It's also all-or-nothing. IIRC there's a hard limit on the > number of sub-ranges in the output anyway via int_range, so > why not use that and always do the first loop over the sub-ranges > of the inputs and the second loop over the range members but > stop when we reach N-1 and then use wi_fold on the remainds? > > Your above code suggests we go up to 112 sub-ranges and once > we'd reach 113 we'd fold down to a single. > > Maybe my "heuristic" wouldn't be much better, but then somehow > breaking the heuristic down to a single magic number would be > better, esp. since .union_ will undo some of the breakup when > reaching N? The <= X in the first case is what I've suggested, the previous behavior didn't suffer from this all or nothing case, but I was worried because the behavior is that we create many subranges and then in the end just merge the last ones into a single subrange such that it fits into the limit. So we could create 255 * 8 subranges and then merge the last 255 * 7 + 1 into one. We could call that wi_fold_in_parts_equiv only if r.num_parts () < m_max_ranges and wi_fold otherwise... Jakub
Re: [PATCH] PR tree-optimization/108359 - Utilize op1 == op2 when invoking range-ops folding.
On 1/16/23 08:19, Richard Biener wrote: On Fri, Jan 13, 2023 at 11:07 PM Andrew MacLeod wrote: On 1/13/23 16:54, Jakub Jelinek wrote: On Fri, Jan 13, 2023 at 04:23:20PM -0500, Andrew MacLeod via Gcc-patches wrote: fold_range() already invokes wi_fold_in_parts to try to get more refined information. If the subranges are quite small, it will do each individual calculation and combine the results. x * y with x = [1,3] and y = [1,3] is broken down and we calculate each possibility and we end up with [1,4][6,6][9,9] instead of [1,9] We limit this as the time is between quadratic to exponential depending on the number of elements in x and y. If we also check the relation and determine that x == y, we don't need to worry about that growth as this process is linear. The above case will be broken down to just 1*1, 2*2 and 3*3, resulting in a range of [1, 1][4,4][9,9]. In the testcase, it happens to be the right_shift operation, but this solution is generic and applies to all range-op operations. I added a testcase which checks >>, *, + and %. I also arbitrarily chose 8 elements as the limit for breaking down individual operations. The overall compile time change for this is negligible. Although this is a regression fix, it will affect all operations where x == y, which is where my initial hesitancy arose. Regardless, bootstrapped on x86_64-pc-linux-gnu with no regressions. OK for trunk? Will defer to Aldy, just some nits. Did you mean Richi? I suppose Aldy, since it's a regression fix it's OK even during stage4. I don't have any issues with the patch. Whatever the release managers agree to, I'm fine with. Aldy I do have a comment as well though, you do + // If op1 and op2 are equivalences, then we don't need a complete cross + // product, just pairs of matching elements. + if (relation_equiv_p (rel) && lh == rh && num_lh <= 16) +{ + int_range_max tmp; + r.set_undefined (); + for (unsigned x = 0; x < num_lh; ++x) + { + wide_int lh_lb = lh.lower_bound (x); + wide_int lh_ub = lh.upper_bound (x); + wi_fold_in_parts_equiv (tmp, type, lh_lb, lh_ub); and that does + widest_int lh_range = wi::sub (widest_int::from (lh_ub, TYPE_SIGN (type)), +widest_int::from (lh_lb, TYPE_SIGN (type))); + // if there are 1 to 8 values in the LH range, split them up. + r.set_undefined (); + if (lh_range >= 0 && lh_range <= 7) +{ + for (unsigned x = 0; x <= lh_range; x++) which in total limits the number of sub-ranges in the output but in an odd way. It's also all-or-nothing. IIRC there's a hard limit on the number of sub-ranges in the output anyway via int_range, so why not use that and always do the first loop over the sub-ranges of the inputs and the second loop over the range members but stop when we reach N-1 and then use wi_fold on the remainds? Your above code suggests we go up to 112 sub-ranges and once we'd reach 113 we'd fold down to a single. Maybe my "heuristic" wouldn't be much better, but then somehow breaking the heuristic down to a single magic number would be better, esp. since .union_ will undo some of the breakup when reaching N? + // if there are 1 to 8 values in the LH range, split them up. + r.set_undefined (); + if (lh_range >= 0 && lh_range <= 7) +{ + unsigned x; + for (x = 0; x <= lh_range; x++) Nothing uses x after the loop, so why not for (unsigned x = 0; x <= lh_range; x++) instead? Just old habits. @@ -234,6 +264,26 @@ range_operator::fold_range (irange , tree type, unsigned num_lh = lh.num_pairs (); unsigned num_rh = rh.num_pairs (); + // If op1 and op2 are equivalences, then we don't need a complete cross + // product, just pairs of matching elements. + if (relation_equiv_p (rel) && (lh == rh)) The ()s around lh == rh look superfluous to me. Yeah I just found it marginally more readable, but it is superfluous +{ + int_range_max tmp; + r.set_undefined (); + for (unsigned x = 0; x < num_lh; ++x) fold_range has an upper bound of num_lh * num_rh > 12, shouldn't something like that be there for this case too? I mean, every wi_fold_in_parts_equiv can result in 8 subranges, but num_lh could be up to 255 here, it is true it is linear and union_ should merge excess ones, but still I wonder if some larger num_lh upper bound like 20 or 32 wouldn't be useful. Up to you... fold_range has the num_lh * num_rh limit because it was quadratic/exponential and changes rapidly. Since this was linear based on the number of sub ranges I didn't think it would matter much, but sure, we can put a similar limit on it.. 16 seems reasonable. +{ + wide_int lh_lb = lh.lower_bound (x); + wide_int lh_ub = lh.upper_bound (x); + wi_fold_in_parts_equiv (tmp, type, lh_lb, lh_ub); + r.union_ (tmp); + if (r.varying_p ()) +break; +} + op1_op2_relation_effect (r, type, lh, rh, rel); +
Re: [PATCH] PR tree-optimization/108359 - Utilize op1 == op2 when invoking range-ops folding.
On Fri, Jan 13, 2023 at 11:07 PM Andrew MacLeod wrote: > > > On 1/13/23 16:54, Jakub Jelinek wrote: > > On Fri, Jan 13, 2023 at 04:23:20PM -0500, Andrew MacLeod via Gcc-patches > > wrote: > >> fold_range() already invokes wi_fold_in_parts to try to get more refined > >> information. If the subranges are quite small, it will do each individual > >> calculation and combine the results. > >> > >> x * y with x = [1,3] and y = [1,3] is broken down and we calculate each > >> possibility and we end up with [1,4][6,6][9,9] instead of [1,9] > >> > >> We limit this as the time is between quadratic to exponential depending on > >> the number of elements in x and y. > >> > >> If we also check the relation and determine that x == y, we don't need to > >> worry about that growth as this process is linear. The above case will be > >> broken down to just 1*1, 2*2 and 3*3, resulting in a range of [1, > >> 1][4,4][9,9]. > >> > >> In the testcase, it happens to be the right_shift operation, but this > >> solution is generic and applies to all range-op operations. I added a > >> testcase which checks >>, *, + and %. > >> > >> I also arbitrarily chose 8 elements as the limit for breaking down > >> individual operations. The overall compile time change for this is > >> negligible. > >> > >> Although this is a regression fix, it will affect all operations where x == > >> y, which is where my initial hesitancy arose. > >> > >> Regardless, bootstrapped on x86_64-pc-linux-gnu with no regressions. OK > >> for > >> trunk? > > Will defer to Aldy, just some nits. > Did you mean Richi? I suppose Aldy, since it's a regression fix it's OK even during stage4. I do have a comment as well though, you do + // If op1 and op2 are equivalences, then we don't need a complete cross + // product, just pairs of matching elements. + if (relation_equiv_p (rel) && lh == rh && num_lh <= 16) +{ + int_range_max tmp; + r.set_undefined (); + for (unsigned x = 0; x < num_lh; ++x) + { + wide_int lh_lb = lh.lower_bound (x); + wide_int lh_ub = lh.upper_bound (x); + wi_fold_in_parts_equiv (tmp, type, lh_lb, lh_ub); and that does + widest_int lh_range = wi::sub (widest_int::from (lh_ub, TYPE_SIGN (type)), +widest_int::from (lh_lb, TYPE_SIGN (type))); + // if there are 1 to 8 values in the LH range, split them up. + r.set_undefined (); + if (lh_range >= 0 && lh_range <= 7) +{ + for (unsigned x = 0; x <= lh_range; x++) which in total limits the number of sub-ranges in the output but in an odd way. It's also all-or-nothing. IIRC there's a hard limit on the number of sub-ranges in the output anyway via int_range, so why not use that and always do the first loop over the sub-ranges of the inputs and the second loop over the range members but stop when we reach N-1 and then use wi_fold on the remainds? Your above code suggests we go up to 112 sub-ranges and once we'd reach 113 we'd fold down to a single. Maybe my "heuristic" wouldn't be much better, but then somehow breaking the heuristic down to a single magic number would be better, esp. since .union_ will undo some of the breakup when reaching N? > > > >> + // if there are 1 to 8 values in the LH range, split them up. > >> + r.set_undefined (); > >> + if (lh_range >= 0 && lh_range <= 7) > >> +{ > >> + unsigned x; > >> + for (x = 0; x <= lh_range; x++) > > Nothing uses x after the loop, so why not > >for (unsigned x = 0; x <= lh_range; x++) > > instead? > > Just old habits. > > > >> @@ -234,6 +264,26 @@ range_operator::fold_range (irange , tree type, > >> unsigned num_lh = lh.num_pairs (); > >> unsigned num_rh = rh.num_pairs (); > >> > >> + // If op1 and op2 are equivalences, then we don't need a complete cross > >> + // product, just pairs of matching elements. > >> + if (relation_equiv_p (rel) && (lh == rh)) > > The ()s around lh == rh look superfluous to me. > Yeah I just found it marginally more readable, but it is superfluous > >> +{ > >> + int_range_max tmp; > >> + r.set_undefined (); > >> + for (unsigned x = 0; x < num_lh; ++x) > > fold_range has an upper bound of num_lh * num_rh > 12, shouldn't something > > like that be there for this case too? > > I mean, every wi_fold_in_parts_equiv can result in 8 subranges, > > but num_lh could be up to 255 here, it is true it is linear and union_ > > should merge excess ones, but still I wonder if some larger num_lh upper > > bound like 20 or 32 wouldn't be useful. Up to you... > fold_range has the num_lh * num_rh limit because it was > quadratic/exponential and changes rapidly. Since this was linear based > on the number of sub ranges I didn't think it would matter much, but > sure, we can put a similar limit on it.. 16 seems reasonable. > >> +{ > >> + wide_int lh_lb = lh.lower_bound (x); > >> + wide_int lh_ub = lh.upper_bound (x); > >> + wi_fold_in_parts_equiv (tmp, type,
Re: [PATCH] PR tree-optimization/108359 - Utilize op1 == op2 when invoking range-ops folding.
On 1/13/23 16:54, Jakub Jelinek wrote: On Fri, Jan 13, 2023 at 04:23:20PM -0500, Andrew MacLeod via Gcc-patches wrote: fold_range() already invokes wi_fold_in_parts to try to get more refined information. If the subranges are quite small, it will do each individual calculation and combine the results. x * y with x = [1,3] and y = [1,3] is broken down and we calculate each possibility and we end up with [1,4][6,6][9,9] instead of [1,9] We limit this as the time is between quadratic to exponential depending on the number of elements in x and y. If we also check the relation and determine that x == y, we don't need to worry about that growth as this process is linear. The above case will be broken down to just 1*1, 2*2 and 3*3, resulting in a range of [1, 1][4,4][9,9]. In the testcase, it happens to be the right_shift operation, but this solution is generic and applies to all range-op operations. I added a testcase which checks >>, *, + and %. I also arbitrarily chose 8 elements as the limit for breaking down individual operations. The overall compile time change for this is negligible. Although this is a regression fix, it will affect all operations where x == y, which is where my initial hesitancy arose. Regardless, bootstrapped on x86_64-pc-linux-gnu with no regressions. OK for trunk? Will defer to Aldy, just some nits. Did you mean Richi? + // if there are 1 to 8 values in the LH range, split them up. + r.set_undefined (); + if (lh_range >= 0 && lh_range <= 7) +{ + unsigned x; + for (x = 0; x <= lh_range; x++) Nothing uses x after the loop, so why not for (unsigned x = 0; x <= lh_range; x++) instead? Just old habits. @@ -234,6 +264,26 @@ range_operator::fold_range (irange , tree type, unsigned num_lh = lh.num_pairs (); unsigned num_rh = rh.num_pairs (); + // If op1 and op2 are equivalences, then we don't need a complete cross + // product, just pairs of matching elements. + if (relation_equiv_p (rel) && (lh == rh)) The ()s around lh == rh look superfluous to me. Yeah I just found it marginally more readable, but it is superfluous +{ + int_range_max tmp; + r.set_undefined (); + for (unsigned x = 0; x < num_lh; ++x) fold_range has an upper bound of num_lh * num_rh > 12, shouldn't something like that be there for this case too? I mean, every wi_fold_in_parts_equiv can result in 8 subranges, but num_lh could be up to 255 here, it is true it is linear and union_ should merge excess ones, but still I wonder if some larger num_lh upper bound like 20 or 32 wouldn't be useful. Up to you... fold_range has the num_lh * num_rh limit because it was quadratic/exponential and changes rapidly. Since this was linear based on the number of sub ranges I didn't think it would matter much, but sure, we can put a similar limit on it.. 16 seems reasonable. + { + wide_int lh_lb = lh.lower_bound (x); + wide_int lh_ub = lh.upper_bound (x); + wi_fold_in_parts_equiv (tmp, type, lh_lb, lh_ub); + r.union_ (tmp); + if (r.varying_p ()) + break; + } + op1_op2_relation_effect (r, type, lh, rh, rel); + update_known_bitmask (r, m_code, lh, rh); + return true; +} + Jakub Updated patch attached... I'll run it through testing whe the current one is done. Andrew From 42010868ff3cdbb5b9ad3484f115b7c23f9e14e7 Mon Sep 17 00:00:00 2001 From: Andrew MacLeod Date: Wed, 11 Jan 2023 18:12:51 -0500 Subject: [PATCH 2/3] Utilize op1 == op2 when invoking range-ops folding. If there exists an equivalence relationship between op1 and op2, any binary operation can be broken into individual operations and unioned if there are sufficently few elements in the set. PR tree-optimization/108359 gcc/ * range-op.cc (range_operator::wi_fold_in_parts_equiv): New. (range_operator::fold_range): If op1 is equivalent to op2 then invoke new fold_in_parts_equiv to operate on sub-components. * range-op.h (wi_fold_in_parts_equiv): New prototype. gcc/testsuite/ * gcc.dg/pr108359.c: New. --- gcc/range-op.cc | 49 gcc/range-op.h | 5 gcc/testsuite/gcc.dg/pr108359.c | 50 + 3 files changed, 104 insertions(+) create mode 100644 gcc/testsuite/gcc.dg/pr108359.c diff --git a/gcc/range-op.cc b/gcc/range-op.cc index ec75e07bc8a..33bc4dcb4b4 100644 --- a/gcc/range-op.cc +++ b/gcc/range-op.cc @@ -160,6 +160,35 @@ range_operator::wi_fold (irange , tree type, r.set_varying (type); } +// Call wi_fold when both op1 and op2 are equivalent. Further split small +// subranges into constants. This can provide better precision. +// For x + y, when x == y with a range of [0,4] instead of [0, 8] produce +// [0,0][2, 2][4,4][6, 6][8, 8] + +void +range_operator::wi_fold_in_parts_equiv (irange , tree type, + const wide_int _lb, + const wide_int _ub) const +{ + int_range_max tmp; +
Re: [PATCH] PR tree-optimization/108359 - Utilize op1 == op2 when invoking range-ops folding.
On Fri, Jan 13, 2023 at 04:23:20PM -0500, Andrew MacLeod via Gcc-patches wrote: > fold_range() already invokes wi_fold_in_parts to try to get more refined > information. If the subranges are quite small, it will do each individual > calculation and combine the results. > > x * y with x = [1,3] and y = [1,3] is broken down and we calculate each > possibility and we end up with [1,4][6,6][9,9] instead of [1,9] > > We limit this as the time is between quadratic to exponential depending on > the number of elements in x and y. > > If we also check the relation and determine that x == y, we don't need to > worry about that growth as this process is linear. The above case will be > broken down to just 1*1, 2*2 and 3*3, resulting in a range of [1, > 1][4,4][9,9]. > > In the testcase, it happens to be the right_shift operation, but this > solution is generic and applies to all range-op operations. I added a > testcase which checks >>, *, + and %. > > I also arbitrarily chose 8 elements as the limit for breaking down > individual operations. The overall compile time change for this is > negligible. > > Although this is a regression fix, it will affect all operations where x == > y, which is where my initial hesitancy arose. > > Regardless, bootstrapped on x86_64-pc-linux-gnu with no regressions. OK for > trunk? Will defer to Aldy, just some nits. > + // if there are 1 to 8 values in the LH range, split them up. > + r.set_undefined (); > + if (lh_range >= 0 && lh_range <= 7) > +{ > + unsigned x; > + for (x = 0; x <= lh_range; x++) Nothing uses x after the loop, so why not for (unsigned x = 0; x <= lh_range; x++) instead? > @@ -234,6 +264,26 @@ range_operator::fold_range (irange , tree type, >unsigned num_lh = lh.num_pairs (); >unsigned num_rh = rh.num_pairs (); > > + // If op1 and op2 are equivalences, then we don't need a complete cross > + // product, just pairs of matching elements. > + if (relation_equiv_p (rel) && (lh == rh)) The ()s around lh == rh look superfluous to me. > +{ > + int_range_max tmp; > + r.set_undefined (); > + for (unsigned x = 0; x < num_lh; ++x) fold_range has an upper bound of num_lh * num_rh > 12, shouldn't something like that be there for this case too? I mean, every wi_fold_in_parts_equiv can result in 8 subranges, but num_lh could be up to 255 here, it is true it is linear and union_ should merge excess ones, but still I wonder if some larger num_lh upper bound like 20 or 32 wouldn't be useful. Up to you... > + { > + wide_int lh_lb = lh.lower_bound (x); > + wide_int lh_ub = lh.upper_bound (x); > + wi_fold_in_parts_equiv (tmp, type, lh_lb, lh_ub); > + r.union_ (tmp); > + if (r.varying_p ()) > + break; > + } > + op1_op2_relation_effect (r, type, lh, rh, rel); > + update_known_bitmask (r, m_code, lh, rh); > + return true; > +} > + Jakub