Message ID | a38f4b49-dd7b-320f-444a-e677f76281fe@redhat.com |
---|---|
Headers | show |
Series | tree-optimization/104530 - proposed re-evaluation. | expand |
On Tue, Feb 22, 2022 at 11:39:41AM -0500, Andrew MacLeod wrote: > I'd like to get clarification on some subtle terminology. I find I am > conflating calls that don't return with calls that may throw, and I think > they have different considerations. > > My experiments with calls that can throw indicate that they always end a > basic block. This makes sense to me as there is the outgoing fall-thru edge > and an outgoing EH edge. Are there any conditions under which this is not > the case? (other than non-call exceptions) Generally, there are 2 kinds of calls that can throw, those that can throw internally and those can throw externally (e.g. there are stmt_could_throw_{in,ex}ternal predicates). Consider e.g. void foo (); struct S { S (); ~S (); }; void bar () { foo (); foo (); } void baz () { S s; foo (); foo (); } void qux () { try { foo (); } catch (...) {} } the calls to foo in bar throw externally, if they throw, execution doesn't continue anywhere in bar but in some bar's caller, or could just terminate if nothing catches it at all. Such calls don't terminate a bb. In baz, the s variable needs destruction if either of the foo calls throw, so those calls do terminate bb and there are normal fallthru edges from those bbs and eh edges to an EH pad which will destruct s and continue propagating the exception. In qux, there is explicit try/catch, so again, foo throws internally, ends bb, has an EH edge to EH landing pad which will do what catch does. That is EH, then there are calls that might not return because they leave in some other way (e.g. longjmp), or might loop forever, might exit, might abort, trap etc. I must say I don't know if we have any call flags that would guarantee the function will always return (exactly once) if called. Perhaps ECF_CONST/EFC_PURE without ECF_LOOPING_CONST_OR_PURE do? Jakub
On 2/22/22 11:56, Jakub Jelinek wrote: > On Tue, Feb 22, 2022 at 11:39:41AM -0500, Andrew MacLeod wrote: >> I'd like to get clarification on some subtle terminology. I find I am >> conflating calls that don't return with calls that may throw, and I think >> they have different considerations. >> >> My experiments with calls that can throw indicate that they always end a >> basic block. This makes sense to me as there is the outgoing fall-thru edge >> and an outgoing EH edge. Are there any conditions under which this is not >> the case? (other than non-call exceptions) > Generally, there are 2 kinds of calls that can throw, those that can throw > internally and those can throw externally (e.g. there are > stmt_could_throw_{in,ex}ternal predicates). > > Consider e.g. > > void foo (); > struct S { S (); ~S (); }; > void bar () { foo (); foo (); } > void baz () { S s; foo (); foo (); } > void qux () { try { foo (); } catch (...) {} } > > the calls to foo in bar throw externally, if they throw, execution doesn't > continue anywhere in bar but in some bar's caller, or could just terminate > if nothing catches it at all. Such calls don't terminate a bb. This is not a problem. > In baz, the s variable needs destruction if either of the foo calls throw, > so those calls do terminate bb and there are normal fallthru edges from > those bbs and eh edges to an EH pad which will destruct s and continue > propagating the exception. > In qux, there is explicit try/catch, so again, foo throws internally, ends > bb, has an EH edge to EH landing pad which will do what catch does. Those also are not a problem, everything should flow fine in these situations as well now that we make non-null adjustments on edges, and don't for EH edges. As far as these patches go, any block which has a call at the exit point will not have any import or exports as there is no range stmt at the end of the block, so we will not be marking anything in those blocks as stale. > > That is EH, then there are calls that might not return because they leave > in some other way (e.g. longjmp), or might loop forever, might exit, might > abort, trap etc. Generally speaking, calls which do not return should not now be a problem... as long as they do not transfer control to somewhere else in the current function. > I must say I don't know if we have any call flags that would guarantee > the function will always return (exactly once) if called. > Perhaps ECF_CONST/EFC_PURE without ECF_LOOPING_CONST_OR_PURE do? > I don't think I actually need that. Andrew
On Tue, Feb 22, 2022 at 12:39:28PM -0500, Andrew MacLeod wrote: > > That is EH, then there are calls that might not return because they leave > > in some other way (e.g. longjmp), or might loop forever, might exit, might > > abort, trap etc. > Generally speaking, calls which do not return should not now be a problem... > as long as they do not transfer control to somewhere else in the current > function. I thought all of those cases are very relevant to PR104530. If we have: _1 = ptr_2(D) == 0; // unrelated code in the same bb _3 = *ptr_2(D); then in light of PR104288, we can optimize ptr_2(D) == 0 into true only if there are no calls inside of "// unrelated code in the same bb" or if all calls in "// unrelated code in the same bb" are guaranteed to return exactly once. Because, if there is a call in there which could exit (that is the PR104288 testcase), or abort, or trap, or loop forever, or throw externally, or longjmp or in any other non-UB way cause the _1 = ptr_2(D) == 0; stmt to be invoked at runtime but _3 = *ptr_2(D) not being invoked, then we can't optimize the earlier comparison because ptr_2(D) could be NULL in a valid program. While if there are no calls (and no problematic inline asms) and no trapping insns in between, we can and PR104530 is asking that we continue to optimize that. Jakub
On 2/22/2022 10:57 AM, Jakub Jelinek via Gcc-patches wrote: > On Tue, Feb 22, 2022 at 12:39:28PM -0500, Andrew MacLeod wrote: >>> That is EH, then there are calls that might not return because they leave >>> in some other way (e.g. longjmp), or might loop forever, might exit, might >>> abort, trap etc. >> Generally speaking, calls which do not return should not now be a problem... >> as long as they do not transfer control to somewhere else in the current >> function. > I thought all of those cases are very relevant to PR104530. > If we have: > _1 = ptr_2(D) == 0; > // unrelated code in the same bb > _3 = *ptr_2(D); > then in light of PR104288, we can optimize ptr_2(D) == 0 into true only if > there are no calls inside of "// unrelated code in the same bb" > or if all calls in "// unrelated code in the same bb" are guaranteed to > return exactly once. Because, if there is a call in there which could > exit (that is the PR104288 testcase), or abort, or trap, or loop forever, > or throw externally, or longjmp or in any other non-UB way > cause the _1 = ptr_2(D) == 0; stmt to be invoked at runtime but > _3 = *ptr_2(D) not being invoked, then we can't optimize the earlier > comparison because ptr_2(D) could be NULL in a valid program. > While if there are no calls (and no problematic inline asms) and no trapping > insns in between, we can and PR104530 is asking that we continue to optimize > that. Right. This is similar to some of the restrictions we deal with in the path isolation pass. Essentially we have a path, when traversed, would result in a *0. We would like to be able to find the edge upon-which the *0 is control dependent and optimize the test so that it always went to the valid path rather than the *0 path. The problem is there may be observable side effects on the *0 path between the test and the actual *0 -- including calls to nonreturning functions, setjmp/longjmp, things that could trap, etc. This case is similar. We can't back-propagate the non-null status through any statements with observable side effects. Jeff
On 2/22/22 13:07, Jeff Law wrote: > > > On 2/22/2022 10:57 AM, Jakub Jelinek via Gcc-patches wrote: >> On Tue, Feb 22, 2022 at 12:39:28PM -0500, Andrew MacLeod wrote: >>>> That is EH, then there are calls that might not return because they >>>> leave >>>> in some other way (e.g. longjmp), or might loop forever, might >>>> exit, might >>>> abort, trap etc. >>> Generally speaking, calls which do not return should not now be a >>> problem... >>> as long as they do not transfer control to somewhere else in the >>> current >>> function. >> I thought all of those cases are very relevant to PR104530. >> If we have: >> _1 = ptr_2(D) == 0; >> // unrelated code in the same bb >> _3 = *ptr_2(D); >> then in light of PR104288, we can optimize ptr_2(D) == 0 into true >> only if >> there are no calls inside of "// unrelated code in the same bb" >> or if all calls in "// unrelated code in the same bb" are guaranteed to >> return exactly once. Because, if there is a call in there which could >> exit (that is the PR104288 testcase), or abort, or trap, or loop >> forever, >> or throw externally, or longjmp or in any other non-UB way >> cause the _1 = ptr_2(D) == 0; stmt to be invoked at runtime but >> _3 = *ptr_2(D) not being invoked, then we can't optimize the earlier >> comparison because ptr_2(D) could be NULL in a valid program. >> While if there are no calls (and no problematic inline asms) and no >> trapping >> insns in between, we can and PR104530 is asking that we continue to >> optimize >> that. > Right. This is similar to some of the restrictions we deal with in > the path isolation pass. Essentially we have a path, when traversed, > would result in a *0. We would like to be able to find the edge > upon-which the *0 is control dependent and optimize the test so that > it always went to the valid path rather than the *0 path. > > The problem is there may be observable side effects on the *0 path > between the test and the actual *0 -- including calls to nonreturning > functions, setjmp/longjmp, things that could trap, etc. This case is > similar. We can't back-propagate the non-null status through any > statements with observable side effects. > > Jeff > We can't back propagate, but we can alter our forward view. Any ssa-name defined before the observable side effect can be recalculated using the updated values, and all uses of those names after the side-effect would then appear to be "up-to-date" This does not actually change anything before the side-effect statement, but the lazy re-evalaution ranger employs makes it appear as if we do a new computation when _1 is used afterwards. ie: _1 = ptr_2(D) == 0; // unrelated code in the same bb _3 = *ptr_2(D); _4 = ptr_2(D) == 0; // ptr_2 is known to be [+1, +INF] now. And we use _4 everywhere _1 was used. This is the effect. so we do not actually change anything in the unrelated code, just observable effects afterwards. We already do these recalculations on outgoing edges in other blocks, just not within the definition block because non-null wasn't visible within the def block. Additionally, In the testcase, there is a store to C before the side effects. these patches get rid of the branch and thus the call in the testcase as requested, but we still have to compute _3 in order to store it into global C since it occurs pre side-effect. b.0_1 = b; _2 = b.0_1 == 0B; _3 = (int) _2; c = _3; _5 = *b.0_1; No matter how you look at it, you are going to need to process a block twice in order to handle any code pre-side-effect. Whether it be assigning stmt uids, or what have you. VRP could pre-process the block, and if it gets to the end of the block, and it had at least one statement with a side effect and no calls which may not return you could process the block with all the side effects already active. I'm not sure if that buys as much as the cost, but it would change the value written to C to be 1, and it would change the global values exported for _2 and _3. Another option would be flag the ssa-names instead of/as well as marking them as stale. If we get to the end of the block and there were no non-returning functions or EH edges, then re-calculate and export those ssa_names using the latest values.. That would export [0,0] for _2 and _3. This would have no tangible impact during the first VRP pass, but the *next* VRP pass, (or any other ranger pass) would pick up the new global ranges, and do all the right things... so we basically let a subsequent pass pick up the info and do the dirty work. Andrew
On Tue, Feb 22, 2022 at 8:19 PM Andrew MacLeod via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > On 2/22/22 13:07, Jeff Law wrote: > > > > > > On 2/22/2022 10:57 AM, Jakub Jelinek via Gcc-patches wrote: > >> On Tue, Feb 22, 2022 at 12:39:28PM -0500, Andrew MacLeod wrote: > >>>> That is EH, then there are calls that might not return because they > >>>> leave > >>>> in some other way (e.g. longjmp), or might loop forever, might > >>>> exit, might > >>>> abort, trap etc. > >>> Generally speaking, calls which do not return should not now be a > >>> problem... > >>> as long as they do not transfer control to somewhere else in the > >>> current > >>> function. > >> I thought all of those cases are very relevant to PR104530. > >> If we have: > >> _1 = ptr_2(D) == 0; > >> // unrelated code in the same bb > >> _3 = *ptr_2(D); > >> then in light of PR104288, we can optimize ptr_2(D) == 0 into true > >> only if > >> there are no calls inside of "// unrelated code in the same bb" > >> or if all calls in "// unrelated code in the same bb" are guaranteed to > >> return exactly once. Because, if there is a call in there which could > >> exit (that is the PR104288 testcase), or abort, or trap, or loop > >> forever, > >> or throw externally, or longjmp or in any other non-UB way > >> cause the _1 = ptr_2(D) == 0; stmt to be invoked at runtime but > >> _3 = *ptr_2(D) not being invoked, then we can't optimize the earlier > >> comparison because ptr_2(D) could be NULL in a valid program. > >> While if there are no calls (and no problematic inline asms) and no > >> trapping > >> insns in between, we can and PR104530 is asking that we continue to > >> optimize > >> that. > > Right. This is similar to some of the restrictions we deal with in > > the path isolation pass. Essentially we have a path, when traversed, > > would result in a *0. We would like to be able to find the edge > > upon-which the *0 is control dependent and optimize the test so that > > it always went to the valid path rather than the *0 path. > > > > The problem is there may be observable side effects on the *0 path > > between the test and the actual *0 -- including calls to nonreturning > > functions, setjmp/longjmp, things that could trap, etc. This case is > > similar. We can't back-propagate the non-null status through any > > statements with observable side effects. > > > > Jeff > > > We can't back propagate, but we can alter our forward view. Any > ssa-name defined before the observable side effect can be recalculated > using the updated values, and all uses of those names after the > side-effect would then appear to be "up-to-date" > > This does not actually change anything before the side-effect statement, > but the lazy re-evalaution ranger employs makes it appear as if we do a > new computation when _1 is used afterwards. ie: > > _1 = ptr_2(D) == 0; > // unrelated code in the same bb > _3 = *ptr_2(D); > _4 = ptr_2(D) == 0; // ptr_2 is known to be [+1, +INF] now. > And we use _4 everywhere _1 was used. This is the effect. > > so we do not actually change anything in the unrelated code, just > observable effects afterwards. We already do these recalculations on > outgoing edges in other blocks, just not within the definition block > because non-null wasn't visible within the def block. > > Additionally, In the testcase, there is a store to C before the side > effects. > these patches get rid of the branch and thus the call in the testcase as > requested, but we still have to compute _3 in order to store it into > global C since it occurs pre side-effect. > > b.0_1 = b; > _2 = b.0_1 == 0B; > _3 = (int) _2; > c = _3; > _5 = *b.0_1; > > No matter how you look at it, you are going to need to process a block > twice in order to handle any code pre-side-effect. Whether it be > assigning stmt uids, or what have you. Yes. I thought that is what ranger already does when it discovers new ranges from edges. Say we have _1 = 10 / _2; if (_2 == 1) { _3 = _1 + 1; then when evaluating _1 + 1 we re-evaluate 10 / _2 using _2 == 1 and can compute _3 to [11, 11]? That obviously extends to any stmt-level ranges we discover for uses (not defs because defs are never used upthread). And doing that is _not_ affected by any function/BB terminating calls or EH or whatnot as long as the updated ranges are only affecting stmts dominating the current one. What complicates all this reasoning is that it is straight-forward when you work with a traditional IL walking pass but it gets hard (and possibly easy to get wrong) with on-demand processing and caching because everything you cache will now be context dependent (valid only starting after stmt X and for stmts dominated by it). > VRP could pre-process the block, and if it gets to the end of the block, > and it had at least one statement with a side effect and no calls which > may not return you could process the block with all the side effects > already active. I'm not sure if that buys as much as the cost, but it > would change the value written to C to be 1, and it would change the > global values exported for _2 and _3. > > Another option would be flag the ssa-names instead of/as well as marking > them as stale. If we get to the end of the block and there were no > non-returning functions or EH edges, then re-calculate and export those > ssa_names using the latest values.. That would export [0,0] for _2 and _3. > > This would have no tangible impact during the first VRP pass, but the > *next* VRP pass, (or any other ranger pass) would pick up the new global > ranges, and do all the right things... so we basically let a subsequent > pass pick up the info and do the dirty work. > > Andrew > > > > > > > > >
On 2/23/22 02:48, Richard Biener wrote: > On Tue, Feb 22, 2022 at 8:19 PM Andrew MacLeod via Gcc-patches > <gcc-patches@gcc.gnu.org> wrote: >> On 2/22/22 13:07, Jeff Law wrote: >>> >>> On 2/22/2022 10:57 AM, Jakub Jelinek via Gcc-patches wrote: >>>> On Tue, Feb 22, 2022 at 12:39:28PM -0500, Andrew MacLeod wrote: >>>>>> That is EH, then there are calls that might not return because they >>>>>> leave >>>>>> in some other way (e.g. longjmp), or might loop forever, might >>>>>> exit, might >>>>>> abort, trap etc. >>>>> Generally speaking, calls which do not return should not now be a >>>>> problem... >>>>> as long as they do not transfer control to somewhere else in the >>>>> current >>>>> function. >>>> I thought all of those cases are very relevant to PR104530. >>>> If we have: >>>> _1 = ptr_2(D) == 0; >>>> // unrelated code in the same bb >>>> _3 = *ptr_2(D); >>>> then in light of PR104288, we can optimize ptr_2(D) == 0 into true >>>> only if >>>> there are no calls inside of "// unrelated code in the same bb" >>>> or if all calls in "// unrelated code in the same bb" are guaranteed to >>>> return exactly once. Because, if there is a call in there which could >>>> exit (that is the PR104288 testcase), or abort, or trap, or loop >>>> forever, >>>> or throw externally, or longjmp or in any other non-UB way >>>> cause the _1 = ptr_2(D) == 0; stmt to be invoked at runtime but >>>> _3 = *ptr_2(D) not being invoked, then we can't optimize the earlier >>>> comparison because ptr_2(D) could be NULL in a valid program. >>>> While if there are no calls (and no problematic inline asms) and no >>>> trapping >>>> insns in between, we can and PR104530 is asking that we continue to >>>> optimize >>>> that. >>> Right. This is similar to some of the restrictions we deal with in >>> the path isolation pass. Essentially we have a path, when traversed, >>> would result in a *0. We would like to be able to find the edge >>> upon-which the *0 is control dependent and optimize the test so that >>> it always went to the valid path rather than the *0 path. >>> >>> The problem is there may be observable side effects on the *0 path >>> between the test and the actual *0 -- including calls to nonreturning >>> functions, setjmp/longjmp, things that could trap, etc. This case is >>> similar. We can't back-propagate the non-null status through any >>> statements with observable side effects. >>> >>> Jeff >>> >> We can't back propagate, but we can alter our forward view. Any >> ssa-name defined before the observable side effect can be recalculated >> using the updated values, and all uses of those names after the >> side-effect would then appear to be "up-to-date" >> >> This does not actually change anything before the side-effect statement, >> but the lazy re-evalaution ranger employs makes it appear as if we do a >> new computation when _1 is used afterwards. ie: >> >> _1 = ptr_2(D) == 0; >> // unrelated code in the same bb >> _3 = *ptr_2(D); >> _4 = ptr_2(D) == 0; // ptr_2 is known to be [+1, +INF] now. >> And we use _4 everywhere _1 was used. This is the effect. >> >> so we do not actually change anything in the unrelated code, just >> observable effects afterwards. We already do these recalculations on >> outgoing edges in other blocks, just not within the definition block >> because non-null wasn't visible within the def block. >> >> Additionally, In the testcase, there is a store to C before the side >> effects. >> these patches get rid of the branch and thus the call in the testcase as >> requested, but we still have to compute _3 in order to store it into >> global C since it occurs pre side-effect. >> >> b.0_1 = b; >> _2 = b.0_1 == 0B; >> _3 = (int) _2; >> c = _3; >> _5 = *b.0_1; >> >> No matter how you look at it, you are going to need to process a block >> twice in order to handle any code pre-side-effect. Whether it be >> assigning stmt uids, or what have you. > Yes. I thought that is what ranger already does when it discovers new > ranges from edges. Say we have > > _1 = 10 / _2; > if (_2 == 1) > { > _3 = _1 + 1; > > then when evaluating _1 + 1 we re-evaluate 10 / _2 using _2 == 1 and > can compute _3 to [11, 11]? Correct, we get most of these first order effects via edges. > > That obviously extends to any stmt-level ranges we discover for uses > (not defs because defs are never used upthread). And doing that is > _not_ affected by any function/BB terminating calls or EH or whatnot > as long as the updated ranges are only affecting stmts dominating the > current one. > > What complicates all this reasoning is that it is straight-forward when > you work with a traditional IL walking pass but it gets hard (and possibly > easy to get wrong) with on-demand processing and caching because > everything you cache will now be context dependent (valid only > starting after stmt X and for stmts dominated by it). Yeah, which is why this particular side effect code only applies to definitions during a dom walk. we know we will not return to a def. The non-null list (and next release the generalized side-effects) are only applied to on-exit ranges via non-EH edges.. so they cant really get us into trouble as we are sure of those values only affecting dominated blocks. Pure on-demand clients will not get any of this intra-block fine tuning.