Hal Finkel via llvm-dev
2017-Jan-25 23:03 UTC
[llvm-dev] RFC: Emitting empty invariant group for vtable loads
Hi Piotr, I think makes sense. Modulo bitcasts, the invariant is identified by a particular pointer SSA value. Given that you can't sensibly have two nonequivalent invariants associated with the same pointer SSA value simultaneously, there's no need to also identify the invariant with a metadata string as well. When we need a new "identifier" for the pointed-to value, we get one using invariant.group.barrier. -Hal On 01/24/2017 01:39 PM, Piotr Padlewski via llvm-dev wrote:> Hi, > I would really like to hear some feedback about this. > > Piotr > > 2017-01-20 17:07 GMT+01:00 Piotr Padlewski <piotr.padlewski at gmail.com > <mailto:piotr.padlewski at gmail.com>>: > > Hi all, > > I would like to propose a new way clang would decorate vtable > loads in order to handle devirtualization better. > > I've added *llvm-dev* also, because this can start a discussion > about changing invariant.group to just invariant. > > PDF version of this RFC can be found here: > > https://drive.google.com/file/d/0B72TmzNsY6Z8ZmpOUnB5dDZfSFU/view?usp=sharing > <https://drive.google.com/file/d/0B72TmzNsY6Z8ZmpOUnB5dDZfSFU/view?usp=sharing> > > > Background: > > Initial old design: > > http://lists.llvm.org/pipermail/cfe-dev/2015-July/044227.html > <http://lists.llvm.org/pipermail/cfe-dev/2015-July/044227.html> > > My talk from LLVM Dev Meeting > > http://llvm.org/devmtg/2016-11/#talk6 > <http://llvm.org/devmtg/2016-11/#talk6> > > > The problem > > Clang with -fstrict-vtable-pointers decorates vtable loads with > metadata corresponding to mangled pointer type name like: > > voidg(A& a){ a.foo();} > > define void at _Z1gR1A(%struct.A* dereferenceable(8) %a) > local_unnamed_addr #0{entry: %0= bitcast %struct.A* %a to > void(%struct.A*)*** %vtable = load void(%struct.A*)**, > void(%struct.A*)*** %0, !invariant.group !7 %1= load > void(%struct.A*)*, void(%struct.A*)** %vtable tail call > void%1(%struct.A* nonnull %a) ret void}!7= !{!"_ZTS1A"} > > This works well if the pointer type doesn’t change, but when it > does, devirtualization might not happen like here: > > structA { A();virtualvoidfoo();};structB : > A{ B();virtualvoidfoo();};voidg(A& > a){ a.foo(); a.foo();}voidclobber(A&);voidf() { B > b; clobber(b); g(b);} > > The other problem is that when we combine 2 instructions with > different invariant.group metadata, then we pick one of them, > because for now we can have only single !invariant.group metadata. > > > The solution > > I had some initial ideas how it can be solved, like > > 1. > > introducing multi invariant groups > > 2. > > having sub invariant groups - like inheritance, so we could > figure out that one group is subgroup of another > > 3. > > decorating all loads with base pointer MD (doesn’t work with > multiple inheritance) > > I consulted my ideas with Krzysztof Pszeniczny, and he proposed > something much simpler: we can decorate every invariant.group md > with empty metadata. > > This should work because the lifetime of the object is strictly > defined by invariant.group.barrier. > > If this holds, we can start discussion about if it makes sense to > keep invariant groups, and instead have just “invariant”, that > would be equivalent to having invariant.group with the same metadata. > > Do you have some thoughts about this approach? I don’t have a > mathematical proof, but I am confident that it should be valid. > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev-- Hal Finkel Lead, Compiler Technology and Programming Languages Leadership Computing Facility Argonne National Laboratory -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170125/7a93830a/attachment.html>
Piotr Padlewski via llvm-dev
2017-Jan-26 00:05 UTC
[llvm-dev] RFC: Emitting empty invariant group for vtable loads
2017-01-26 0:03 GMT+01:00 Hal Finkel <hfinkel at anl.gov>:> Hi Piotr, > > I think makes sense. Modulo bitcasts, the invariant is identified by a > particular pointer SSA value. Given that you can't sensibly have two > nonequivalent invariants associated with the same pointer SSA value > simultaneously, there's no need to also identify the invariant with a > metadata string as well. When we need a new "identifier" for the pointed-to > value, we get one using invariant.group.barrier. > > -Hal >What is your opinion about changing invariant.group to just invariant? Piotr> On 01/24/2017 01:39 PM, Piotr Padlewski via llvm-dev wrote: > > Hi, > I would really like to hear some feedback about this. > > Piotr > > 2017-01-20 17:07 GMT+01:00 Piotr Padlewski <piotr.padlewski at gmail.com>: > >> Hi all, >> >> I would like to propose a new way clang would decorate vtable loads in >> order to handle devirtualization better. >> >> I've added *llvm-dev* also, because this can start a discussion about >> changing invariant.group to just invariant. >> >> PDF version of this RFC can be found here: >> >> https://drive.google.com/file/d/0B72TmzNsY6Z8ZmpOUnB5dDZfSFU >> /view?usp=sharing >> >> Background: >> >> Initial old design: >> >> http://lists.llvm.org/pipermail/cfe-dev/2015-July/044227.html >> >> My talk from LLVM Dev Meeting >> >> http://llvm.org/devmtg/2016-11/#talk6 >> The problem >> >> Clang with -fstrict-vtable-pointers decorates vtable loads with metadata >> corresponding to mangled pointer type name like: >> >> void g(A& a){ a.foo(); } >> >> define void @_Z1gR1A(%struct.A* dereferenceable(8) %a) >> local_unnamed_addr #0 { entry: %0 = bitcast %struct.A* %a to void (% >> struct.A*)*** %vtable = load void (%struct.A*)**, void (%struct.A*)*** % >> 0, !invariant.group !7 %1 = load void (%struct.A*)*, void (%struct.A*)** >> %vtable tail call void %1(%struct.A* nonnull %a) ret void } !7 = !{! >> "_ZTS1A"} >> >> This works well if the pointer type doesn’t change, but when it does, >> devirtualization might not happen like here: >> >> struct A { A(); virtual void foo(); }; struct B : A{ B(); >> virtual void foo(); }; void g(A& a){ a.foo(); a.foo(); } void >> clobber(A&); void f() { B b; clobber(b); g(b); } >> >> The other problem is that when we combine 2 instructions with different >> invariant.group metadata, then we pick one of them, because for now we can >> have only single !invariant.group metadata. >> The solution >> >> I had some initial ideas how it can be solved, like >> >> 1. >> >> introducing multi invariant groups >> 2. >> >> having sub invariant groups - like inheritance, so we could figure >> out that one group is subgroup of another >> 3. >> >> decorating all loads with base pointer MD (doesn’t work with multiple >> inheritance) >> >> I consulted my ideas with Krzysztof Pszeniczny, and he proposed something >> much simpler: we can decorate every invariant.group md with empty metadata. >> >> This should work because the lifetime of the object is strictly defined >> by invariant.group.barrier. >> >> If this holds, we can start discussion about if it makes sense to keep >> invariant groups, and instead have just “invariant”, that would be >> equivalent to having invariant.group with the same metadata. >> >> Do you have some thoughts about this approach? I don’t have a >> mathematical proof, but I am confident that it should be valid. >> > _______________________________________________ > LLVM Developers mailing listllvm-dev at lists.llvm.orghttp://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev > > -- > Hal Finkel > Lead, Compiler Technology and Programming Languages > Leadership Computing Facility > Argonne National Laboratory > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170126/96c7e7ad/attachment-0001.html>
Hal Finkel via llvm-dev
2017-Jan-26 00:11 UTC
[llvm-dev] RFC: Emitting empty invariant group for vtable loads
On 01/25/2017 06:05 PM, Piotr Padlewski wrote:> > > 2017-01-26 0:03 GMT+01:00 Hal Finkel <hfinkel at anl.gov > <mailto:hfinkel at anl.gov>>: > > Hi Piotr, > > I think makes sense. Modulo bitcasts, the invariant is identified > by a particular pointer SSA value. Given that you can't sensibly > have two nonequivalent invariants associated with the same pointer > SSA value simultaneously, there's no need to also identify the > invariant with a metadata string as well. When we need a new > "identifier" for the pointed-to value, we get one using > invariant.group.barrier. > > -Hal > > What is your opinion about changing invariant.group to just invariant?You mean changing !invariant.group -> !invariant and changing @llvm.invariant.group.barrier to @llvm.invariant.barrier? I don't have a strong opinion, but I'm inclined to leave it as-is (saying "group" implies that there might be things outside the group, which is true in this case). -Hal> > Piotr > > On 01/24/2017 01:39 PM, Piotr Padlewski via llvm-dev wrote: >> Hi, >> I would really like to hear some feedback about this. >> >> Piotr >> >> 2017-01-20 17:07 GMT+01:00 Piotr Padlewski >> <piotr.padlewski at gmail.com <mailto:piotr.padlewski at gmail.com>>: >> >> Hi all, >> >> I would like to propose a new way clang would decorate vtable >> loads in order to handle devirtualization better. >> >> I've added *llvm-dev* also, because this can start a >> discussion about changing invariant.group to just invariant. >> >> PDF version of this RFC can be found here: >> >> https://drive.google.com/file/d/0B72TmzNsY6Z8ZmpOUnB5dDZfSFU/view?usp=sharing >> <https://drive.google.com/file/d/0B72TmzNsY6Z8ZmpOUnB5dDZfSFU/view?usp=sharing> >> >> >> Background: >> >> Initial old design: >> >> http://lists.llvm.org/pipermail/cfe-dev/2015-July/044227.html >> <http://lists.llvm.org/pipermail/cfe-dev/2015-July/044227.html> >> >> My talk from LLVM Dev Meeting >> >> http://llvm.org/devmtg/2016-11/#talk6 >> <http://llvm.org/devmtg/2016-11/#talk6> >> >> >> The problem >> >> Clang with -fstrict-vtable-pointers decorates vtable loads >> with metadata corresponding to mangled pointer type name like: >> >> voidg(A& a){ a.foo();} >> >> define void at _Z1gR1A(%struct.A* dereferenceable(8) %a) >> local_unnamed_addr #0{entry: %0= bitcast %struct.A* %a to >> void(%struct.A*)*** %vtable = load void(%struct.A*)**, >> void(%struct.A*)*** %0, !invariant.group !7 %1= load >> void(%struct.A*)*, void(%struct.A*)** %vtable tail call >> void%1(%struct.A* nonnull %a) ret void}!7= !{!"_ZTS1A"} >> >> This works well if the pointer type doesn’t change, but when >> it does, devirtualization might not happen like here: >> >> structA { A();virtualvoidfoo();};structB : >> A{ B();virtualvoidfoo();};voidg(A& >> a){ a.foo(); a.foo();}voidclobber(A&);voidf() { B >> b; clobber(b); g(b);} >> >> The other problem is that when we combine 2 instructions with >> different invariant.group metadata, then we pick one of them, >> because for now we can have only single !invariant.group >> metadata. >> >> >> The solution >> >> I had some initial ideas how it can be solved, like >> >> 1. >> >> introducing multi invariant groups >> >> 2. >> >> having sub invariant groups - like inheritance, so we >> could figure out that one group is subgroup of another >> >> 3. >> >> decorating all loads with base pointer MD (doesn’t work >> with multiple inheritance) >> >> I consulted my ideas with Krzysztof Pszeniczny, and he >> proposed something much simpler: we can decorate every >> invariant.group md with empty metadata. >> >> This should work because the lifetime of the object is >> strictly defined by invariant.group.barrier. >> >> If this holds, we can start discussion about if it makes >> sense to keep invariant groups, and instead have just >> “invariant”, that would be equivalent to having >> invariant.group with the same metadata. >> >> Do you have some thoughts about this approach? I don’t have a >> mathematical proof, but I am confident that it should be valid. >> >> _______________________________________________ >> LLVM Developers mailing list >> llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> >> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >> <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev> > > -- > Hal Finkel > Lead, Compiler Technology and Programming Languages > Leadership Computing Facility > Argonne National Laboratory >-- Hal Finkel Lead, Compiler Technology and Programming Languages Leadership Computing Facility Argonne National Laboratory -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170125/fa1526c4/attachment.html>
Richard Smith via llvm-dev
2017-Jan-26 02:28 UTC
[llvm-dev] [cfe-dev] RFC: Emitting empty invariant group for vtable loads
On 25 January 2017 at 15:03, Hal Finkel via cfe-dev <cfe-dev at lists.llvm.org> wrote:> Hi Piotr, > > I think makes sense. Modulo bitcasts, the invariant is identified by a > particular pointer SSA value. Given that you can't sensibly have two > nonequivalent invariants associated with the same pointer SSA value > simultaneously, there's no need to also identify the invariant with a > metadata string as well. When we need a new "identifier" for the pointed-to > value, we get one using invariant.group.barrier. >As I recall, the original motivation for the identifier was to support cases where the "invariant" region's value changes and then changes back (remember that invariant.group does not imply the storage doesn't change, just that a particular set of loads and stores witness the same value): struct A { void *vptr; /*...*/ }; struct B { void *vptr; /*...*/ }; union U { A a; B b; }; void f(U *u) { // #1, load A vptr load u->a.vptr, !invariant "A::vptr" // #2, change union member to B ... store u->b.vptr, !invariant "B::vptr" // #3, change union member back to A ... f(u); // performs: store u->a.vptr, !invariant "A::vptr" // #4, load A vptr again, can be forwarded from #1 but not from #2 load u->a.vptr, !invariant "A::vptr" } However, I don't immediately see a way in which the C++ object model would require us to track multiple distinct groups of loads and stores, so if it isn't useful to be able to do that outside of C++ vptr / const member invariant tracking, I think we can remove it.> -Hal > > On 01/24/2017 01:39 PM, Piotr Padlewski via llvm-dev wrote: > > Hi, > I would really like to hear some feedback about this. > > Piotr > > 2017-01-20 17:07 GMT+01:00 Piotr Padlewski <piotr.padlewski at gmail.com>: > >> Hi all, >> >> I would like to propose a new way clang would decorate vtable loads in >> order to handle devirtualization better. >> >> I've added *llvm-dev* also, because this can start a discussion about >> changing invariant.group to just invariant. >> >> PDF version of this RFC can be found here: >> >> https://drive.google.com/file/d/0B72TmzNsY6Z8ZmpOUnB5dDZfSFU >> /view?usp=sharing >> >> Background: >> >> Initial old design: >> >> http://lists.llvm.org/pipermail/cfe-dev/2015-July/044227.html >> >> My talk from LLVM Dev Meeting >> >> http://llvm.org/devmtg/2016-11/#talk6 >> The problem >> >> Clang with -fstrict-vtable-pointers decorates vtable loads with metadata >> corresponding to mangled pointer type name like: >> >> void g(A& a){ a.foo(); } >> >> define void @_Z1gR1A(%struct.A* dereferenceable(8) %a) >> local_unnamed_addr #0 { entry: %0 = bitcast %struct.A* %a to void (% >> struct.A*)*** %vtable = load void (%struct.A*)**, void (%struct.A*)*** % >> 0, !invariant.group !7 %1 = load void (%struct.A*)*, void (%struct.A*)** >> %vtable tail call void %1(%struct.A* nonnull %a) ret void } !7 = !{! >> "_ZTS1A"} >> >> This works well if the pointer type doesn’t change, but when it does, >> devirtualization might not happen like here: >> >> struct A { A(); virtual void foo(); }; struct B : A{ B(); >> virtual void foo(); }; void g(A& a){ a.foo(); a.foo(); } void >> clobber(A&); void f() { B b; clobber(b); g(b); } >> >> The other problem is that when we combine 2 instructions with different >> invariant.group metadata, then we pick one of them, because for now we can >> have only single !invariant.group metadata. >> The solution >> >> I had some initial ideas how it can be solved, like >> >> 1. >> >> introducing multi invariant groups >> 2. >> >> having sub invariant groups - like inheritance, so we could figure >> out that one group is subgroup of another >> 3. >> >> decorating all loads with base pointer MD (doesn’t work with multiple >> inheritance) >> >> I consulted my ideas with Krzysztof Pszeniczny, and he proposed something >> much simpler: we can decorate every invariant.group md with empty metadata. >> >> This should work because the lifetime of the object is strictly defined >> by invariant.group.barrier. >> >> If this holds, we can start discussion about if it makes sense to keep >> invariant groups, and instead have just “invariant”, that would be >> equivalent to having invariant.group with the same metadata. >> >> Do you have some thoughts about this approach? I don’t have a >> mathematical proof, but I am confident that it should be valid. >> > _______________________________________________ > LLVM Developers mailing listllvm-dev at lists.llvm.orghttp://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev > > -- > Hal Finkel > Lead, Compiler Technology and Programming Languages > Leadership Computing Facility > Argonne National Laboratory > > > _______________________________________________ > cfe-dev mailing list > cfe-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170125/de4f87a0/attachment.html>
Piotr Padlewski via llvm-dev
2017-Jan-26 12:44 UTC
[llvm-dev] [cfe-dev] RFC: Emitting empty invariant group for vtable loads
2017-01-26 3:28 GMT+01:00 Richard Smith <richard at metafoo.co.uk>:> On 25 January 2017 at 15:03, Hal Finkel via cfe-dev < > cfe-dev at lists.llvm.org> wrote: > >> Hi Piotr, >> >> I think makes sense. Modulo bitcasts, the invariant is identified by a >> particular pointer SSA value. Given that you can't sensibly have two >> nonequivalent invariants associated with the same pointer SSA value >> simultaneously, there's no need to also identify the invariant with a >> metadata string as well. When we need a new "identifier" for the pointed-to >> value, we get one using invariant.group.barrier. >> > > As I recall, the original motivation for the identifier was to support > cases where the "invariant" region's value changes and then changes back > (remember that invariant.group does not imply the storage doesn't change, > just that a particular set of loads and stores witness the same value): > > struct A { void *vptr; /*...*/ }; > struct B { void *vptr; /*...*/ }; > union U { A a; B b; }; > > void f(U *u) { > // #1, load A vptr > load u->a.vptr, !invariant "A::vptr" > // #2, change union member to B ... > store u->b.vptr, !invariant "B::vptr" > // #3, change union member back to A ... > f(u); // performs: store u->a.vptr, !invariant "A::vptr" > // #4, load A vptr again, can be forwarded from #1 but not from #2 > load u->a.vptr, !invariant "A::vptr" > } > > Excellent point, I forgot that one don't have to call placement new inorder to emplace different type. I discussed it with Krzysztof and we belive the best way to solve it is to introduce barriers before every use of union if it contains any polymorphic class (recursively for each class in union). This might look very bad, but assuming we will be able skip barrier for optimizations not relying on !invariant.group, then it will not pessimize anything.> However, I don't immediately see a way in which the C++ object model would > require us to track multiple distinct groups of loads and stores, so if it > isn't useful to be able to do that outside of C++ vptr / const member > invariant tracking, I think we can remove it. >It should not make optimizations in LLVM any harder with group or without, so I will postpone the removal of groups for some time.> -Hal >> >> On 01/24/2017 01:39 PM, Piotr Padlewski via llvm-dev wrote: >> >> Hi, >> I would really like to hear some feedback about this. >> >> Piotr >> >> 2017-01-20 17:07 GMT+01:00 Piotr Padlewski <piotr.padlewski at gmail.com>: >> >>> Hi all, >>> >>> I would like to propose a new way clang would decorate vtable loads in >>> order to handle devirtualization better. >>> >>> I've added *llvm-dev* also, because this can start a discussion about >>> changing invariant.group to just invariant. >>> >>> PDF version of this RFC can be found here: >>> >>> https://drive.google.com/file/d/0B72TmzNsY6Z8ZmpOUnB5dDZfSFU >>> /view?usp=sharing >>> >>> Background: >>> >>> Initial old design: >>> >>> http://lists.llvm.org/pipermail/cfe-dev/2015-July/044227.html >>> >>> My talk from LLVM Dev Meeting >>> >>> http://llvm.org/devmtg/2016-11/#talk6 >>> The problem >>> >>> Clang with -fstrict-vtable-pointers decorates vtable loads with >>> metadata corresponding to mangled pointer type name like: >>> >>> void g(A& a){ a.foo(); } >>> >>> define void @_Z1gR1A(%struct.A* dereferenceable(8) %a) >>> local_unnamed_addr #0 { entry: %0 = bitcast %struct.A* %a to void (% >>> struct.A*)*** %vtable = load void (%struct.A*)**, void (%struct.A*)*** >>> %0, !invariant.group !7 %1 = load void (%struct.A*)*, void (%struct.A*)** >>> %vtable tail call void %1(%struct.A* nonnull %a) ret void } !7 = !{! >>> "_ZTS1A"} >>> >>> This works well if the pointer type doesn’t change, but when it does, >>> devirtualization might not happen like here: >>> >>> struct A { A(); virtual void foo(); }; struct B : A{ B(); >>> virtual void foo(); }; void g(A& a){ a.foo(); a.foo(); } void >>> clobber(A&); void f() { B b; clobber(b); g(b); } >>> >>> The other problem is that when we combine 2 instructions with different >>> invariant.group metadata, then we pick one of them, because for now we can >>> have only single !invariant.group metadata. >>> The solution >>> >>> I had some initial ideas how it can be solved, like >>> >>> 1. >>> >>> introducing multi invariant groups >>> 2. >>> >>> having sub invariant groups - like inheritance, so we could figure >>> out that one group is subgroup of another >>> 3. >>> >>> decorating all loads with base pointer MD (doesn’t work with >>> multiple inheritance) >>> >>> I consulted my ideas with Krzysztof Pszeniczny, and he proposed >>> something much simpler: we can decorate every invariant.group md with empty >>> metadata. >>> >>> This should work because the lifetime of the object is strictly defined >>> by invariant.group.barrier. >>> >>> If this holds, we can start discussion about if it makes sense to keep >>> invariant groups, and instead have just “invariant”, that would be >>> equivalent to having invariant.group with the same metadata. >>> >>> Do you have some thoughts about this approach? I don’t have a >>> mathematical proof, but I am confident that it should be valid. >>> >> _______________________________________________ >> LLVM Developers mailing listllvm-dev at lists.llvm.orghttp://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >> >> -- >> Hal Finkel >> Lead, Compiler Technology and Programming Languages >> Leadership Computing Facility >> Argonne National Laboratory >> >> >> _______________________________________________ >> cfe-dev mailing list >> cfe-dev at lists.llvm.org >> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev >> >> >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170126/305b3c31/attachment-0001.html>