Philip Reames via llvm-dev
2015-Sep-24 23:34 UTC
[llvm-dev] [PATCH] D12923: Add support for function attribute "notail"
On 09/24/2015 03:04 PM, Akira Hatanaka wrote:> On Thu, Sep 24, 2015 at 1:53 PM, Philip Reames > <listmail at philipreames.com <mailto:listmail at philipreames.com>> wrote: > > > > On 09/24/2015 01:47 PM, Akira Hatanaka wrote: >> On Wed, Sep 23, 2015 at 11:13 AM, Philip Reames >> <listmail at philipreames.com <mailto:listmail at philipreames.com>> wrote: >> >> >> >> On 09/23/2015 08:48 AM, Akira Hatanaka wrote: >>> On Tue, Sep 22, 2015 at 8:31 AM, Philip Reames >>> <listmail at philipreames.com >>> <mailto:listmail at philipreames.com>> wrote: >>> >>> To be clear, this is a debuging aid only? It's not >>> something required for correctness? I'm somewhat >>> bothered by that because it seems like it would be a >>> useful implementation tool for higher level languages. >>> >>> >>> It's not purely a debugging aid that helps when you are >>> using the debugger. There are projects (that are not >>> debuggers) that rely on not missing frames to produce >>> results that are useful. >> If it's not simply best effort, that constrains our choices. >>> >>> A couple of thoughts in no particular order: >>> 1) Can we always annotate the call site rather than the >>> function? That removes the unpredictability due to >>> optimization. >>> >>> >>> Annotating the call site should be fine. For the use cases >>> that we care about, it probably isn't important to prevent >>> tail calls on indirect calls. >> Given this, I would lean towards a notail value being added >> as an alternative to "tail" and "musttail". This seems to fit >> the existing uses, doesn't have any obvious loop holes or >> best effort semantics, and solves the problem at hand. >>> >>> 2) Calling it something like "no-direct-tail-call" or >>> "prefer-no-tail" would remove some of the confusion >>> value. When I see "notail", I expect that to always be >>> respected; the best effort semantics come as a bit of a >>> surprise. >>> >>> >>> I agree. A name that indicates it's only a best effort >>> option or it's an option that affects only direct calls >>> would be good. >> (This only applies if we're talking about a function >> annotation. The call site annotation applies to both direct >> and indirect calls.) >>> >>> >>> 3) This seems analogous to the "tail" marker in that it >>> indicates a preference/option. Whatever we end up with, >>> it needs to be a verifier option to have a "tail" or >>> "musttail" call site which is also "notail". It also >>> needs to be an error to have a mustail callsite to a >>> notail function (if such ends up existing.) >>> >>> >>> If we are going to annotate the function, I think we should >>> have the verifier catch incompatibilities between the >>> markers on the call sites and the function attribute on the >>> called functions. >>> >>> If we are annotating the call site, the verifier check isn't >>> needed since the tail-call related markers are enums that >>> are mutually exclusive. >> Yep. >>> >>> 4) It somewhat feels like there are two concepts being >>> intermixed here. 1) A call site which will never be a >>> tail call. 2) A function which we prefer not to tail >>> call to. Does it make sense to separate them? >>> >>> >>> Yes, it makes sense to separate them. For the use case we >>> care about, either 1) or 2) will do. We don't have to have >>> support for both. >> I would lean toward doing (1) for now. We can come back and >> implement (2) at a later time if we find it's needed. After >> (1), each call site will have four states: >> - "notail" - Can not be a tail call. >> - "" - May be a tail call if analysis finds it legal, >> profitable, and desirable* >> - "tail" - May be a tail call, profitability hinted >> - "musttail" - Must be a tail call. >> >> * (2) would basically just change the desirability of moving >> from "" to "tail". >> >> >> OK. I'm considering changing the direction of this patch and >> marking the call site instead of the called function. >> >> We should also discuss what kinds of source level attributes >> we'll need. My plan is to attach an attribute that indicates >> notail (something like no_direct_tail) to the called function >> declaration and definition and then mark all the direct call >> sites in the IR that call the function as notaill. In addition to >> that, it seems like we want to have a way to attach the attribute >> directly to the call site: >> >> void (*indirectCall)(int, int, int); >> >> void foo1(int a, int b) { >> (*indirectCall)(a, b, c) __attribute__((notail)); >> } > I think you're going to want to have the same dichotomy between > (1) and (2) at the source level as in the IR. The same issues > apply in both cases. > > > I think you were suggesting we always annotate the call site rather > than the function. Are you suggesting we do the same thing at the > source level, i.e. allow marking the call site but not the function? I > have to confirm, but I believe the people who requested this feature > wanted to use a function attribute rather than marking each call site > that calls the target function.What I'm suggesting is that the function attribute should be a best effort semantic. If we can tell something is a direct call, we should avoid the tail call, but we don't guarantee to never emit a tail call to the function in question. If you need an absolutely guarantee that a particular call (even if indirect) will not be a tail call, you'd need to mark the call site. (In practice, the "prefer_no_tail" function attribute will probably work pretty reliability, but I'm concerned about "promising" that it will work. Doing so for a subset of calls (i.e. "statically direct calls") and may work for others is somewhat defensible, but I've learned not to make promises the optimizer has a hard time keeping. Does the distinction make sense?)> >>> >>> Philip >>> >>> >>> On 09/21/2015 06:22 PM, Akira Hatanaka wrote: >>>> Several users have been asking for this function >>>> attribute to prevent losing the calling functions's >>>> information in the backtrace. If we attach the >>>> attribute to a function, ideally we would want to >>>> prevent tail call optimization on all call sites that >>>> call the function. However, the compiler cannot always >>>> tell which function is called from a call site if it's >>>> an indirect call, so it's fine if an indirect call to >>>> the marked function ends up being tail-call optimized. >>>> For direct calls, we want the function attribute to >>>> prevent tail call 100% of the time. >>>> >>>> We can also use a "notail" marker on the call >>>> instruction instead of using a function attribute. The >>>> only downside of using a marker is that we probably >>>> will never be able to prevent tail call optimization on >>>> indirect calls even when the compiler can turn it into >>>> a direct call (for example, via inlining). I'm not sure >>>> at the moment how important this is. >>>> >>>> On Thu, Sep 17, 2015 at 9:47 AM, Philip Reames via >>>> llvm-commits <llvm-commits at lists.llvm.org >>>> <mailto:llvm-commits at lists.llvm.org>> wrote: >>>> >>>> +llvm-dev >>>> >>>> Can you give a bit of background on what you're >>>> trying to address here? After reading through the >>>> discussion and seeing that this is a best effort >>>> flag, I'm not sure that a function attribute is the >>>> best way to describe this. I'm open to being >>>> convinced it is, but I'd like to hear a bit more >>>> about the use case and get broader visibility on >>>> the proposal first. >>>> >>>> Philip >>>> >>>> >>>> On 09/16/2015 07:27 PM, Akira Hatanaka via >>>> llvm-commits wrote: >>>>> ahatanak created this revision. >>>>> ahatanak added a subscriber: llvm-commits. >>>>> >>>>> This patch adds support for a new IR function attribute "notail". The attribute is used to disable tail call optimization on calls to functions marked with the attribute. >>>>> >>>>> This attribute is different from the existing attribute "disable-tail-calls", which disables tail call optimizations on all call sites within the marked function. >>>>> >>>>> The patch to add support for the corresponding source-level function attribute is here: >>>>> http://reviews.llvm.org/D12922 >>>>> >>>>> http://reviews.llvm.org/D12923 >>>>> >>>>> Files: >>>>> docs/LangRef.rst >>>>> include/llvm/Bitcode/LLVMBitCodes.h >>>>> include/llvm/IR/Attributes.h >>>>> include/llvm/IR/Instructions.h >>>>> lib/AsmParser/LLLexer.cpp >>>>> lib/AsmParser/LLParser.cpp >>>>> lib/AsmParser/LLToken.h >>>>> lib/Bitcode/Reader/BitcodeReader.cpp >>>>> lib/Bitcode/Writer/BitcodeWriter.cpp >>>>> lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp >>>>> lib/IR/Attributes.cpp >>>>> lib/IR/Verifier.cpp >>>>> lib/Transforms/Scalar/TailRecursionElimination.cpp >>>>> test/Bindings/llvm-c/Inputs/invalid.ll.bc >>>>> test/Bindings/llvm-c/invalid-bitcode.test >>>>> test/Bitcode/attributes.ll >>>>> test/Bitcode/invalid.ll >>>>> test/Bitcode/invalid.ll.bc >>>>> test/CodeGen/X86/attr-notail.ll >>>>> test/Transforms/TailCallElim/notail.ll >>>>> >>>>> >>>>> >>>>> _______________________________________________ >>>>> llvm-commits mailing list >>>>> llvm-commits at lists.llvm.org >>>>> <mailto:llvm-commits at lists.llvm.org> >>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits >>>> >>>> >>>> _______________________________________________ >>>> llvm-commits mailing list >>>> llvm-commits at lists.llvm.org >>>> <mailto:llvm-commits at lists.llvm.org> >>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits >>>> >>>> >>> >>> >> >> > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150924/e3eea620/attachment.html>
Akira Hatanaka via llvm-dev
2015-Sep-29 00:49 UTC
[llvm-dev] [PATCH] D12923: Add support for function attribute "notail"
On Thu, Sep 24, 2015 at 4:34 PM, Philip Reames <listmail at philipreames.com> wrote:> > > On 09/24/2015 03:04 PM, Akira Hatanaka wrote: > > On Thu, Sep 24, 2015 at 1:53 PM, Philip Reames <listmail at philipreames.com> > wrote: > >> >> >> On 09/24/2015 01:47 PM, Akira Hatanaka wrote: >> >> On Wed, Sep 23, 2015 at 11:13 AM, Philip Reames < >> <listmail at philipreames.com>listmail at philipreames.com> wrote: >> >>> >>> >>> On 09/23/2015 08:48 AM, Akira Hatanaka wrote: >>> >>> On Tue, Sep 22, 2015 at 8:31 AM, Philip Reames < >>> <listmail at philipreames.com>listmail at philipreames.com> wrote: >>> >>>> To be clear, this is a debuging aid only? It's not something required >>>> for correctness? I'm somewhat bothered by that because it seems like it >>>> would be a useful implementation tool for higher level languages. >>>> >>>> >>> It's not purely a debugging aid that helps when you are using the >>> debugger. There are projects (that are not debuggers) that rely on not >>> missing frames to produce results that are useful. >>> >>> If it's not simply best effort, that constrains our choices. >>> >>> >>> >>>> A couple of thoughts in no particular order: >>>> 1) Can we always annotate the call site rather than the function? That >>>> removes the unpredictability due to optimization. >>>> >>> >>> Annotating the call site should be fine. For the use cases that we care >>> about, it probably isn't important to prevent tail calls on indirect calls. >>> >>> Given this, I would lean towards a notail value being added as an >>> alternative to "tail" and "musttail". This seems to fit the existing uses, >>> doesn't have any obvious loop holes or best effort semantics, and solves >>> the problem at hand. >>> >>> >>> >>>> 2) Calling it something like "no-direct-tail-call" or "prefer-no-tail" >>>> would remove some of the confusion value. When I see "notail", I expect >>>> that to always be respected; the best effort semantics come as a bit of a >>>> surprise. >>>> >>> >>> I agree. A name that indicates it's only a best effort option or it's an >>> option that affects only direct calls would be good. >>> >>> (This only applies if we're talking about a function annotation. The >>> call site annotation applies to both direct and indirect calls.) >>> >>> >>> >>>> >>>> 3) This seems analogous to the "tail" marker in that it indicates a >>>> preference/option. Whatever we end up with, it needs to be a verifier >>>> option to have a "tail" or "musttail" call site which is also "notail". It >>>> also needs to be an error to have a mustail callsite to a notail function >>>> (if such ends up existing.) >>>> >>> >>> If we are going to annotate the function, I think we should have the >>> verifier catch incompatibilities between the markers on the call sites and >>> the function attribute on the called functions. >>> >>> If we are annotating the call site, the verifier check isn't needed >>> since the tail-call related markers are enums that are mutually exclusive. >>> >>> Yep. >>> >>> >>> >>>> 4) It somewhat feels like there are two concepts being intermixed >>>> here. 1) A call site which will never be a tail call. 2) A function which >>>> we prefer not to tail call to. Does it make sense to separate them? >>>> >>>> >>> Yes, it makes sense to separate them. For the use case we care about, >>> either 1) or 2) will do. We don't have to have support for both. >>> >>> I would lean toward doing (1) for now. We can come back and implement >>> (2) at a later time if we find it's needed. After (1), each call site will >>> have four states: >>> - "notail" - Can not be a tail call. >>> - "" - May be a tail call if analysis finds it legal, profitable, and >>> desirable* >>> - "tail" - May be a tail call, profitability hinted >>> - "musttail" - Must be a tail call. >>> >>> * (2) would basically just change the desirability of moving from "" to >>> "tail". >>> >>> >> OK. I'm considering changing the direction of this patch and marking the >> call site instead of the called function. >> >> We should also discuss what kinds of source level attributes we'll need. >> My plan is to attach an attribute that indicates notail (something like >> no_direct_tail) to the called function declaration and definition and then >> mark all the direct call sites in the IR that call the function as notaill. >> In addition to that, it seems like we want to have a way to attach the >> attribute directly to the call site: >> >> void (*indirectCall)(int, int, int); >> >> void foo1(int a, int b) { >> (*indirectCall)(a, b, c) __attribute__((notail)); >> } >> >> I think you're going to want to have the same dichotomy between (1) and >> (2) at the source level as in the IR. The same issues apply in both >> cases. >> > > I think you were suggesting we always annotate the call site rather than > the function. Are you suggesting we do the same thing at the source level, > i.e. allow marking the call site but not the function? I have to confirm, > but I believe the people who requested this feature wanted to use a > function attribute rather than marking each call site that calls the target > function. > > What I'm suggesting is that the function attribute should be a best effort > semantic. If we can tell something is a direct call, we should avoid the > tail call, but we don't guarantee to never emit a tail call to the function > in question. If you need an absolutely guarantee that a particular call > (even if indirect) will not be a tail call, you'd need to mark the call > site. > > (In practice, the "prefer_no_tail" function attribute will probably work > pretty reliability, but I'm concerned about "promising" that it will work. > Doing so for a subset of calls (i.e. "statically direct calls") and may > work for others is somewhat defensible, but I've learned not to make > promises the optimizer has a hard time keeping. Does the distinction make > sense?) > >That was what I had in mind: the function attribute blocks tail call for statically direct calls but doesn't promise anything (in fact, does nothing) for indirect calls. Do you think we shouldn't make any promises for statically direct calls either? I don't see why it's hard to keep the promise that direct tail calls will be blocked. Do you have a particular optimization in mind that would be difficult or impossible to implement if we promised to block direct calls?> Philip >>>> >>>> >>>> On 09/21/2015 06:22 PM, Akira Hatanaka wrote: >>>> >>>> Several users have been asking for this function attribute to prevent >>>> losing the calling functions's information in the backtrace. If we attach >>>> the attribute to a function, ideally we would want to prevent tail call >>>> optimization on all call sites that call the function. However, the >>>> compiler cannot always tell which function is called from a call site if >>>> it's an indirect call, so it's fine if an indirect call to the marked >>>> function ends up being tail-call optimized. For direct calls, we want the >>>> function attribute to prevent tail call 100% of the time. >>>> >>>> We can also use a "notail" marker on the call instruction instead of >>>> using a function attribute. The only downside of using a marker is that we >>>> probably will never be able to prevent tail call optimization on indirect >>>> calls even when the compiler can turn it into a direct call (for example, >>>> via inlining). I'm not sure at the moment how important this is. >>>> >>>> On Thu, Sep 17, 2015 at 9:47 AM, Philip Reames via llvm-commits < >>>> <llvm-commits at lists.llvm.org>llvm-commits at lists.llvm.org> wrote: >>>> >>>>> +llvm-dev >>>>> >>>>> Can you give a bit of background on what you're trying to address >>>>> here? After reading through the discussion and seeing that this is a best >>>>> effort flag, I'm not sure that a function attribute is the best way to >>>>> describe this. I'm open to being convinced it is, but I'd like to hear a >>>>> bit more about the use case and get broader visibility on the proposal >>>>> first. >>>>> >>>>> Philip >>>>> >>>>> >>>>> On 09/16/2015 07:27 PM, Akira Hatanaka via llvm-commits wrote: >>>>> >>>>> ahatanak created this revision. >>>>> ahatanak added a subscriber: llvm-commits. >>>>> >>>>> This patch adds support for a new IR function attribute "notail". The attribute is used to disable tail call optimization on calls to functions marked with the attribute. >>>>> >>>>> This attribute is different from the existing attribute "disable-tail-calls", which disables tail call optimizations on all call sites within the marked function. >>>>> >>>>> The patch to add support for the corresponding source-level function attribute is here:http://reviews.llvm.org/D12922 >>>>> http://reviews.llvm.org/D12923 >>>>> >>>>> Files: >>>>> docs/LangRef.rst >>>>> include/llvm/Bitcode/LLVMBitCodes.h >>>>> include/llvm/IR/Attributes.h >>>>> include/llvm/IR/Instructions.h >>>>> lib/AsmParser/LLLexer.cpp >>>>> lib/AsmParser/LLParser.cpp >>>>> lib/AsmParser/LLToken.h >>>>> lib/Bitcode/Reader/BitcodeReader.cpp >>>>> lib/Bitcode/Writer/BitcodeWriter.cpp >>>>> lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp >>>>> lib/IR/Attributes.cpp >>>>> lib/IR/Verifier.cpp >>>>> lib/Transforms/Scalar/TailRecursionElimination.cpp >>>>> test/Bindings/llvm-c/Inputs/invalid.ll.bc >>>>> test/Bindings/llvm-c/invalid-bitcode.test >>>>> test/Bitcode/attributes.ll >>>>> test/Bitcode/invalid.ll >>>>> test/Bitcode/invalid.ll.bc >>>>> test/CodeGen/X86/attr-notail.ll >>>>> test/Transforms/TailCallElim/notail.ll >>>>> >>>>> >>>>> >>>>> >>>>> _______________________________________________ >>>>> llvm-commits mailing listllvm-commits at lists.llvm.orghttp://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits >>>>> >>>>> >>>>> >>>>> _______________________________________________ >>>>> llvm-commits mailing list >>>>> <llvm-commits at lists.llvm.org>llvm-commits at lists.llvm.org >>>>> <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits> >>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits >>>>> >>>>> >>>> >>>> >>> >>> >> >> > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150928/e8553bb1/attachment-0001.html>
Sanjoy Das via llvm-dev
2015-Sep-29 05:38 UTC
[llvm-dev] [PATCH] D12923: Add support for function attribute "notail"
> That was what I had in mind: the function attribute blocks tail call for statically direct calls but doesn't promise> anything (in fact, does nothing) for indirect calls. > > Do you think we shouldn't make any promises for statically direct calls either? I don't see why it's hard to keep the > promise that direct tail calls will be blocked. Do you have a particular optimization in mind that would be difficult or > impossible to implement if we promised to block direct calls? I think in this scheme we'll have problems around devirtualization. For instance, you could start with a indirect call to a notail target that would get TCO'ed (as an indirect call), after which devirtualization or load/store motion would turn the indirect call into a direct call; and you'd end up with a direct tail call to a notail target. I don't know what your requirements are, but if I were you I'd design the `notail` marker to be logically part of a function's calling convention. That way functions marked `notail` get to do things that are legal only if the call to them wasn't a tail call, because all well-defined calls constrain the caller to match callee's expectation with respect to tail calls. In the CFE, you could make the `notail` attribute have the same restrictions / conventions as things like `fastcall`. -- Sanjoy