Florian Hahn via llvm-dev
2021-Mar-27 18:30 UTC
[llvm-dev] [RFC] Adding range metadata to array subscripts.
> On Mar 24, 2021, at 19:32, Johannes Doerfert <johannesdoerfert at gmail.com> wrote: > > > On 3/24/21 12:47 PM, Florian Hahn wrote: >> >>> On Mar 24, 2021, at 15:16, Johannes Doerfert via llvm-dev <llvm-dev at lists.llvm.org> wrote: >>> >>> >>> On 3/24/21 9:06 AM, Clement Courbet wrote: >>>> On Wed, Mar 24, 2021 at 2:20 PM Johannes Doerfert < >>>> johannesdoerfert at gmail.com> wrote: >>>> >>>>> I really like encoding more (range) information in the IR, >>>>> more thoughts inlined. >>>>> >>>>> On 3/24/21 4:14 AM, Clement Courbet via llvm-dev wrote: >>>>>> Hi everyone, >>>>>> >>>>>> tl;dr: I would like to teach clang to output range metadata so that LLVM >>>>>> can do better alias analysis. I have a proposal as D99248 >>>>>> <https://reviews.llvm.org/D99248> (clang part) and D99247 >>>>>> <https://reviews.llvm.org/D99247> (llvm part). But there are other >>>>> possible >>>>>> options that I'm detailing below. >>>>>> >>>>>> Consider the following code, adapted from brotli >>>>>> <https://en.wikipedia.org/wiki/Brotli>: >>>>>> >>>>>> ``` >>>>>> >>>>>> struct Histogram { >>>>>> >>>>>> int values[256]; >>>>>> >>>>>> int total; >>>>>> >>>>>> }; >>>>>> >>>>>> Histogram DoIt(const int* image, int size) { >>>>>> >>>>>> Histogram histogram; >>>>>> >>>>>> for (int i = 0; i < size; ++i) { >>>>>> >>>>>> ++histogram.values[image[i]]; // (A) >>>>>> >>>>>> ++histogram.total; // (B) >>>>>> >>>>>> } >>>>>> >>>>>> return histogram; >>>>>> >>>>>> } >>>>>> ``` >>>>>> >>>>>> In this code, the compiler does not know anything about the values of >>>>>> images[i], so it assumes that 256 is a possible value for it. In that >>>>> case, >>>>>> (A) would change the value of histogram.total, so (B) has to load, add >>>>> one >>>>>> and store [godbolt <https://godbolt.org/z/KxE343>]. >>>>>> >>>>>> Fortunately, C/C++ has a rule that it is invalid (actually, UB) to use >>>>>> values to form a pointer to total and dereference it. What valid C/C++ >>>>> code >>>>>> is allowed to do with values is: >>>>>> - Form any pointer in [values, values + 256]. >>>>>> - Form and dereference any pointer in [values, values + 256) >>>>>> >>>>>> Note that the LLVM memory model is much laxer than that of C/C++. It has >>>>> no >>>>>> notion of types. In particular, given an LLVM aggregate definition: >>>>>> >>>>>> ``` >>>>>> %struct.S = type { [42 x i32], i32, i32 } >>>>>> ``` >>>>>> >>>>>> It is perfectly valid to use an address derived from a GEP(0,0,%i) [gep >>>>>> reference] representing indexing into the [42 x i32] array to load the >>>>> i32 >>>>>> member at index 2. It is also valid for %i to be 43 (though not 44 if an >>>>>> inbound GEP is used). >>>>>> So clang has to give LLVM more information about the C/C++ rules. >>>>>> >>>>>> *IR representation:* >>>>>> LLVM has several ways of representing ranges of values: >>>>>> - *!range* metadata can be attached to integer call and load >>>>> instructions >>>>>> to indicate the allowed range of values of the result. LLVM's >>>>> ValueTracking >>>>>> provides a function for querying the range for any llvm::Variable. >>>>>> - The *llvm.assume* intrinsic takes a boolean condition that can also >>>>> be >>>>>> used by ValueTracking to infer range of values. >>>>>> - The *inrange* attribute of GEP can be used to indicate C-like >>>>> semantics >>>>>> for the structure field marked with the inrange attribute. It can only be >>>>>> used for GEP constantexprs (ie.e. GEPs defined inline), but not for >>>>>> standalone GEPs defining instructions. relevant discussion >>>>>> <https://reviews.llvm.org/D22793?id=65626#inline-194653>. >>>>>> >>>>>> Alternatives: >>>>>> *(1) *Annotate each array subscript index value with a range, e.g.: >>>>>> ``` >>>>>> %i = i64 … >>>>>> %ri = call i64 @llvm.annotation.i64(%index), !range !0 >>>>>> %gep1 = getelementptr inbounds %struct.S, %struct.S* %s, i64 0, i32 0, >>>>> i32 >>>>>> %ri >>>>>> ... >>>>>> !0 = !{i64 0, i64 42} >>>>>> ``` >>>>>> *(2) *(variant of 1) relax the constraint that !range metadata can only >>>>> be >>>>>> set on call and load instructions, and set the !range metadata on the >>>>> index >>>>>> expression. We still need annotations for function parameters though: >>>>>> ``` >>>>>> %i = i64 … , !range !0 >>>>>> %gep1 = getelementptr inbounds %struct.S, %struct.S* %s, i64 0, i32 0, >>>>> i32 >>>>>> %i >>>>>> ... >>>>>> !0 = !{i64 0, i64 42} >>>>>> ``` >>>>>> This is slightly more compact. >>>>>> >>>>>> *(3)* Same as (1), with llvm.assume. This feels inferior to annotations. >>>>>> *(4)* Extend inrange to non-constantexprs GEPs. It is unclear how this >>>>> will >>>>>> interfere with optimizations. >>>>> I would very much like not to introduce another way to encode >>>>> assumptions other than `llvm.assume`. If you want to avoid the extra >>>>> instructions, use `llvm.assume(i1 true) ["range"(%val, %lb, %ub)]`, >>>>> which is in line with our move towards operand bundle use. >>>>> >>>> Thanks, I did not know about that. I've just tried it but it appears that >>>> tags have to be attribute names, and `!range` is not a valid attribute, >>>> it's a metadata node. Is there a way to encode this ? >>> We need to get rid of that assertion. There are other non-attributes >>> to be used in assume operand bundles in the (near) future, so the this >>> work has to be done anyway. >> >> +1 on trying to use assume, rather than adding another way. >> >> But are value ranges special for assumes, so that we need to handle them in a bundle? Is that just so we can easier skip ‘artificial’ assume users? > > It would make users explicit and we will have non-attribute bundles anyway. > I find it also "conceptually nicer", would you prefer explicit instructions?One disadvantage of using a bundle (or !range metadata) is that we treat ranges for certain values in a special way and differently to how we treat range information expressed by the user e.g. via conditions (or builtin assume). This means we have to handle multiple variants across the codebase, which can lead to situations where only one or the other is handled, which in turn can lead to surprising results (of the form: why does a transformation apply if information provided in a certain way, but does not apply of the equivalent info is provided in a different way). Using instruction potentially also allows us to specify more complex ranges, in relation to other values. But I realize that there are some practical consideration that make the instruction approach less appealing and I am all in favor of the more pragmatic & practical solution to start with. Cheers, Florian -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210327/ec6db754/attachment.html>
Johannes Doerfert via llvm-dev
2021-Mar-27 20:37 UTC
[llvm-dev] [RFC] Adding range metadata to array subscripts.
On 3/27/21 1:30 PM, Florian Hahn wrote:> >> On Mar 24, 2021, at 19:32, Johannes Doerfert <johannesdoerfert at gmail.com> wrote: >> >> >> On 3/24/21 12:47 PM, Florian Hahn wrote: >>>> On Mar 24, 2021, at 15:16, Johannes Doerfert via llvm-dev <llvm-dev at lists.llvm.org> wrote: >>>> >>>> >>>> On 3/24/21 9:06 AM, Clement Courbet wrote: >>>>> On Wed, Mar 24, 2021 at 2:20 PM Johannes Doerfert < >>>>> johannesdoerfert at gmail.com> wrote: >>>>> >>>>>> I really like encoding more (range) information in the IR, >>>>>> more thoughts inlined. >>>>>> >>>>>> On 3/24/21 4:14 AM, Clement Courbet via llvm-dev wrote: >>>>>>> Hi everyone, >>>>>>> >>>>>>> tl;dr: I would like to teach clang to output range metadata so that LLVM >>>>>>> can do better alias analysis. I have a proposal as D99248 >>>>>>> <https://reviews.llvm.org/D99248> (clang part) and D99247 >>>>>>> <https://reviews.llvm.org/D99247> (llvm part). But there are other >>>>>> possible >>>>>>> options that I'm detailing below. >>>>>>> >>>>>>> Consider the following code, adapted from brotli >>>>>>> <https://en.wikipedia.org/wiki/Brotli>: >>>>>>> >>>>>>> ``` >>>>>>> >>>>>>> struct Histogram { >>>>>>> >>>>>>> int values[256]; >>>>>>> >>>>>>> int total; >>>>>>> >>>>>>> }; >>>>>>> >>>>>>> Histogram DoIt(const int* image, int size) { >>>>>>> >>>>>>> Histogram histogram; >>>>>>> >>>>>>> for (int i = 0; i < size; ++i) { >>>>>>> >>>>>>> ++histogram.values[image[i]]; // (A) >>>>>>> >>>>>>> ++histogram.total; // (B) >>>>>>> >>>>>>> } >>>>>>> >>>>>>> return histogram; >>>>>>> >>>>>>> } >>>>>>> ``` >>>>>>> >>>>>>> In this code, the compiler does not know anything about the values of >>>>>>> images[i], so it assumes that 256 is a possible value for it. In that >>>>>> case, >>>>>>> (A) would change the value of histogram.total, so (B) has to load, add >>>>>> one >>>>>>> and store [godbolt <https://godbolt.org/z/KxE343>]. >>>>>>> >>>>>>> Fortunately, C/C++ has a rule that it is invalid (actually, UB) to use >>>>>>> values to form a pointer to total and dereference it. What valid C/C++ >>>>>> code >>>>>>> is allowed to do with values is: >>>>>>> - Form any pointer in [values, values + 256]. >>>>>>> - Form and dereference any pointer in [values, values + 256) >>>>>>> >>>>>>> Note that the LLVM memory model is much laxer than that of C/C++. It has >>>>>> no >>>>>>> notion of types. In particular, given an LLVM aggregate definition: >>>>>>> >>>>>>> ``` >>>>>>> %struct.S = type { [42 x i32], i32, i32 } >>>>>>> ``` >>>>>>> >>>>>>> It is perfectly valid to use an address derived from a GEP(0,0,%i) [gep >>>>>>> reference] representing indexing into the [42 x i32] array to load the >>>>>> i32 >>>>>>> member at index 2. It is also valid for %i to be 43 (though not 44 if an >>>>>>> inbound GEP is used). >>>>>>> So clang has to give LLVM more information about the C/C++ rules. >>>>>>> >>>>>>> *IR representation:* >>>>>>> LLVM has several ways of representing ranges of values: >>>>>>> - *!range* metadata can be attached to integer call and load >>>>>> instructions >>>>>>> to indicate the allowed range of values of the result. LLVM's >>>>>> ValueTracking >>>>>>> provides a function for querying the range for any llvm::Variable. >>>>>>> - The *llvm.assume* intrinsic takes a boolean condition that can also >>>>>> be >>>>>>> used by ValueTracking to infer range of values. >>>>>>> - The *inrange* attribute of GEP can be used to indicate C-like >>>>>> semantics >>>>>>> for the structure field marked with the inrange attribute. It can only be >>>>>>> used for GEP constantexprs (ie.e. GEPs defined inline), but not for >>>>>>> standalone GEPs defining instructions. relevant discussion >>>>>>> <https://reviews.llvm.org/D22793?id=65626#inline-194653>. >>>>>>> >>>>>>> Alternatives: >>>>>>> *(1) *Annotate each array subscript index value with a range, e.g.: >>>>>>> ``` >>>>>>> %i = i64 … >>>>>>> %ri = call i64 @llvm.annotation.i64(%index), !range !0 >>>>>>> %gep1 = getelementptr inbounds %struct.S, %struct.S* %s, i64 0, i32 0, >>>>>> i32 >>>>>>> %ri >>>>>>> ... >>>>>>> !0 = !{i64 0, i64 42} >>>>>>> ``` >>>>>>> *(2) *(variant of 1) relax the constraint that !range metadata can only >>>>>> be >>>>>>> set on call and load instructions, and set the !range metadata on the >>>>>> index >>>>>>> expression. We still need annotations for function parameters though: >>>>>>> ``` >>>>>>> %i = i64 … , !range !0 >>>>>>> %gep1 = getelementptr inbounds %struct.S, %struct.S* %s, i64 0, i32 0, >>>>>> i32 >>>>>>> %i >>>>>>> ... >>>>>>> !0 = !{i64 0, i64 42} >>>>>>> ``` >>>>>>> This is slightly more compact. >>>>>>> >>>>>>> *(3)* Same as (1), with llvm.assume. This feels inferior to annotations. >>>>>>> *(4)* Extend inrange to non-constantexprs GEPs. It is unclear how this >>>>>> will >>>>>>> interfere with optimizations. >>>>>> I would very much like not to introduce another way to encode >>>>>> assumptions other than `llvm.assume`. If you want to avoid the extra >>>>>> instructions, use `llvm.assume(i1 true) ["range"(%val, %lb, %ub)]`, >>>>>> which is in line with our move towards operand bundle use. >>>>>> >>>>> Thanks, I did not know about that. I've just tried it but it appears that >>>>> tags have to be attribute names, and `!range` is not a valid attribute, >>>>> it's a metadata node. Is there a way to encode this ? >>>> We need to get rid of that assertion. There are other non-attributes >>>> to be used in assume operand bundles in the (near) future, so the this >>>> work has to be done anyway. >>> +1 on trying to use assume, rather than adding another way. >>> >>> But are value ranges special for assumes, so that we need to handle them in a bundle? Is that just so we can easier skip ‘artificial’ assume users? >> It would make users explicit and we will have non-attribute bundles anyway. >> I find it also "conceptually nicer", would you prefer explicit instructions? > One disadvantage of using a bundle (or !range metadata) is that we treat ranges for certain values in a special way and differently to how we treat range information expressed by the user e.g. via conditions (or builtin assume).I don't think this is necessarily accurate. We can, and already do (https://godbolt.org/z/MaMEb1Koo), generate bundles from conditions. If we can interpret a condition, why could we not rewrite it into a bundle? I'm not sure why this is any different form other normalization we do. (And bundles have benefits over implicit instruction encodings, for example use tracking and #instructions.)> This means we have to handle multiple variants across the codebase, which can lead to situations where only one or the other is handled, which in turn can lead to surprising results (of the form: why does a transformation apply if information provided in a certain way, but does not apply of the equivalent info is provided in a different way). > Using instruction potentially also allows us to specify more complex ranges, in relation to other values.I don't see how bundles would restrict us in any way. I mean, if we want to express property XYZ for %v and %q, `llvm.assume(i1 true) ["XYZ"(%v, %q)]`, makes it really easy and it is arguably as generic as you want it to be.> But I realize that there are some practical consideration that make the instruction approach less appealing and I am all in favor of the more pragmatic & practical solution to start with.I think bundles, and more generic assumptions, are what we need in the future. I still believe we should use them to encode information in assertions [0], among other things, without running into the risk of having side-effects that influence the compilation. ~ Johannes [0] https://lists.llvm.org/pipermail/llvm-dev/2019-December/137632.html> Cheers, > Florian