[llvm-dev] Inlining with different target features
Thomas Lively via llvm-dev
llvm-dev at lists.llvm.org
Mon Aug 31 13:23:38 PDT 2020
Thanks for your reply! Using module metadata is an interesting idea, but it
would require frontends to make wasm-specific changes to how they handle
target features, which would be unfortunate. Working around that by
extracting target features into metadata somewhere seems like it would be
at least as intrusive as having the inliner update target features. We
would also lose out on other valuable utilities like the ability to limit
target intrinsics to certain features.
On Mon, Aug 31, 2020 at 9:52 AM David Blaikie <dblaikie at gmail.com> wrote:
> +echristo for thoughts on subtarget features and function inlining.
> Maybe target features aren't the right tool to model the WebAssembly
> situation? Perhaps you could model those with mergeable module-level
> metadata instead? Then the module would always have all the features and
> that would match the "we're allowed to union all features across all
> functions anyway" without it being a delayed pass that happens later.
> On Mon, Aug 17, 2020 at 1:49 PM Thomas Lively via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>> Hi llvm-dev,
>> I recently updated the WebAssembly TargetTransformInfo to allow functions
>> with different target feature sets to be inlined into each other, but I ran
>> into an issue I want to get the community's opinion on.
>> Since WebAssembly modules have to be validated before they are run, it
>> only makes sense to talk about WebAssembly features at module granularity
>> rather than function granularity. The WebAssembly backend even runs a pass
>> that applies the union of all used features to each function. That means
>> that ideally inlining for the WebAssembly target would be able to disregard
>> features entirely, since they will all be the same in the end.
>> However, right now I have to be more conservative than that and only
>> allow a callee to be inlined into a caller if the callee has a subset of
>> the caller's features. Otherwise, a target intrinsic might end up being
>> used in a function that does not have the necessary target features
>> enabled, which would cause a validation failure.
>> The best solution I can think of for this problem would be to allow
>> targets to opt-in to having a caller's feature set updated to include the
>> callee's feature set when the callee is inlined into the caller. This could
>> be implemented via a new TTI hook, but a more general solution might be to
>> change the return type of `areInlineCompatible` to allow targets to control
>> this behavior on a case-by-case basis. Does this general direction sound
>> ok, and if so, would it be better to add a new hook or add functionality to
>> the existing one?
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev