Capture variables from successful combination of sub-matches

[Declan-ARM]

Hi there (cross-post from re2),

Hopefully not a silly question, but is there a way to capture the variables of all possible matches and perform checks with them?

For example, if I use the following made-up SPIR-V:

    %end = OpVariable %_ptr_Function_int Function
     %15 = OpLoad %int %n
     %16 = OpLoad %int %n
           OpStore %end %16

...the following checks will return false...

const std::string check = R"(
    CHECK:    [[END:%\w+]] = OpVariable
    CHECK: [[LOADED:%\w+]] = OpLoad %int %n
    CHECK:                   OpStore [[END]] [[LOADED]]
)";

...despite OpStore %end %16 being a possibility.

Is there a better syntactical way to capture all possible versions of what LOADED might be and return success if a match is found with one of these versions (instead of stopping after the first match)?

[dneto0]

sorry for the long inattention.

This is a really good question. I don't know how to achieve this with the current set of rules. I also don't know how to do it with upstream FileCheck.

The requirement is that a rule would have to be matched one-or-more times, like the + modifier in a regular expression. The way effcee operates, it builds a list of checks, and then 'resolves' each as they are matched. (There's extra complexity with CHECK-NOTs, but set that aside). Currently a check is fully consumed and forgotten once it is resolved once.

If I were to address this, I'd probably add -LAST variants for all the postive directives: CHECK-LAST certainly and maybe CHECK-NEXT-LAST CHECK-SAME-LAST, CHECK-DAG-LAST, CHECK-LABEL-LAST (this last one seems excessive and redundant) The semantics would be: resolve the check but remain available to be resolved again. The trickiness comes in the interaction with CHECK-NOT, described here: https://github.com/google/effcee/blob/e63a164aa0a40a04ddca2c18976819668b5a47a8/effcee/match.cc#L56

[dneto0]

But I'm not sure that would really solve your problem, because if you wrote patterns:

const std::string check = R"(
    CHECK:         [[END:%\w+]] = OpVariable
    CHECK-LAST: [[LOADED:%\w+]] = OpLoad %int %n
    CHECK:                        OpStore [[END]] [[LOADED]]
)";

then it would fail if presented with input:

    %end = OpVariable %_ptr_Function_int Function
     %15 = OpLoad %int %n
     %16 = OpLoad %int %n
           OpStore %end %15

[Declan-ARM]

You're a gentleman, David - thank you for responding, regardless.

It was a really interesting problem, at the time - from memory, I believe we managed to find a way to make things work(ish) with CHECK-DAG.

Perhaps, some kind of CHECK-MULTIPLE might be useful, in the future, where the LOADED variable would be something like a std::unordered_set<T> that allows for checks for the existence of at least one of the potential combinations.

Happy to close now, unless you want to keep open for future work.

[dneto0]

Thanks.

Let's keep this open.

Clearly we don't update this much (a.k.a. "it's stable and does its job"), but I'd like to keep the idea around.