[llvm-dev] Conflicting check prefix detection not working in update_llc_test_checks.py?

Wed Feb 3 11:31:32 PST 2021

IIUC, what would happen in the future change case is that you end up with
some (not all) functions with no asserts, is that correct? And so we'd want
a warning listing those functions?

On Wed, Feb 3, 2021 at 11:25 AM Craig Topper <craig.topper at gmail.com> wrote:

> Maybe. I'm not quite sure what you mean by "functions with no asm".
>
> X86 especially has a lot of tests with a half dozen or more RUN lines.
> There is a common prefix for all RUN lines or at least all AVX/AVX512 run
> lines and then successively more precise check prefixes grouping smaller
> and smaller subsets of RUN lines. The goal is to minimize the number of
> check lines in the file by grouping as much as possible. But allowing
> precise checks for individual functions that have different behavior on
> many targets.
>
> Since we can't have unused prefixes in the test files any more we might
> lose some of the finer grained check prefixes if multiple RUN lines produce
> the same results for all functions today. A future change may break that
> and require finer grained prefixes to be added. That's the important case
> to warn about since it requires the user to change the RUN lines.
>
> ~Craig
>
>
> On Tue, Feb 2, 2021 at 9:39 PM Mircea Trofin <mtrofin at google.com> wrote:
>
>> I see, so maybe there are some levels of warning to observe -
>> "everything", functions with no asm, prefixes with no use. And maybe the
>> default is the second?
>>
>> On Tue, Feb 2, 2021, 21:35 Craig Topper <craig.topper at gmail.com> wrote:
>>
>>> Hi Mircea,
>>>
>>> It looks like the script is now reporting warnings even when there is a
>>> set of prefixes to update the test for all functions. For example, if you
>>> add -check-prefixes=CHECK,RV32I to the first command in alu32.ll and
>>> -check-prefixes=CHECK,RV64I to the second you'll get this, but it looks
>>> like there's a valid solution for the test and no user intervention is
>>> required.
>>>
>>> WARNING: Function slti had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> WARNING: Function sltiu had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> WARNING: Function srli had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> WARNING: Function srai had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> WARNING: Function add had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> WARNING: Function sub had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> WARNING: Function sll had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> WARNING: Function slt had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> WARNING: Function sltu had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> WARNING: Function srl had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> WARNING: Function sra had conflicting output from different RUN lines
>>> for prefix CHECK
>>>
>>> ~Craig
>>>
>>>
>>> On Mon, Feb 1, 2021 at 3:19 PM Craig Topper <craig.topper at gmail.com>
>>> wrote:
>>>
>>>> Yes, that matches my expectations. Thanks!
>>>>
>>>> ~Craig
>>>>
>>>>
>>>> On Mon, Feb 1, 2021 at 3:05 PM Mircea Trofin <mtrofin at google.com>
>>>> wrote:
>>>>
>>>>> Indeed, we're now not output-ing the case where some functions have
>>>>> conflicting asm, just the case when all functions lose their asm.
>>>>>
>>>>> I have a fix ready; to confirm, for this example (i.e. taking all
>>>>> (both) the "--check-prefix"-es in alu32.ll), would this output match your
>>>>> expectations?
>>>>>
>>>>> WARNING: Function slti had conflicting output from different RUN lines
>>>>> for prefix CHECK
>>>>> WARNING: Function sltiu had conflicting output from different RUN
>>>>> lines for prefix CHECK
>>>>> WARNING: Function srli had conflicting output from different RUN lines
>>>>> for prefix CHECK
>>>>> WARNING: Function srai had conflicting output from different RUN lines
>>>>> for prefix CHECK
>>>>> WARNING: Function add had conflicting output from different RUN lines
>>>>> for prefix CHECK
>>>>> WARNING: Function sub had conflicting output from different RUN lines
>>>>> for prefix CHECK
>>>>> WARNING: Function sll had conflicting output from different RUN lines
>>>>> for prefix CHECK
>>>>> WARNING: Function slt had conflicting output from different RUN lines
>>>>> for prefix CHECK
>>>>> WARNING: Function sltu had conflicting output from different RUN lines
>>>>> for prefix CHECK
>>>>> WARNING: Function srl had conflicting output from different RUN lines
>>>>> for prefix CHECK
>>>>> WARNING: Function sra had conflicting output from different RUN lines
>>>>> for prefix CHECK
>>>>>
>>>>> On Mon, Feb 1, 2021 at 2:12 PM Mircea Trofin <mtrofin at google.com>
>>>>> wrote:
>>>>>
>>>>>> looking
>>>>>>
>>>>>> On Mon, Feb 1, 2021 at 2:11 PM Craig Topper <craig.topper at gmail.com>
>>>>>> wrote:
>>>>>>
>>>>>>> update_llc_test_checks.py seems to not be telling me about assembly
>>>>>>> that differs under the same prefix anymore.
>>>>>>>
>>>>>>> An easy way to see this is to just remove the --check-prefix from
>>>>>>> test/CodeGen/RISCV/alu32.ll and run the script. You'll get no error about
>>>>>>> conflicts. And if you look at the resulting file only some functions will
>>>>>>> have been updated to use CHECK as the prefix.
>>>>>>>
>>>>>>> Reverting some commits to update_llc_test_checks.py suggest this may
>>>>>>> have been broken by e2dc306b1ac71258e6ce40a66e778527f282c355 [utils] Fix
>>>>>>> UpdateTestChecks case where 2 runs differ for last label
>>>>>>>
>>>>>>> ~Craig
>>>>>>>
>>>>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210203/acce1d3e/attachment-0001.html>