[FileCheck] Error for unused check prefixes (was: [RFC] FileCheck: (dis)allowing unused prefixes)

Project Infrastructure LLVM Dev List Archives
#1

Last October Mircea sent an RFC about (dis)allowing unused check prefixes in FileCheck: https://lists.llvm.org/pipermail/llvm-dev/2020-October/146162.html
(In short, if FileCheck --check-prefix or --check-prefixes specifies a prefix which does not occur in the test, FileCheck will error. Note: a -NOT pattern is also counted as an occurrence.)

Mircea created a worksheet https://docs.google.com/spreadsheets/d/1o6q3XH1n3DDyyccnYZ_kVfhFbTDzC_S09e973_cwYuw/edit#gid=0 which may be a bit stale now but llvm-project has reached a state where all issues have been identified and fixed, or worked around (by opting in FileCheck --allow-unused-prefixes in some test directories, such as: clang/test/OpenMP, llvm/test/Transforms/Attributor and llvm/test/FileCheck).

We can make a switch if the community thinks that not allowing unused prefixes is the better default: https://reviews.llvm.org/D95849
For downstream projects using FileCheck and lit, it should be easy to restore the previous permissive behavior

from lit.llvm.subst import ToolSubst

fc = ToolSubst('FileCheck', unresolved='fatal')
config.substitutions.insert(0, (fc.regex, 'FileCheck --allow-unused-prefixes'))

# Note: if multiple --allow-unused-prefixes options are specified, the last wins.

Personally I am strongly in favor of disallowing unused check prefixes by default. We have identified numerous issues:

(1) typo. A misspelled check prefix does not test what it intends to test.
(2) bitrot tests due to refactoring
(3) unspecified -NOT patterns. Sometimes a test uses something like --check-prefixes=COMMON,TRUE and --check-prefixes=COMMON,FALSE to test both behaviors but they forget to include a FALSE-NO: pattern to test that some string does not appear.

(1) and (2) are especially common. There are indeed tests where --allow-unused-prefixes is more suitable - but they are sufficiently few that I think the default should be --allow-unused-prefixes=false.

So, what do folks think?

#2

In my opinion, the cost of a bit of inconvenience for users who want to allow unused prefixes is far outweighed by the potential to catch bugs (both test bugs and more importantly genuine bugs hidden by test bugs), so +1 to the default being false for the option.

#3

Agreed, in particular this half of the catching-typos problem is definitely worth solving.

(The other half would be typos in the CHECK directives themselves, which is a distinctly harder problem and one that FileCheck cannot solve on its own.)

–paulr

#4

Just my 2c, but I think we should allow unused prefixes. This does catch the occasional typo, but also has a cost: Historically, certain kinds of tests simply used a certain boilerplate of check lines, because differences are common, even if they don’t occur for each test. For X86 vector tests, it makes more sense to simply always include AVX1 and AVX2 test prefixes, even if it so happens that for this particular test, codegen is identical and only the AVX prefix ends up being used. This means that whenever codegen changes in a minor way (e.g. due to a target-independent SimplyDemandedBits change that has no direct relation to X86) and a difference is introduced, you need to now figure out which new prefixes you have to add. Or drop prefixes if a codegen difference goes away. Having to manually adjust check prefixes takes away from the usual experience of “Just run update_(llc_)test_checks”.

At least I personally have found the gradual migration towards disallowing unused prefixes to be more annoying than useful. I guess ergonomics could be improved if update_test_checks automatically dropped unused prefixes, but there’s really no way to automatically add prefixes, without domain-specific knowledge.

Regards,
Nikita

#5

Last October Mircea sent an RFC about (dis)allowing unused check prefixes in FileCheck: https://lists.llvm.org/pipermail/llvm-dev/2020-October/146162.html
(In short, if FileCheck --check-prefix or --check-prefixes specifies a prefix which does not occur in the test, FileCheck will error. Note: a `-NOT` pattern is also counted as an occurrence.)

Mircea created a worksheet https://docs.google.com/spreadsheets/d/1o6q3XH1n3DDyyccnYZ_kVfhFbTDzC_S09e973_cwYuw/edit#gid=0 which may be a bit stale now but llvm-project has reached a state where all issues have been identified and fixed, or worked around (by opting in FileCheck --allow-unused-prefixes in some test directories, such as: clang/test/OpenMP, llvm/test/Transforms/Attributor and llvm/test/FileCheck).

We can make a switch if the community thinks that not allowing unused prefixes is the better default: https://reviews.llvm.org/D95849
For downstream projects using FileCheck and lit, it should be easy to restore the previous permissive behavior

from lit.llvm.subst import ToolSubst

fc = ToolSubst('FileCheck', unresolved='fatal')
config.substitutions.insert(0, (fc.regex, 'FileCheck --allow-unused-prefixes'))

# Note: if multiple --allow-unused-prefixes options are specified, the last wins.

Personally I am strongly in favor of disallowing unused check prefixes by default. We have identified numerous issues:

(1) typo. A misspelled check prefix does not test what it intends to test.
(2) bitrot tests due to refactoring
(3) unspecified `-NOT` patterns. Sometimes a test uses something like `--check-prefixes=COMMON,TRUE` and `--check-prefixes=COMMON,FALSE` to test both behaviors but they forget to include a `FALSE-NO:` pattern to test that some string does not appear.

(1) and (2) are especially common. There are indeed tests where --allow-unused-prefixes is more suitable - but they are sufficiently few that I think the default should be --allow-unused-prefixes=false.

So, what do folks think?

Just my 2c, but I think we should allow unused prefixes. This does catch the occasional typo, but also has a cost: Historically, certain kinds of tests simply used a certain boilerplate of check lines, because differences are common, even if they don't occur for each test. For X86 vector tests, it makes more sense to simply always include AVX1 and AVX2 test prefixes, even if it so happens that for *this* particular test, codegen is identical and only the AVX prefix ends up being used.

Is it particularly burdensome for those tests in particular to opt-in
to allowing unused prefixes on the RUN/FileCheck line? (if that's the
right tradeoff/makes things smoother compared to finding/adding the
prefixes only when they become necessary)

#6

Last October Mircea sent an RFC about (dis)allowing unused check prefixes in FileCheck: https://lists.llvm.org/pipermail/llvm-dev/2020-October/146162.html
(In short, if FileCheck --check-prefix or --check-prefixes specifies a prefix which does not occur in the test, FileCheck will error. Note: a -NOT pattern is also counted as an occurrence.)

Mircea created a worksheet https://docs.google.com/spreadsheets/d/1o6q3XH1n3DDyyccnYZ_kVfhFbTDzC_S09e973_cwYuw/edit#gid=0 which may be a bit stale now but llvm-project has reached a state where all issues have been identified and fixed, or worked around (by opting in FileCheck --allow-unused-prefixes in some test directories, such as: clang/test/OpenMP, llvm/test/Transforms/Attributor and llvm/test/FileCheck).

We can make a switch if the community thinks that not allowing unused prefixes is the better default: https://reviews.llvm.org/D95849
For downstream projects using FileCheck and lit, it should be easy to restore the previous permissive behavior

from lit.llvm.subst import ToolSubst

fc = ToolSubst(‘FileCheck’, unresolved=‘fatal’)
config.substitutions.insert(0, (fc.regex, ‘FileCheck --allow-unused-prefixes’))

Note: if multiple --allow-unused-prefixes options are specified, the last wins.

Personally I am strongly in favor of disallowing unused check prefixes by default. We have identified numerous issues:

(1) typo. A misspelled check prefix does not test what it intends to test.
(2) bitrot tests due to refactoring
(3) unspecified -NOT patterns. Sometimes a test uses something like --check-prefixes=COMMON,TRUE and --check-prefixes=COMMON,FALSE to test both behaviors but they forget to include a FALSE-NO: pattern to test that some string does not appear.

(1) and (2) are especially common. There are indeed tests where --allow-unused-prefixes is more suitable - but they are sufficiently few that I think the default should be --allow-unused-prefixes=false.

So, what do folks think?

Just my 2c, but I think we should allow unused prefixes. This does catch the occasional typo, but also has a cost: Historically, certain kinds of tests simply used a certain boilerplate of check lines, because differences are common, even if they don’t occur for each test. For X86 vector tests, it makes more sense to simply always include AVX1 and AVX2 test prefixes, even if it so happens that for this particular test, codegen is identical and only the AVX prefix ends up being used.

Is it particularly burdensome for those tests in particular to opt-in
to allowing unused prefixes on the RUN/FileCheck line? (if that’s the
right tradeoff/makes things smoother compared to finding/adding the
prefixes only when they become necessary)

Another option available is to whole-sale opt in those directories with tests wishing to make this tradeoff, via lit.local.cfg, like we already did for the Attributor (and @maskray summarized above)

#7

ISTM That the tests using update_test_checks should probably opt into allowing unused prefixes as a matter of course.

With a set of auto-updated checks, an unused prefix is unlikely to be due to a mistake, like it would indicate be for a manually maintained test. Maybe update_test_checks could even add the right argument to FileCheck automatically.

#8

Last October Mircea sent an RFC about (dis)allowing unused check prefixes
in FileCheck:
[llvm-dev] [RFC] FileCheck: (dis)allowing unused prefixes
(In short, if FileCheck --check-prefix or --check-prefixes specifies a
prefix which does not occur in the test, FileCheck will error. Note: a
`-NOT` pattern is also counted as an occurrence.)

Mircea created a worksheet
FileCheck: impact of --allow-unused-prefixes=false - Google Sheets which
may be a bit stale now but llvm-project has reached a state where all
issues have been identified and fixed, or worked around (by opting in
FileCheck --allow-unused-prefixes in some test directories, such as:
clang/test/OpenMP, llvm/test/Transforms/Attributor and llvm/test/FileCheck).

We can make a switch if the community thinks that not allowing unused
prefixes is the better default: ⚙ D95849 [FileCheck] Default --allow-unused-prefixes to false
For downstream projects using FileCheck and lit, it should be easy to
restore the previous permissive behavior

from lit.llvm.subst import ToolSubst

fc = ToolSubst('FileCheck', unresolved='fatal')
config.substitutions.insert(0, (fc.regex, 'FileCheck --allow-unused-prefixes'))

# Note: if multiple --allow-unused-prefixes options are specified, the last wins.

Personally I am strongly in favor of disallowing unused check prefixes by
default. We have identified numerous issues:

(1) typo. A misspelled check prefix does not test what it intends to test.
(2) bitrot tests due to refactoring
(3) unspecified `-NOT` patterns. Sometimes a test uses something like
`--check-prefixes=COMMON,TRUE` and `--check-prefixes=COMMON,FALSE` to test
both behaviors but they forget to include a `FALSE-NO:` pattern to test
that some string does not appear.

(1) and (2) are especially common. There are indeed tests where
--allow-unused-prefixes is more suitable - but they are sufficiently few
that I think the default should be --allow-unused-prefixes=false.

So, what do folks think?

Just my 2c, but I think we should allow unused prefixes. This does catch
the occasional typo, but also has a cost: Historically, certain kinds of
tests simply used a certain boilerplate of check lines, because differences
are common, even if they don't occur for each test. For X86 vector tests,
it makes more sense to simply always include AVX1 and AVX2 test prefixes,
even if it so happens that for *this* particular test, codegen is identical
and only the AVX prefix ends up being used. This means that whenever
codegen changes in a minor way (e.g. due to a target-independent
SimplyDemandedBits change that has no direct relation to X86) and a
difference is introduced, you need to now figure out which new prefixes you
have to add. Or drop prefixes if a codegen difference goes away. Having to
manually adjust check prefixes takes away from the usual experience of
"Just run update_(llc_)test_checks".

The comment may be related to fixes such as

7979f24954ed6c1bfac8a9961fee69a44822f2b6 "[test] Fix some unused check prefixes in test/Analysis/CostModel/X86"
and
f4467c4d3b6c65d2a0d799badb1edf233e829162 "[CodeGen][X86] Remove some unused check-prefixes and regenerate tests."

In the updates, a common pattern is to replace multi-level

   --check-prefixes=CHECK,SSE,SSE1
   --check-prefixes=CHECK,SSE,SSE2

with

   --check-prefixes=CHECK,SSE1
   --check-prefixes=CHECK,SSE2

because SSE ends up to be unneeded.

I can see arguments that 'SSE' may be needed - if the code generation
changes and SSE1/SSE2 happen to have new common, keeping 'SSE' can allow
shared check lines with 'SSE:'.

However, I personally don't consider it as good use of sharing.
Tree-style sharing is fairly brittle and difficult for automatic tools
like update{,_llc,_analyzer}_test_checks.py to update.
Many times two-level sharing is sufficient and sometimes just
duplicating everything is more robust - in the cost of verbose check
prefixes.

#9

That’s a great observation. If update_test_checks is used, then typos aren’t really possible and unused check prefixes are only unnecessary at worst (and usually “currently unnecessary”). Making a distinction based on that seems like a nice solution to me.

Regards,

Nikita

#10

Just my 2c, but I think we should allow unused prefixes. This does catch the occasional typo, but also has a cost: Historically, certain kinds of tests simply used a certain boilerplate of check lines, because differences are common, even if they don't occur for each test. For X86 vector tests, it makes more sense to simply always include AVX1 and AVX2 test prefixes, even if it so happens that for *this* particular test, codegen is identical and only the AVX prefix ends up being used. This means that whenever codegen changes in a minor way (e.g. due to a target-independent SimplyDemandedBits change that has no direct relation to X86) and a difference is introduced, you need to now figure out which new prefixes you have to add. Or drop prefixes if a codegen difference goes away. Having to manually adjust check prefixes takes away from the usual experience of "Just run update_(llc_)test_checks".

At least I personally have found the gradual migration towards disallowing unused prefixes to be more annoying than useful. I guess ergonomics could be improved if update_test_checks automatically dropped unused prefixes, but there's really no way to automatically add prefixes, without domain-specific knowledge.

ISTM That the tests using update_test_checks should probably opt into allowing unused prefixes as a matter of course.

With a set of auto-updated checks, an unused prefix is unlikely to be due to a mistake, like it would indicate be for a manually maintained test. Maybe update_test_checks could even add the right argument to FileCheck automatically.

That's a great observation. If update_test_checks is used, then typos aren't really possible and unused check prefixes are only unnecessary at worst (and usually "currently unnecessary"). Making a distinction based on that seems like a nice solution to me.

Regards,
Nikita

Thanks for the feedback! llvm/test/CodeGen has already defaulted to
--allow-unused-prefixes=false.

From this discussion, update{,_llc,_analyzer}_test_checks can be

improved to add --allow-unused-prefixes on demand.
(Such tests do exist but are not too many.)
So I will assume that the blocker has been resolved.

#11

I'll submit ⚙ D95849 [FileCheck] Default --allow-unused-prefixes to false next week.

(We actually have another thread discussing the
update_*_test_checks.py's problem
([llvm-dev] Conflicting check prefix detection not working in update_llc_test_checks.py?).
Hope we can find folks who are willing to improve the situation. (Not
affected by this FileCheck default change as far as I can see :)))