llvm dev - May 2018 - [RFC] Formalizing FileCheck Features

Background
----------

FileCheck [0] is a cornerstone testing tool for the LLVM project.  It
has grown new features over the years to meet new needs, but these
sometimes have surprising and counter-intuitive behavior [1].  This
has become even more evident in Joel Denny's recent quest to repair
what seemed like an obvious defect [2] but which led me to the
conclusion [3] that FileCheck sorely needed a clear, intuitive
conceptual model.  And then someone to make it work that way (hi
Joel!).

Basic Conceptual Model
----------------------

FileCheck should operate on the basis of these three fundamental
concepts.

(1) Search range.  This is some substring of the input text where one
or more directives will do their pattern-matching magic.

(2) Match range.  This is a substring of a search range where a
directive (or in one case, a group of directives) has matched a
pattern.

(3) Directive groups.  These are sequences of adjacent directives that
operate in a related way on a search range.  Directives within a group
are processed in order, except as noted in the directive description.

Finally we add The Rule:  No match ranges may overlap.

(This is largely formalizing what FileCheck already does, except that
it didn't have The Rule with respect to DAG matches.  That's the bug
that Joel was originally trying to fix, until I stuck my nose into
it.)

Directive Descriptions Based On Conceptual Model
------------------------------------------------

Given the conceptual model, all directives can be defined in terms of
it. This is possibly going overboard with the formalism but hey, we're
all compiler geeks here.

CHECK: Scans the search range for a pattern match. Fails if no match
is found.  The end of the match range becomes the start of the search
range for subsequent directives.

CHECK-SAME: Like CHECK, plus there must be zero newlines prior to the
start of the match range.

CHECK-NEXT: Like CHECK, plus there must be exactly one newline prior
to the start of the match range.

CHECK-LABEL: All LABEL directives are processed before any other
directives.  These directives have two effects.  First, they act like
CHECK directives, but also partition the input text into disjoint
search ranges, delimited by the match ranges of the LABEL directives.
Second, they partition the remaining directives into Label Groups,
each of which operates on the corresponding search range.  For truly
pedantic formalism, we can say there are implicit LABEL directives
matching the start and end of the entire input text, thus all
non-LABEL directives are always in some Label Group.

CHECK-NOT: A sequence of NOT directives forms a NOT Group. The group
is not executed immediately; instead the next non-NOT directive is
executed first, and the start of that directive's match range becomes
the end of the NOT Group's search range.  (If the next directive is
LABEL, it has already executed and has a match range, which is already
the end of the search range.)  After the NOT Group's search range is
defined, each NOT directive scans the range for a match, and fails if
a match is found.

CHECK-DAG: A sequence of DAG directives forms a DAG Group. The group
is not executed immediately; instead the next non-DAG directive is
executed first, and the start of that directive's match range becomes
the end of the DAG Group's search range.  If the next directive is
CHECK-NOT, the end of the DAG Group's search range is
unaffected. (This might or might not be FileCheck's historical
behavior; I didn't check.)  After the DAG Group's search range is
defined, each DAG directive scans the range for a match, and fails if
a match is not found.  Per The Rule, match ranges for DAG directives
may not overlap. (This is not historical FileCheck behavior, and the
bug Joel Denny wanted to fix.)  After all DAG directives run, the
match range for the entire DAG Group extends from the start of the
earliest match to the end of the latest match.  The end of that match
range becomes the start of the search range for subsequent directives.

Observations
------------

A CHECK-NOT still separates surrounding CHECK-DAG directives into
disjoint groups, and does not permit matches from the two groups to
overlap. DAG was originally implemented to detect and diagnose an
overlap, but this worked only for the first DAG after a NOT. This can
lead to counter-intuitive behavior and potentially makes certain kinds
of matches impossible.

Putting CHECK-SAME and CHECK-NEXT after CHECK-DAG now has defined
behavior, but it's unlikely to be useful.  Putting SAME or NEXT as the
first directive in a file likewise has defined behavior, matching
precisely the first or second line (respectively) of the input text.


References
----------
[0] https://llvm.org/docs/CommandGuide/FileCheck.html
[1] https://www.youtube.com/watch?v=4rhW8knj0L8
[2] https://lists.llvm.org/pipermail/llvm-dev/2018-May/123010.html
[3] https://reviews.llvm.org/D47106

On 05/24/2018 08:46 AM, via llvm-dev wrote:
> Background
> ----------
>
> FileCheck [0] is a cornerstone testing tool for the LLVM project.  It
> has grown new features over the years to meet new needs, but these
> sometimes have surprising and counter-intuitive behavior [1].  This
> has become even more evident in Joel Denny's recent quest to repair
> what seemed like an obvious defect [2] but which led me to the
> conclusion [3] that FileCheck sorely needed a clear, intuitive
> conceptual model.
Thanks for writing this up. I definitely think that it will be good to
add this to FileCheck's documentation.
>   And then someone to make it work that way (hi
> Joel!).
>
> Basic Conceptual Model
> ----------------------
>
> FileCheck should operate on the basis of these three fundamental
> concepts.
>
> (1) Search range.  This is some substring of the input text where one
> or more directives will do their pattern-matching magic.
>
> (2) Match range.  This is a substring of a search range where a
> directive (or in one case, a group of directives) has matched a
> pattern.
>
> (3) Directive groups.  These are sequences of adjacent directives that
> operate in a related way on a search range.  Directives within a group
> are processed in order, except as noted in the directive description.
>
> Finally we add The Rule:  No match ranges may overlap.
>
> (This is largely formalizing what FileCheck already does, except that
> it didn't have The Rule with respect to DAG matches.  That's the
bug
> that Joel was originally trying to fix, until I stuck my nose into
> it.)
>
> Directive Descriptions Based On Conceptual Model
> ------------------------------------------------
>
> Given the conceptual model, all directives can be defined in terms of
> it. This is possibly going overboard with the formalism but hey, we're
> all compiler geeks here.
>
> CHECK: Scans the search range for a pattern match. Fails if no match
> is found.  The end of the match range becomes the start of the search
> range for subsequent directives.
>
> CHECK-SAME: Like CHECK, plus there must be zero newlines prior to the
> start of the match range.
>
> CHECK-NEXT: Like CHECK, plus there must be exactly one newline prior
> to the start of the match range.
>
> CHECK-LABEL: All LABEL directives are processed before any other
> directives.  These directives have two effects.  First, they act like
> CHECK directives, but also partition the input text into disjoint
> search ranges, delimited by the match ranges of the LABEL directives.
> Second, they partition the remaining directives into Label Groups,
> each of which operates on the corresponding search range.  For truly
> pedantic formalism, we can say there are implicit LABEL directives
> matching the start and end of the entire input text, thus all
> non-LABEL directives are always in some Label Group.
>
> CHECK-NOT: A sequence of NOT directives forms a NOT Group. The group
> is not executed immediately; instead the next non-NOT directive is
> executed first, and the start of that directive's match range becomes
> the end of the NOT Group's search range.
Both here, and for CHECK-DAG, we should say something about reaching the
end of the input.

 -Hal
>   (If the next directive is
> LABEL, it has already executed and has a match range, which is already
> the end of the search range.)  After the NOT Group's search range is
> defined, each NOT directive scans the range for a match, and fails if
> a match is found.
>
> CHECK-DAG: A sequence of DAG directives forms a DAG Group. The group
> is not executed immediately; instead the next non-DAG directive is
> executed first, and the start of that directive's match range becomes
> the end of the DAG Group's search range.  If the next directive is
> CHECK-NOT, the end of the DAG Group's search range is
> unaffected. (This might or might not be FileCheck's historical
> behavior; I didn't check.)  After the DAG Group's search range is
> defined, each DAG directive scans the range for a match, and fails if
> a match is not found.  Per The Rule, match ranges for DAG directives
> may not overlap. (This is not historical FileCheck behavior, and the
> bug Joel Denny wanted to fix.)  After all DAG directives run, the
> match range for the entire DAG Group extends from the start of the
> earliest match to the end of the latest match.  The end of that match
> range becomes the start of the search range for subsequent directives.
>
> Observations
> ------------
>
> A CHECK-NOT still separates surrounding CHECK-DAG directives into
> disjoint groups, and does not permit matches from the two groups to
> overlap. DAG was originally implemented to detect and diagnose an
> overlap, but this worked only for the first DAG after a NOT. This can
> lead to counter-intuitive behavior and potentially makes certain kinds
> of matches impossible.
>
> Putting CHECK-SAME and CHECK-NEXT after CHECK-DAG now has defined
> behavior, but it's unlikely to be useful.  Putting SAME or NEXT as the
> first directive in a file likewise has defined behavior, matching
> precisely the first or second line (respectively) of the input text.
>
>
> References
> ----------
> [0] https://llvm.org/docs/CommandGuide/FileCheck.html
> [1] https://www.youtube.com/watch?v=4rhW8knj0L8
> [2] https://lists.llvm.org/pipermail/llvm-dev/2018-May/123010.html
> [3] https://reviews.llvm.org/D47106
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-- 
Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory

> -----Original Message-----
> From: Hal Finkel [mailto:hfinkel at anl.gov]
> Sent: Thursday, May 24, 2018 10:33 AM
> To: Robinson, Paul; llvm-dev at lists.llvm.org
> Subject: Re: [llvm-dev] [RFC] Formalizing FileCheck Features
> 
> 
> On 05/24/2018 08:46 AM, via llvm-dev wrote:
> > Background
> > ----------
> >
> > FileCheck [0] is a cornerstone testing tool for the LLVM project.  It
> > has grown new features over the years to meet new needs, but these
> > sometimes have surprising and counter-intuitive behavior [1].  This
> > has become even more evident in Joel Denny's recent quest to
repair
> > what seemed like an obvious defect [2] but which led me to the
> > conclusion [3] that FileCheck sorely needed a clear, intuitive
> > conceptual model.
> 
> Thanks for writing this up. I definitely think that it will be good to
> add this to FileCheck's documentation.
> 
> >   And then someone to make it work that way (hi
> > Joel!).
> >
> > Basic Conceptual Model
> > ----------------------
> >
> > FileCheck should operate on the basis of these three fundamental
> > concepts.
> >
> > (1) Search range.  This is some substring of the input text where one
> > or more directives will do their pattern-matching magic.
> >
> > (2) Match range.  This is a substring of a search range where a
> > directive (or in one case, a group of directives) has matched a
> > pattern.
> >
> > (3) Directive groups.  These are sequences of adjacent directives that
> > operate in a related way on a search range.  Directives within a group
> > are processed in order, except as noted in the directive description.
> >
> > Finally we add The Rule:  No match ranges may overlap.
> >
> > (This is largely formalizing what FileCheck already does, except that
> > it didn't have The Rule with respect to DAG matches.  That's
the bug
> > that Joel was originally trying to fix, until I stuck my nose into
> > it.)
> >
> > Directive Descriptions Based On Conceptual Model
> > ------------------------------------------------
> >
> > Given the conceptual model, all directives can be defined in terms of
> > it. This is possibly going overboard with the formalism but hey,
we're
> > all compiler geeks here.
> >
> > CHECK: Scans the search range for a pattern match. Fails if no match
> > is found.  The end of the match range becomes the start of the search
> > range for subsequent directives.
> >
> > CHECK-SAME: Like CHECK, plus there must be zero newlines prior to the
> > start of the match range.
> >
> > CHECK-NEXT: Like CHECK, plus there must be exactly one newline prior
> > to the start of the match range.
> >
> > CHECK-LABEL: All LABEL directives are processed before any other
> > directives.  These directives have two effects.  First, they act like
> > CHECK directives, but also partition the input text into disjoint
> > search ranges, delimited by the match ranges of the LABEL directives.
> > Second, they partition the remaining directives into Label Groups,
> > each of which operates on the corresponding search range.  For truly
> > pedantic formalism, we can say there are implicit LABEL directives
> > matching the start and end of the entire input text, thus all
> > non-LABEL directives are always in some Label Group.
> >
> > CHECK-NOT: A sequence of NOT directives forms a NOT Group. The group
> > is not executed immediately; instead the next non-NOT directive is
> > executed first, and the start of that directive's match range
becomes
> > the end of the NOT Group's search range.
> 
> Both here, and for CHECK-DAG, we should say something about reaching the
> end of the input.
> 
>  -Hal
It seemed intuitive to me that a range can't extend past the end of the
input, and under CHECK-LABEL I did say there are implicit directives 
matching the start and end of the input; but it does no harm to add some 
words about that to DAG and NOT.
--paulr
> 
> >   (If the next directive is
> > LABEL, it has already executed and has a match range, which is already
> > the end of the search range.)  After the NOT Group's search range
is
> > defined, each NOT directive scans the range for a match, and fails if
> > a match is found.
> >
> > CHECK-DAG: A sequence of DAG directives forms a DAG Group. The group
> > is not executed immediately; instead the next non-DAG directive is
> > executed first, and the start of that directive's match range
becomes
> > the end of the DAG Group's search range.  If the next directive is
> > CHECK-NOT, the end of the DAG Group's search range is
> > unaffected. (This might or might not be FileCheck's historical
> > behavior; I didn't check.)  After the DAG Group's search range
is
> > defined, each DAG directive scans the range for a match, and fails if
> > a match is not found.  Per The Rule, match ranges for DAG directives
> > may not overlap. (This is not historical FileCheck behavior, and the
> > bug Joel Denny wanted to fix.)  After all DAG directives run, the
> > match range for the entire DAG Group extends from the start of the
> > earliest match to the end of the latest match.  The end of that match
> > range becomes the start of the search range for subsequent directives.
> >
> > Observations
> > ------------
> >
> > A CHECK-NOT still separates surrounding CHECK-DAG directives into
> > disjoint groups, and does not permit matches from the two groups to
> > overlap. DAG was originally implemented to detect and diagnose an
> > overlap, but this worked only for the first DAG after a NOT. This can
> > lead to counter-intuitive behavior and potentially makes certain kinds
> > of matches impossible.
> >
> > Putting CHECK-SAME and CHECK-NEXT after CHECK-DAG now has defined
> > behavior, but it's unlikely to be useful.  Putting SAME or NEXT as
the
> > first directive in a file likewise has defined behavior, matching
> > precisely the first or second line (respectively) of the input text.
> >
> >
> > References
> > ----------
> > [0] https://llvm.org/docs/CommandGuide/FileCheck.html
> > [1] https://www.youtube.com/watch?v=4rhW8knj0L8
> > [2] https://lists.llvm.org/pipermail/llvm-dev/2018-May/123010.html
> > [3] https://reviews.llvm.org/D47106
> >
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> 
> --
> Hal Finkel
> Lead, Compiler Technology and Programming Languages
> Leadership Computing Facility
> Argonne National Laboratory

Hi Paul,

On Thu, May 24, 2018 at 9:46 AM, <paul.robinson at sony.com> wrote:
> Background
> ----------
>
> FileCheck [0] is a cornerstone testing tool for the LLVM project.  It
> has grown new features over the years to meet new needs, but these
> sometimes have surprising and counter-intuitive behavior [1].  This
> has become even more evident in Joel Denny's recent quest to repair
> what seemed like an obvious defect [2] but which led me to the
> conclusion [3] that FileCheck sorely needed a clear, intuitive
> conceptual model.

Agreed.  Thanks for doing this.


> And then someone to make it work that way (hi
> Joel!).
>
Sure, I can help with the implementation given that I'm running into these
issues a lot in my own work.  As I'm a bit too close to the FileCheck
implementation at this point, I would suggest that someone else write the
initial specification-based tests to find the deviations from the
description we arrive at.  Paul, you're the obvious person for that one.  I
can of course work on further implementation-based testing.

I also recommend we make changes toward the new specification incrementally.

> Basic Conceptual Model
> ----------------------
>
> FileCheck should operate on the basis of these three fundamental
> concepts.
>
> (1) Search range.  This is some substring of the input text where one
> or more directives will do their pattern-matching magic.
>
> (2) Match range.  This is a substring of a search range where a
> directive (or in one case, a group of directives) has matched a
> pattern.
>
> (3) Directive groups.  These are sequences of adjacent directives that
> operate in a related way on a search range.  Directives within a group
> are processed in order, except as noted in the directive description.
>
> Finally we add The Rule:  No match ranges may overlap.
>
> (This is largely formalizing what FileCheck already does, except that
> it didn't have The Rule with respect to DAG matches.  That's the
bug
> that Joel was originally trying to fix, until I stuck my nose into
> it.)
>
I agree with The Rule.  I haven't found any real use case yet that needs to
violate that rule.

> Directive Descriptions Based On Conceptual Model
> ------------------------------------------------
>
> Given the conceptual model, all directives can be defined in terms of
> it. This is possibly going overboard with the formalism but hey, we're
> all compiler geeks here.
>
I think it's great.

>
> CHECK: Scans the search range for a pattern match. Fails if no match
> is found.  The end of the match range becomes the start of the search
> range for subsequent directives.
>
> CHECK-SAME: Like CHECK, plus there must be zero newlines prior to the
> start of the match range.
>
... within the search range.

Should it be possible for CHECK-SAME match range to include newlines?

>
> CHECK-NEXT: Like CHECK, plus there must be exactly one newline prior
> to the start of the match range.
>
... within the search range.

Your choice to talk about the match range rather than the search range for
CHECK-SAME and CHECK-NEXT implies you like the current behavior that
extends the search range beyond these match range restrictions and then
complains if the match range restrictions aren't met.  For example,
CHECK-SAME searches past the newline and then complains if the match range
starts after the newline.  Is that what you prefer?

I'd note that, in the case of CHECK-NEXT, that choice can restrict what
CHECK-NEXT can match.  That is, it will complain about a match on the
previous line rather than skip it and look on the next line.

>
> CHECK-LABEL: All LABEL directives are processed before any other
> directives.  These directives have two effects.  First, they act like
> CHECK directives, but also partition the input text into disjoint
> search ranges, delimited by the match ranges of the LABEL directives.
> Second, they partition the remaining directives into Label Groups,
> each of which operates on the corresponding search range.  For truly
> pedantic formalism, we can say there are implicit LABEL directives
> matching the start and end of the entire input text, thus all
> non-LABEL directives are always in some Label Group.
>
> CHECK-NOT: A sequence of NOT directives forms a NOT Group. The group
> is not executed immediately; instead the next non-NOT directive is
> executed first, and the start of that directive's match range becomes
> the end of the NOT Group's search range.

Based on the following, that wording is not quite right when a DAG group
follows, so there should probably be some note about that here.


>   (If the next directive is
> LABEL, it has already executed and has a match range, which is already
> the end of the search range.)  After the NOT Group's search range is
> defined, each NOT directive scans the range for a match, and fails if
> a match is found.
>
> CHECK-DAG: A sequence of DAG directives forms a DAG Group. The group
> is not executed immediately; instead the next non-DAG directive is
> executed first, and the start of that directive's match range becomes
> the end of the DAG Group's search range.

That's definitely a change from the current behavior.  Currently, the DAG
group finds its own end based on the farthest match.

>   If the next directive is
> CHECK-NOT, the end of the DAG Group's search range is
> unaffected.

Unaffected means that it's as if there's no following directive?  So
next
CHECK-LABEL (possibly the implicit one at EOF)?  What if there's a CHECK,
CHECK-NEXT, or CHECK-SAME after all the DAGs and NOTs?

> (This might or might not be FileCheck's historical
> behavior; I didn't check.)  After the DAG Group's search range is
> defined, each DAG directive scans the range for a match, and fails if
> a match is not found.  Per The Rule, match ranges for DAG directives
> may not overlap. (This is not historical FileCheck behavior, and the
> bug Joel Denny wanted to fix.)  After all DAG directives run, the
> match range for the entire DAG Group extends from the start of the
> earliest match to the end of the latest match.  The end of that match
> range becomes the start of the search range for subsequent directives.
>
That last sentence contradicts the first few sentences: the subsequent
directive has already been matched.

One point not addressed here is the start of the DAG group's search range.
Currently, if the DAG group is preceded by a NOT group preceded by a DAG
group, the last DAG group's search range starts at the start of the first
DAG group's match range.  Any matches in the first DAG group's match
range
produces a reordering error.  This is somewhat similar to the CHECK-SAME
and CHECK-NEXT behavior I mentioned earlier: the search ranges permit
invalid match ranges and then complain about them in an effort to diagnose
mistakes.  However, that restricts what can be matched.

I'm not claiming that either behavior is best.  It's not clear to me. 
The
best use of DAG-NOT-DAG is very confusing to me.  An effort to prescribe
the right semantics to it needs to be informed by real use cases, in my
opinion.

>
> Observations
> ------------
>
> A CHECK-NOT still separates surrounding CHECK-DAG directives into
> disjoint groups, and does not permit matches from the two groups to
> overlap. DAG was originally implemented to detect and diagnose an
> overlap, but this worked only for the first DAG after a NOT. This can
> lead to counter-intuitive behavior and potentially makes certain kinds
> of matches impossible.
>
I definitely agree it shouldn't be just the first DAG.  The reordering
detection should happen for all consecutive DAGs after the NOT or none of
them.

>
> Putting CHECK-SAME and CHECK-NEXT after CHECK-DAG now has defined
> behavior, but it's unlikely to be useful.

I believe they had predictable behavior before (their search ranges started
at the end of the match range for the entire CHECK-DAG), but it's different
with the above description (they define the end of the search range for the
preceding CHECK-DAG group).

Thanks.

Joel

>   Putting SAME or NEXT as the
> first directive in a file likewise has defined behavior, matching
> precisely the first or second line (respectively) of the input text.
>
>
> References
> ----------
> [0] https://llvm.org/docs/CommandGuide/FileCheck.html
> [1] https://www.youtube.com/watch?v=4rhW8knj0L8
> [2] https://lists.llvm.org/pipermail/llvm-dev/2018-May/123010.html
> [3] https://reviews.llvm.org/D47106
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180524/bfe1bed4/attachment-0001.html>

Awesome, thanks for helping to improve the conceptual model here, I agree that
this is really important - particularly for a tool like FileCheck.

One observation: if certain combinations of directives have dubious or
surprising behavior, it is perfectly fine for FileCheck to reject them.

-Chris

> On May 24, 2018, at 6:46 AM, via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> Background
> ----------
> 
> FileCheck [0] is a cornerstone testing tool for the LLVM project.  It
> has grown new features over the years to meet new needs, but these
> sometimes have surprising and counter-intuitive behavior [1].  This
> has become even more evident in Joel Denny's recent quest to repair
> what seemed like an obvious defect [2] but which led me to the
> conclusion [3] that FileCheck sorely needed a clear, intuitive
> conceptual model.  And then someone to make it work that way (hi
> Joel!).
> 
> Basic Conceptual Model
> ----------------------
> 
> FileCheck should operate on the basis of these three fundamental
> concepts.
> 
> (1) Search range.  This is some substring of the input text where one
> or more directives will do their pattern-matching magic.
> 
> (2) Match range.  This is a substring of a search range where a
> directive (or in one case, a group of directives) has matched a
> pattern.
> 
> (3) Directive groups.  These are sequences of adjacent directives that
> operate in a related way on a search range.  Directives within a group
> are processed in order, except as noted in the directive description.
> 
> Finally we add The Rule:  No match ranges may overlap.
> 
> (This is largely formalizing what FileCheck already does, except that
> it didn't have The Rule with respect to DAG matches.  That's the
bug
> that Joel was originally trying to fix, until I stuck my nose into
> it.)
> 
> Directive Descriptions Based On Conceptual Model
> ------------------------------------------------
> 
> Given the conceptual model, all directives can be defined in terms of
> it. This is possibly going overboard with the formalism but hey, we're
> all compiler geeks here.
> 
> CHECK: Scans the search range for a pattern match. Fails if no match
> is found.  The end of the match range becomes the start of the search
> range for subsequent directives.
> 
> CHECK-SAME: Like CHECK, plus there must be zero newlines prior to the
> start of the match range.
> 
> CHECK-NEXT: Like CHECK, plus there must be exactly one newline prior
> to the start of the match range.
> 
> CHECK-LABEL: All LABEL directives are processed before any other
> directives.  These directives have two effects.  First, they act like
> CHECK directives, but also partition the input text into disjoint
> search ranges, delimited by the match ranges of the LABEL directives.
> Second, they partition the remaining directives into Label Groups,
> each of which operates on the corresponding search range.  For truly
> pedantic formalism, we can say there are implicit LABEL directives
> matching the start and end of the entire input text, thus all
> non-LABEL directives are always in some Label Group.
> 
> CHECK-NOT: A sequence of NOT directives forms a NOT Group. The group
> is not executed immediately; instead the next non-NOT directive is
> executed first, and the start of that directive's match range becomes
> the end of the NOT Group's search range.  (If the next directive is
> LABEL, it has already executed and has a match range, which is already
> the end of the search range.)  After the NOT Group's search range is
> defined, each NOT directive scans the range for a match, and fails if
> a match is found.
> 
> CHECK-DAG: A sequence of DAG directives forms a DAG Group. The group
> is not executed immediately; instead the next non-DAG directive is
> executed first, and the start of that directive's match range becomes
> the end of the DAG Group's search range.  If the next directive is
> CHECK-NOT, the end of the DAG Group's search range is
> unaffected. (This might or might not be FileCheck's historical
> behavior; I didn't check.)  After the DAG Group's search range is
> defined, each DAG directive scans the range for a match, and fails if
> a match is not found.  Per The Rule, match ranges for DAG directives
> may not overlap. (This is not historical FileCheck behavior, and the
> bug Joel Denny wanted to fix.)  After all DAG directives run, the
> match range for the entire DAG Group extends from the start of the
> earliest match to the end of the latest match.  The end of that match
> range becomes the start of the search range for subsequent directives.
> 
> Observations
> ------------
> 
> A CHECK-NOT still separates surrounding CHECK-DAG directives into
> disjoint groups, and does not permit matches from the two groups to
> overlap. DAG was originally implemented to detect and diagnose an
> overlap, but this worked only for the first DAG after a NOT. This can
> lead to counter-intuitive behavior and potentially makes certain kinds
> of matches impossible.
> 
> Putting CHECK-SAME and CHECK-NEXT after CHECK-DAG now has defined
> behavior, but it's unlikely to be useful.  Putting SAME or NEXT as the
> first directive in a file likewise has defined behavior, matching
> precisely the first or second line (respectively) of the input text.
> 
> 
> References
> ----------
> [0] https://llvm.org/docs/CommandGuide/FileCheck.html
> [1] https://www.youtube.com/watch?v=4rhW8knj0L8
> [2] https://lists.llvm.org/pipermail/llvm-dev/2018-May/123010.html
> [3] https://reviews.llvm.org/D47106
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

Thanks Joel and Chris, comments inline.
>> CHECK: Scans the search range for a pattern match. Fails if no match
>> is found.  The end of the match range becomes the start of the search
>> range for subsequent directives.
>>
>> CHECK-SAME: Like CHECK, plus there must be zero newlines prior to the
>> start of the match range.
>
> ... within the search range.
Yes, thanks.
> Should it be possible for CHECK-SAME match range to include newlines?
 
It is possible to write a regex that matches newlines.  Doing that in
CHECK-SAME seems a bit odd but I don't think it's worth trying to forbid
it.
>> CHECK-NEXT: Like CHECK, plus there must be exactly one newline prior
>> to the start of the match range.
>
> ... within the search range.
Again yes.
> Your choice to talk about the match range rather than the search range
> for CHECK-SAME and CHECK-NEXT implies you like the current behavior
> that extends the search range beyond these match range restrictions and
> then complains if the match range restrictions aren't met.  For
example,
> CHECK-SAME searches past the newline and then complains if the match
> range starts after the newline.  Is that what you prefer?
>
> I'd note that, in the case of CHECK-NEXT, that choice can restrict what
> CHECK-NEXT can match.  That is, it will complain about a match on the
> previous line rather than skip it and look on the next line.
Ah, so we could define CHECK-NEXT as: move the start of the search
range past the first newline, then behaves as CHECK-SAME?

But, appending {{.*$}} to the previous pattern should have the same
effect if you have a CHECK-NEXT that runs into that problem.  And I
do think it's valuable for SAME and NEXT to tell you they found
matches but not on the line you asked for.  So I'd prefer to leave
these defined as they are.
>> CHECK-NOT: A sequence of NOT directives forms a NOT Group. The group
>> is not executed immediately; instead the next non-NOT directive is
>> executed first, and the start of that directive's match range
becomes
>> the end of the NOT Group's search range.
>
> Based on the following, that wording is not quite right when a DAG
> group follows, so there should probably be some note about that here.
So, "the next non-NOT directive or DAG group is executed ... the start
of that directive or group's match range ..." ?
 
>>  (If the next directive is
>> LABEL, it has already executed and has a match range, which is already
>> the end of the search range.)  After the NOT Group's search range
is
>> defined, each NOT directive scans the range for a match, and fails if
>> a match is found.
>>
>> CHECK-DAG: A sequence of DAG directives forms a DAG Group. The group
>> is not executed immediately; instead the next non-DAG directive is
>> executed first, and the start of that directive's match range
becomes
>> the end of the DAG Group's search range.
>
> That's definitely a change from the current behavior.  Currently, the
> DAG group finds its own end based on the farthest match.
Oh good catch.  Copy-thinko from the NOT description.  NOT is the only
kind of directive that has deferred execution.
 
>>  If the next directive is
>> CHECK-NOT, the end of the DAG Group's search range is
>> unaffected.
>
> Unaffected means that it's as if there's no following directive? 
So
> next CHECK-LABEL (possibly the implicit one at EOF)?  What if there's
> a CHECK, CHECK-NEXT, or CHECK-SAME after all the DAGs and NOTs?
If DAG doesn't have deferred execution then the end of the search range
is the next (explicit or implicit) CHECK-LABEL point, end of story.
 
>>  After all DAG directives run, the
>> match range for the entire DAG Group extends from the start of the
>> earliest match to the end of the latest match.  The end of that match
>> range becomes the start of the search range for subsequent directives.
>
> That last sentence contradicts the first few sentences: the subsequent
> directive has already been matched.
Right, fixing the previous bug means this sentence says the right thing.
> One point not addressed here is the start of the DAG group's search
> range.  Currently, if the DAG group is preceded by a NOT group
> preceded by a DAG group, the last DAG group's search range starts at
> the start of the first DAG group's match range.  Any matches in the
> first DAG group's match range produces a reordering error.  This is
> somewhat similar to the CHECK-SAME and CHECK-NEXT behavior I mentioned
> earlier: the search ranges permit invalid match ranges and then
> complain about them in an effort to diagnose mistakes.  However, that
> restricts what can be matched.
>
> I'm not claiming that either behavior is best.  It's not clear to
me.
> The best use of DAG-NOT-DAG is very confusing to me.  An effort to
> prescribe the right semantics to it needs to be informed by real use
> cases, in my opinion.
I did some email archaeology, and found this exchange on llvm-dev between
myself and Michael Liao (original DAG implementor) 13 Mar 2016:

pr> Commentary in FileCheck itself can easily be interpreted to mean the
pr> intent was that –NOT would scan the region between the points defined
pr> by the last match of the preceding DAG group (which the code gets
pr> right) and the first match of the following DAG group (which the code
pr> does not get right). But the commentary is not really that clear.

ml> That's the intention of the original design. CHECK-NOT never occurs
ml> before we find the start point (the start of file by default) and end
ml> point (the end of file by default.) All other points are through other
ml> CHECKs, including CHECK-DAG but excluding CHECK-NOT.  So that, if you
ml> use CHECK-NOT, you need to be aware of how that range is defined. As
ml> CHECK-DAG pattern matches a group of pattern in any order, the match
ml> point of that group of CHECK-DAG (a consecutive CHECK-DAGs without any
ml> other CHECKs interleaved) is always the point where one of that pgroup
ml> is matched. If one CHECK-DAG is separated by any other CHECKs
ml> (including CHECK-NOT) from preceding CHECK-DAGs, it is not in the
ml> preceding group of CHECK-DAG. That's way how we could check the order
ml> where a group of patterns should never occur before another group of
ml> patterns.

So, I believe my specification for the interaction between DAG and NOT
does match the original intent.  Regarding the diagnostic aid, it does
make some sequences really hard to match, and I don't have a general
idea how to fix that (versus {{.*$}} for the similar NEXT situation).
It's also a reasonable continuation of the behavior of plain CHECK, in 
that a second CHECK doesn't search the prior text to complain about 
ordering issues.

SAME and NEXT are, I think, a different category; that has to do with
line-breaks that are not explicitly described by user-written patterns,
and my own experience is that it's helpful to be told that something
matches but isn't on the line I expected.

So, I don't have a definitive answer for changing DAG-NOT-DAG, but
intuitively the spec makes sense to me and my inclination is to think
the diagnostic isn't hugely valuable.
>> Putting CHECK-SAME and CHECK-NEXT after CHECK-DAG now has defined
>> behavior, but it's unlikely to be useful.
>
> I believe they had predictable behavior before (their search ranges
> started at the end of the match range for the entire CHECK-DAG), but
> it's different with the above description (they define the end of the
> search range for the preceding CHECK-DAG group).
You're right, it was predictable before, and I am fixing the bug where
the directive after DAG gets executed first so the range isn't affected.

Taking Chris Lattner's point into consideration, we might want to say
SAME or NEXT after a DAG should be an error.  But we could also leave
that for a later round.

--paulr

P.S. I am away next week but expect to keep an eye on the lists.