[llvm-dev] FileCheck idiom difficulties

[James Henderson via llvm-dev]

 

Hi all,

Many of our lit tests use FileCheck and a tool like llvm-readobj to check properties of a section header/symbol/etc. A typical (pseudoised for brevity) output to match against might be something like the following:

Symbols [

  Symbol {

    Name: foo

    Value: 0

    Type: Function

    Section: .foo (1)

  }

  Symbol {

    Name: bar

    Value: 1

    Type: Object

    Section: .foo (1)

  }

]

and your lit test might want to check the properties of the foo symbol like so:

# CHECK:      Name: foo

# CHECK-NEXT: Value: 0

# CHECK-NEXT: Type: Function

# CHECK-NEXT: Section: .foo (1)

This is fine. But what if you only care about the section of a symbol, and not the value or type etc? You could do the following:

# CHECK: Name: foo

# CHECK: Section: .foo (1)

Hopefully some of you will already notice the problem with this approach: if foo was in, say, the .baz section, the test will spuriously pass, because the Section line will match the Section line for .bar. One alternative to this is to explicitly match each field in between, using CHECK-NEXT:

# CHECK:      Name: foo

# CHECK-NEXT: Value:

# CHECK-NEXT: Type:

# CHECK-NEXT: Section: .foo (1)

This works, but somewhat hides what is really being tested by adding extra noise to the checks. In reality, there are actually other fields too that need to be listed, meaning the "interesting" parts of the test are even more hidden.

I recently started using yet another approach:

# CHECK: Name: foo

# CHECK: Section:

# CHECK-SAME:     .foo (1)

This works because the Section: matched will be the first one found, i.e. the one belonging to foo, and then .foo will be looked for on the same line. However, I noticed today that this pattern has its own problem, namely that there could be something between the Section tag and .foo. In other words, the above pattern would match "Section: .bar.foo". A couple of solutions to this are:

# CHECK: Section:

# CHECK-NOT: {{[:graph:]}}

# CHECK-SAME: .foo (1)

# CHECK: Section:

# CHECK-SAME: {{^}} .foo (1)

The first one ensures that there's no non-whitespace between the end of "Section:" and the start of ".foo (1)". The second ensures that the start of the CHECK-SAME match is the "start of line", and since the first half of the line has already been consumed, it means " .foo (1)" must immediately follow "Section:". However, the first is even less readable than the current CHECK-SAME approach, whilst the second is somewhat confusing if you don't realise that FileCheck effectively consumes the things it has matched already, so that they effectively don't exist any more.

Does anybody have any other suggestions/thoughts/comments? One idea I had was for a new directive something like "CHECK-IMMEDIATE" which is implicitly the same as the final approach I suggested above, but maybe adding a new directive to achieve this isn't the right approach?

James

[George Rimar via llvm-dev]

> One idea I had was for a new directive something like "CHECK-IMMEDIATE" which

> is implicitly the same as the final approach I suggested above, but maybe adding a new

> directive to achieve this isn't the right approach?

"CHECK-IMMEDIATE"​ (or "CHECK-CONTINUE"/better name) sound like a clean and fine approach to me.

​I'd go with it probably.

Best regards,
George | Developer | Access Softek, Inc

---

От: James Henderson <jh737...@my.bristol.ac.uk>
Отправлено: 6 ноября 2019 г. 15:17
Кому: llvm-dev; George Rimar; Fāng-ruì Sòng; Paul Robinson; tho...@graphcore.ai
Тема: FileCheck idiom difficulties

 

CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender and know the content is safe.  If you suspect potential phishing or spam email, report it to Repor...@accesssoftek.com

[Thomas Preud'homme via llvm-dev]

Aren't the name lines unique? If they are you could use the good old CHECK-LABEL:

# CHECK-LABEL: Name: foo

# CHECK: Section: .foo (1)

---

From: George Rimar <gri...@accesssoftek.com>
Sent: 06 November 2019 13:01
To: jh737...@my.bristol.ac.uk <jh737...@my.bristol.ac.uk>
Cc: Fāng-ruì Sòng <mas...@google.com>; llvm-dev <llvm...@lists.llvm.org>; Thomas Preud'homme <tho...@graphcore.ai>; Paul Robinson <paul.r...@sony.com>
Subject: Re: FileCheck idiom difficulties

 

** We have updated our privacy policy, which contains important information about how we collect and process your personal data. To read the policy, please click here **

This email and its attachments are intended solely for the addressed recipients and may contain confidential or legally privileged information.
If you are not the intended recipient you must not copy, distribute or disseminate this email in any way; to do so may be unlawful.

Any personal data/special category personal data herein are processed in accordance with UK data protection legislation.
All associated feasible security measures are in place. Further details are available from the Privacy Notice on the website and/or from the Company.

Graphcore Limited (registered in England and Wales with registration number 10185006) is registered at 107 Cheapside, London, UK, EC2V 6DN.
This message was scanned for viruses upon transmission. However Graphcore accepts no liability for any such transmission.

[David Blaikie via llvm-dev]

I'd usually write this sort of test (& there are many in the DWARF emission tests):

# CHECK:      Name: foo

# CHECK-NOT: }

# CHECK: Section: .foo (1)

Or, if you're checking every section, you could use a --implicit-check-not=Symbol or similar.

_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

[James Henderson via llvm-dev]

I keep forgetting about CHECK-LABEL, thanks! In many cases, the name lines are unique, so this would probably work, although with sections, for example, they don't have to be (but in 99% of tests they can be). I think in the limited cases where CHECK-LABEL isn't possible, we can use one of the uglier syntaxes.

[James Henderson via llvm-dev]

Thanks, I had considered using the CHECK-NOT approach. That would work in most cases, I believe, but I'm always nervous to use it because CHECK-NOTs can start failing to check the right thing if output changes. For example, George Rimar recently made changes that turned some list-like structs into dictionaries (using curly brackets instead of square brackets) in the readobj output, which could theoretically have affected the '}' style.

However, the implicit-check-not approach probably works for many cases. Thanks.

[Reid Kleckner via llvm-dev]

I think we should think about changing llvm-readobj to produce more filecheck friendly output.

This was a problem we used to have with LLVM debug info metadata: it wasn't very structured, and it was printed in some arbitrary order. Then Duncan (I think) added the DI* classes, which made it easier to match something semantic, like `DILocalVariable.*name: "foo"`, and standardized on a topological output ordering, so you could start from the variable, then match the type, and then find the type metadata number later.

If we printed `Section (.foo) {` for example, that would help some.

It still doesn't help establish delimited regions for properties printed across multiple lines for readability, though... Should we add some kind of ad-hoc delimiter balancing to FileCheck? Something like:

CHECK-SCOPE: Section {

CHECK: Name: asdf

CHECK: Name: asdf  

CHECK-ENDSCOPE: }

A SCOPE directive line would have to end in a known delimiter, '(', '{', or '['. The ENDSCOPE directive would only match lines with delimiters that balance with the opening delimiter. It would kind of work for filechecking JSON, for example.

The scope idea here is pretty half-baked, but it's food for thought.

On Wed, Nov 6, 2019 at 4:18 AM James Henderson via llvm-dev <llvm...@lists.llvm.org> wrote:

[Robinson, Paul via llvm-dev]

I have to say, I’m not super excited about the answer always being “let’s add another FileCheck directive.”  FileCheck is basically a meta-grep, that knows how to execute some sub-searches in different orders.  Asking it to do things like implicit paren balancing is really out of scope for the tool.

--paulr

[David Blaikie via llvm-dev]

On Wed, Nov 6, 2019 at 11:12 AM Reid Kleckner via llvm-dev <llvm...@lists.llvm.org> wrote:

I think we should think about changing llvm-readobj to produce more filecheck friendly output.

This was a problem we used to have with LLVM debug info metadata: it wasn't very structured, and it was printed in some arbitrary order. Then Duncan (I think) added the DI* classes, which made it easier to match something semantic, like `DILocalVariable.*name: "foo"`, and standardized on a topological output ordering, so you could start from the variable, then match the type, and then find the type metadata number later.

If we printed `Section (.foo) {` for example, that would help some.

It still doesn't help establish delimited regions for properties printed across multiple lines for readability, though... Should we add some kind of ad-hoc delimiter balancing to FileCheck? Something like:

CHECK-SCOPE: Section {

CHECK: Name: asdf

CHECK: Name: asdf  

CHECK-ENDSCOPE: }

There's some overlap between ^ and this existing construct \/

CHECK: Section {
CHECK-DAG: Name: asdf
CHECK-DAG: Name: asdf
CHECK: }
 

[George Rimar via llvm-dev]

> Or, if you're checking every section, you could use a --implicit-check-not=Symbol or similar.

'implicit-check-not' should be used with caution, btw. It can cause matches against the file path in some situations.

See: https://reviews.llvm.org/rL361621, https://reviews.llvm.org/rL361563​

George.

[James Henderson via llvm-dev]

FWIW, I did a BoF on the llvm binutils earlier this year at Euro LLVM, to discuss the direction of the binutils and one of the suggestions that seemed to get broad appeal was a more machine-readable output format option (e.g. JSON/XML). llvm-readobj's output is relatively human readable, but isn't quite JSON. You could possibly then write a straightforward python script to load the json and test it's properties exactly, if you're interested in the property values rather than how they are printed. I haven't got around to this yet, and probably won't for quite some time, so if anybody else is interested in the idea, I'd be happy for them to take it forward.

James