Testing GCC

https://developers.redhat.com/blog/2017/02/13/testing-testing-gcc/

by dmalcolm • 9 years ago

84 39 9 years ago

developers.redhat.com

saurik• 9 years ago

So, in GCC 7, we’ve extended the C frontend so that we can embed fragments of GIMPLE and RTL dumps as the bodies of functions.

This additionally sounds extremely interesting for reasons that have nothing to do with testing (though, I bet most of the things I'd want to do this with this will continue to require the expanded platform support for __attribute__((__naked__)) that has blocked a lot of my use cases for inline assembly and which does not seem to be something that is "wanted" by GCC, though I should verify it isn't just due to no one providing a patch... it isn't as if there aren't other people asking for it).

monochromatic• 9 years ago

Took forever to load, then gave me "Page Does Not Exist." Anybody else having issues?

pirocks• 9 years ago

I too am having issues.

twiss• 9 years ago

I filed a report through the "report a website issue" button at the blog homepage. Hope they fix it soon.

smhenderson• 9 years ago

I am getting the same. Tried about 15 minutes ago and then again just now. Still showing the Page Does Not Exist error.

ParadoxOryx• 9 years ago

Here's the Google cached version - http://webcache.googleusercontent.com/search?q=cache:https:/...

egwynn• 9 years ago

For me, that version refreshes indefinitely. I recommend reading the plain-text version: http://webcache.googleusercontent.com/search?q=cache:https:/...

dmalcolm6• 9 years ago

[author of the post here]

Thanks; am chasing it up at my end

i336_2• 9 years ago

Thanks for chasing it up!

Would I be right in guessing that we're looking at a well-built security system responding to high page load?

Saying the page doesn't exist in these types of situations is interesting. Users consider 404s permanent, so some proportion is unlikely to come back. Furthermore, because each user is hitting the page individually, there's going to be very little out-of-band "no it doesn't work", so it won't reduce server load by much.

From a security standpoint, if high load like this is causing "lp0 on fire"-type issues, a generic 404 may be the correct answer, depending on the situation. A 404 is quite generic, but I'm sure it could still leak useful information depending on the scenario.

Whatever's going on, hearing what actually happened would be very cool (and educational, most people here are Web developers!).

drfuchs• 9 years ago

Great, but now how will you find out that a particular optimization pass hasn't been completely obviated by earlier passes, since you won't be looking for C code that actually triggers it?

correnos• 9 years ago

Well, you could embed a machine-checked proof that instruction sequences of the targeted form will sometimes survive previous optimization passes and therefore the given optimization will always be contributing something.

But that's hard, so every compiler I've used leaves it to human engineering instead.

astrange• 9 years ago

You can emit stats for time and # of instructions changed by each pass. gcc certainly does the first, don't remember the second.

bonzini• 9 years ago

You can do both kinds of tests: check that a pass still optimizes some GIMPLE or RTL input the same way, and check that the C code is optimized the same way. In turn, the latter can either look at per-pass dumps, or at the final assembly code for one or more architectures.

greglindahl4• 9 years ago

I never got around to actually implementing the following in the PathScale compiler, but my idea was to add logging in all the places where allocations do something, and then create lists of optimizations that I expect to have successfully fired in various functions or programs. This would be especially handy for checking that new optimizations didn't disable important things for various SPECcpu benchmarks.

greglindahl4• 9 years ago

Sorry, "the places where optimizations do something". Didn't notice until after the edit window was long passed. I blame autocorrect :-)

wolf550e4• 9 years ago

So is GCC adding a toolchain similar to the one LLVM has around its bitcode?

In this instance, a way to write a text representation of an internal data structure that represents code, run a single optimization pass over that data structure, dump the result, compare the result with serialized text representation of "expected value"?

Will they add the other parts, to extract an llvm-like core out of gcc and implement the C and C++ compilers and link time optimizer as users of that core?

ori_b• 9 years ago

So is GCC adding a toolchain similar to the one LLVM has around its bitcode?

Not as far as I can tell.

In this instance, a way to write a text representation of an internal data structure that represents code, run a single optimization pass over that data structure, dump the result, compare the result with serialized text representation of "expected value"?

That's already existed for ages. See the -fdump-tree-* family of functions. It also doesn't do anything llvm-like.

wolf550e4• 9 years ago

So all that hasn't existed for ages. Only dumping the intermediate representation has existed for ages, but not loading it and not running the middle-end of the compiler on the loaded IR.

namelezz• 9 years ago

As a relative newcomer to the project, one of my “pain points” learning GCC’s internals was the custom garbage collector it uses to manage memory

It's interesting that GCC has a custom GC.

wolf550e4• 9 years ago

https://en.wikipedia.org/wiki/Greenspun's_tenth_rule ?

dmalcolm6• 9 years ago

[author of the article]

The RTL dump format does look a lot like Lisp, FWIW; see: https://gcc.gnu.org/onlinedocs/gccint/RTL.html or this article I wrote a while back on it: http://developerblog.redhat.com/2014/09/16/improving-gccs-in...

(This aspect of GCC is probably the oldest, written by RMS back in the late 1980s, I believe).

pekk• 9 years ago

Middlebrow dismissal? Memory management is a common concern across languages, and languages which do not include garbage collection tend to host a variety of approaches depending on the author and what sort of thing they were building. Writing C code that uses garbage collection does not equal Common Lisp.

greglindahl4• 9 years ago

I'm be surprised if it didn't have something related to memory allocation and free. Open64 (previously the SGI compiler) has an "arena" memory allocator to improve memory locality and make freeing scratch memory easier. Looks like LLVM uses arena allocation, too.

Joky• 9 years ago

LLVM does it only for very particular things. Otherwise it is using malloc a lot. The ownership is interesting as well, as some pieces are "leaked" to the LLVMContext, which has some granularity: it owns the module you're working on, so can't be destroyed before you don't need the module anymore.

pizlonator2• 9 years ago

Some people swear by arenas, some people swear by GCs, some swear by stacks, and others swear by malloc. There are probably people who swear by things I didn't list. Compiler writers are people, too, and so they swear by many memory allocation styles.

When I first started writing compilers, I learned the craft from people who happened to also be GC zealots, and they insisted that it was not possible to write a compiler any other way.

I remember once talking to a VM hacker who wrote compilers in C++ and was shocked that it was even possible to write a compiler in Java.

Then I saw LLVM, which mostly just relies on new/delete and an intuitive ownership model. I've done that ever since.

But as a memory management nerd I have to say that there are some advantages to every approach. Arenas are cheap and intuitive. GCs mean you don't have to care about lifetimes. Malloc/free minimizes object drag. Pick what you like and don't let anyone tell you that any one of these approaches is inherently better than the others.

dmalcolm6• 9 years ago

[author of the post here]

Yes: memory management in the compiler is interesting. There's a complicated graph of pointer references. Most of the time the compiler is building something relatively small, so we don't need to bother cleaning up, we just exit without freeing it all (for speed). But when e.g. building with Link Time Optimization, we can use large amounts of RAM, so a garbage-collection can be needed.

raverbashing• 9 years ago

Well, besides the big testing of compiling an entire linux distribution, yes, it is good to have specific tests for some particular issues in the compiler (manageable pass/fail tests that can identify a problem quickly)

greglindahl4• 9 years ago

gcc has always had a huge unit test suite. Not a new thing.

sanxiyn• 9 years ago

gcc had a huge integration test suite, which is different from unit test suite.

hermitdev• 9 years ago

Yeah...I got hit by a regression in a bug fix in RH's GCC back in the 4.1.2 era. The revision/build number only moved by a dozen or 2, but they completely broke C++ anonymous namespaces at global scope (caused an ICE). Glad I didn't let Ops upgrade all of my boxes at the same time like they wanted...

boris• 9 years ago

Interesting to see complex pieces of software including features that are primarily used for testing. I guess this is a natural next step after having testing micro-formats[1].

One thing I don't like about GCC's testing framework is the number of files. If you need to test 10 related things (say, diagnostics for a particular language construct), then you will have 10 different test files. I would much prefer for them to be in a single file. This is one of the goals of Testscript[2].

[1] https://blog.nelhage.com/2016/12/how-i-test/

[2] https://build2.org/build2/doc/build2-testscript-manual.xhtml

dmalcolm6• 9 years ago

[author of the post]

Thanks for the links; a lot of interesting material in there.

FWIW, DejaGnu gives some flexibility over how granular the test files can be. For example, here's part of the test case for the -Wmisleading-indentation warning I added in gcc 6: https://github.com/gcc-mirror/gcc/blob/master/gcc/testsuite/...

i.e. that's one 700 line test source file, expressing most of the test coverage for one feature.

I realize now that the post should have had an example of what such a test case has; see the "dg-warning" directives in that file.

So that's one "single file" test case. That said, I had to put some of the other test cases for that feature into different test files: some of them interact with the preprocessor in "interesting" ways, so it was simplest to split them out.

However, when a test starts failing, it's handy if the failing test is small (especially when stepping through the latter in the debugger...).

So in our testing we have a mix of both big test files covering a lot of material, and small test files, covering very specific corner cases.

Another aspect of granularity is how long the testsuite takes to run: each test source file in the DejaGnu suite involves exec-ing the compiler, which adds a fixed amount of overhead. It parallelizes well, but it's still advantageous to stuff more testing into fewer files (with the obvious tradeoff w.r.t. debugability if something starts failing). This was another reason for adding the unit-testing suite: this part is all done in-process, and so these ~30,000 tests run in less than a second.

pizlonator2• 9 years ago

I'm super impressed by the size of their test suite. I wish I had a test suite like that!

Testing GCC

Related Stories