[validation-test] More granular FixedPoint tests

[modocache]

FixedPoint.swift.gyb generates a massive amount of assertions: one for each Int type (such as UInt8 and Int64) as a source, to one of each Int type as a destination, for each of 27 bit patterns. By my math that's 8 * 8 * 27 == 1728 assertions. As a result, when one of those assertions fails, it's difficult to
tell what went wrong.

Split each assertion into its own test case. This makes the test take a little longer to run, but it produces much more valuable output when it fails.

[modocache]

I would mark this as "NFC", but it actually fixes a SIGSEGV when the test is run on android-armv7. I think the change is beneficial for all of the supported Swift architectures regardless, but for the curious, I've written up what I learned debugging android-armv7 here: SwiftAndroid#17. Feedback welcome!

[gribozavr]

@modocache I don't see a problem accepting this PR, but the fact that long functions cause silent miscompiles is alarming. I recommend you to investigate further, taking the .swift input file and trying to reduce it to a minimal input that crashes. You can use delta (http://delta.tigris.org/) for that.

[gribozavr]

@swift-ci Please test

[modocache]

@gribozavr Absolutely will do. Thanks for the delta tip!

[modocache]

Weird, it looks like CI encountered an error in test/IRGen/protocol_resilience.sil, which should be unrelated. Perhaps a stochastic failure?

[gribozavr]

I think master is broken... I'm going to revert something and try again.

[gribozavr]

I have fixed test/IRGen/protocol_resilience.sil in c8a3579

[modocache]

@gribozavr I believe I may be running into the same issue in another test: validation-test/stdlib/Sort.swift.gyb. That test mentions that gyb has trouble with nested loops and references rdar://problem/1890035. This test references rdar://problem/17548877. I'm going to try using delta to narrow things down, but if those radars have some relevant information could you share it with me? 🙇

[gribozavr]

rdar://problem/17548877 is fixed. It was about this code:

$ cat ../a.gyb
% foo = 1
% for bar in [ 10, 20 ]:
twice
%   foo += 1
% end

once

rdar://problem/18900352 is also fixed. It was about this code:

% for i1 in [1, 2]:
%   for i2 in [10, 20]:
%     i3 = i1 + i2
${i1} + ${i2} = ${i3}
% end

Now both snippets seem to execute correctly.

[gribozavr]

@modocache It is likely that you are hitting a different bug in the same area.

[modocache]

@gribozavr Excellent, thanks for the tips. I'll continue to look into it!