Rewrite std.meta.staticMap using AliasAssign

[WalterBright]

Try out dlang/dmd#12029

[dlang-bot]

Thanks for your pull request, @WalterBright!

Bugzilla references

Your PR doesn't reference any Bugzilla issue.

If your PR contains non-trivial changes, please reference a Bugzilla issue or create a manual changelog.

Testing this PR locally

If you don't have a local development environment setup, you can use Digger to test this PR:

dub run digger -- build "master + phobos#7756"

[WalterBright]

The error is:

std/conv.d(5710): Error: static assert:  `!__traits(compiles, emplace(& s2, 1))` is false

but I'm not seeing where staticMap is being called?

[WalterBright]

@9il can you please have a look at the buildkite failures in Mir?

[WalterBright]

dscanner fails with the new syntax (buildkite)

[WalterBright]

The buildkite druntime error is:

generated/linux/release/64/unittest/test_runner core.sys.windows.dde
--| generated/linux/release/64/benchmark.o:benchmark/runbench.d:function _D8runbench6runCmdFAyabIAaZQi: error: undefined reference to '_D3std7process12executeShellFNeMAxaxHAyaAyaEQBqQBp6ConfigmMQBbQzZSQCm8typecons__T5TupleTiVQCba6_737461747573TQCuVQCya6_6f7574707574ZQBz'

[WalterBright]

@s-ludwig In buildkite taggedalgebraic I'm getting an error:

source/taggedalgebraic/taggedunion.d(449,13): Error: function `taggedalgebraic.taggedunion.TaggedUnion!(E).TaggedUnion.set!E.text.set()` is not callable using argument types `(string)`

[9il]

@WalterBright

Looks like the following mir-core unittest is failed because of a syntax error (regresssion), that is probably is unrelated to this MR. The CI compiler can't handle static F[3][] vals(F) = ... expression anymore.

@safe @nogc nothrow version(mir_core_test) unittest
{
    import std.meta: AliasSeq;
    static F[3][] vals(F) =    // value,exp,ldexp
    [
    [    0,    0,    0],
    [    1,    0,    1],
    [    -1,    0,    -1],
    [    1,    1,    2],
    [    123,    10,    125952],
    [    F.max,    int.max,    F.infinity],
    [    F.max,    -int.max,    0],
    [    F.min_normal,    -int.max,    0],
    ];
    static foreach(F; AliasSeq!(double, real))
    {{
        int i;

        for (i = 0; i < vals!F.length; i++)
        {
            F x = vals!F[i][0];
            int exp = cast(int) vals!F[i][1];
            F z = vals!F[i][2];
            F l = ldexp(x, exp);
            assert(feqrel(z, l) >= 23);
        }
    }}
}

mir-algorithm is failed because of another thing, I did a new release to so it will print at least what UDA keys are reflected with the new staticMap.

[s-ludwig]

source/taggedalgebraic/taggedunion.d(449,13): Error: function `taggedalgebraic.taggedunion.TaggedUnion!(E).TaggedUnion.set!E.text.set()` is not callable using argument types `(string)`

It looks like there is a dependency cycle between staticMap!(TypeOf, EnumMembers!FieldEnum) (which is indexed to declare the parameter type for that set method) and the template TypeOf. This works right now, but seems fishy anyway. I'll have to take a closer look tomorrow.

[WalterBright]

@s-ludwig thanks! Though this should behave identically to the old implementation of staticMap, warts and all.

[maxhaton]

out.pdf

Plot of dumb timings of staticMap!(Alias, int, int, int [up to 8192 times]). Starts off better, scales worse (I assume due to memory usage, even with lowmem the heap sizes are enormous)

I'll do some heavier operations later.

[WalterBright]

@maxhaton Good information. I bet the result is because the old is using a binary method to build long alias sequences, while the new is using the sequential one. Changing the implementation of staticMap to a binary method should fix that.

[maxhaton]

@WalterBright more perf plots

Same N values for fullyQualifiedName (slower, more memory usage) fullyQualifiedNameData.pdf

First few 150 plotted in more detail fullyQualifiedNameData_linear.pdf

aliasplots_linear.pdf Very clear division in the log-log plot

[WalterBright]

The old staticMap had changed since I last worked with it. Instead of calling itself, it used a function to generate a string that enumerated all the cases, and did one big (up to 150) case AliasSeq. Unfortunately, that implementation traded off the recursive call for a rather large set of compile-time allocations in generating the string.

[WalterBright]

@9il @s-ludwig thank you for fixing buildkite!

[9il]

@9il @s-ludwig thank you for fixing buildkite!

Something has been fixed in DMD / Phobos. I didn't really fix anything.

[9il]

I have accidentally edited @WalterBright's comment and was surprised that I am able to do so. That is a bit weird.

[WalterBright]

@9il that's because you have privileged access. It's a superpower, I only used it once when the comment writer asked me to modify it.

[s-ludwig]

Same here, might be because all of the sequences in those taggedalgebraic unit tests are <= 8 entries and the issue doesn't manifest for the first assignment.

[RazvanN7]

@WalterBright please rebase? Also, what are the plans with AliasAssign and phobos?

[RazvanN7]

I'm going to close this as #8039 seems to be a superior implementation.