Add "automatic" JSON facilities

[daurnimator]

Recreating #3155 (github doesn't seem to allow reopening PRs)

• dump json of arbitrary types
• load json of arbitrary types

[data-man]

Vectors?

[daurnimator]

I don't understand the CI failure

204/1301 json.test "std-aarch64-linux-none-Debug-bare-multi parseFree descends into tagged union"...access of inactive union field
/home/vsts/work/1/s/lib/std/json.zig:1672:53: 0x37d473 in json.test "std-aarch64-linux-none-Debug-bare-multi parseFree descends into tagged union" (test)
    testing.expectEqualSlices(u8, "withąunicode", r.string);
                                                    ^

This line is directly following a check of the tag type:

    testing.expectEqual(@TagType(T).string, @as(@TagType(T), r));
    testing.expectEqualSlices(u8, "withąunicode", r.string);

[Vexu]

Might be #4295? No idea why the previous line works.

[daurnimator]

Bah; rebased on master and CI is still failing on that one line of code.

The failure seems to only happen on aarch64.

[LemonBoy]

Bad news: you're hitting a LLVM bug!

Consider this small piece of LLVM IR:

; ModuleID = 'foo'
source_filename = "foo"
target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
target triple = "aarch64v8.1a-unknown-linux-musl"

define internal fastcc void @foobar(i2*) unnamed_addr {
Entry:
  store i2 -2, i2* %0
  ret void
}

LLVM 9 assembles it as:

	mov	w8, #254
	strb	w8, [x0]
	ret

Notice the wrong integer literal 😱 every other backend gets it right so nobody hit this bug before:

	movb	$2, (%rdi)
	retq

The good news is that LLVM 10 already fixed this problem and produces the following asm:

	mov	w8, #2
	strb	w8, [x0]
	ret

[daurnimator]

@LemonBoy seems like you were looking into this at same time as me; Perhaps move your comment to #4492?

Is there a simple workaround to the LLVM bug so we can merge this PR?

[LemonBoy]

Is there a simple workaround so the LLVM bug so we can merge this PR?

Not really, we'd have to widen the enum base type to a minumum of 8 bits... but that'd wreak havoc on pretty much every other piece of code :\

[andrewrk]

This is some really nice work, and a useful abstraction as well.

This is very close to merge-ready. I'm here to help get it landed; reach out if you have any questions or want to hand it off.

[andrewrk]

you could take inspiration from

zig/lib/std/json/write_stream.zig

Lines 24 to 31 in e8a8492

	/// The string used for indenting.
	one_indent: []const u8 = " ",
	/// The string used as a newline character.
	newline: []const u8 = "\n",
	/// The string used as spacing.
	space: []const u8 = " ",

[andrewrk]

This seems like it should be somehow configurable on a per-field basis. Not necessary to solve this problem for this PR though.

[daurnimator]

I looked into adding support for overriding on a per-type basis:

    pub const Override = struct {
        comptime T: type,
        comptime jsonStringify: fn (
            value: T,
            options: StringifyOptions,
            context: var,
            comptime Errors: type,
            comptime output: fn (@TypeOf(context), []const u8) Errors!void,
        ) Errors!void,
    };
    comptime overrides: []Override = &[_]Override,

However this hit a few issues:

• The "comptime struct field missing initialization value" error codepath doesn't seem to consider default values
• I am unable to refer to the T field from the function type declaration: "error: use of undeclared identifier 'T'"
• I get error: use of undeclared identifier 'context' for that function type declaration

[andrewrk]

this is so ugly, that I'm motivated to try to solve #2727

(not a merge blocker though)