new intrinsic: @isComptime(value: var) bool

I need this to optimize fmt.charToDigit:

pub fn charToDigit(c: u8, radix: u8) (error{InvalidCharacter}!u4) {
    // These are manual optimizations for the common cases.
    // They only require a single branch instruction.
    if (radix == 10) { <== I need to test @isComptime(radix)
        if (std.ascii.isDigit(c)) {
            return c - '0';
        } else {
            return error.InvalidCharacter;
        }
    } else (radix == 16) {
        const NOT: u8 = 0xff;
        const swtch = []u8{
        //    0,   1,   2,   3,   4,   5,   6,   7,   8,   9,  10,  11,  12,  13,  14,  15
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
              0,   1,   2,   3,   4,   5,   6,   7,   8,   9, NOT, NOT, NOT, NOT, NOT, NOT,

            NOT,  10,  11,  12,  13,  14,  15, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT,  10,  11,  12,  13,  14,  15, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,

            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,

            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
            NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
        };
        const digit = swtch[c];
        if (digit == NOT) {
            return error.InvalidCharacter;
        }
        return u4(digit);
    }

    const value = switch (c) {
        '0'...'9' => c - '0',
        'A'...'Z' => c - 'A' + 10,
        'a'...'z' => c - 'a' + 10,
        else => return error.InvalidCharacter,
    };

    if (value >= radix) return error.InvalidCharacter;

    return value;
}

what code did you use to test the performance of charToDigit?

https://gist.github.com/shawnl/41ecc272cf77e6e7490589bee630063f

$ zig run ~/projects/shared/bench_charToDigit.zig  --release-fast
optimized: 279.4MiB/sec, checksum: 962597485
simple: 160.3MiB/sec, checksum: 962597485
optimized: 222.8MiB/sec, checksum: 921599919
simple: 149.8MiB/sec, checksum: 921599919

nice, however some safety check is making the optimized hex code very slow. Is there a way to unconditionally optimize a block of code?:

$ zig run ~/projects/shared/bench_charToDigit.zig 
optimized: 39.4MiB/sec, checksum: 962597485
simple: 36.0MiB/sec, checksum: 962597485
optimized: 1.7MiB/sec, checksum: 921599919
simple: 35.0MiB/sec, checksum: 921599919

This particular case you can actually fix by pulling the switch table out and making it global.

For example:

const std = @import("std");
const Utf8ByteSequenceLengthError = error{Utf8InvalidStartByte};

const NOT = 0xff;
const swtch = []u8{
    //    0,   1,   2,   3,   4,   5,   6,   7,   8,   9,  10,  11,  12,  13,  14,  15
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    0, 1, 2, 3, 4, 5, 6, 7, 8, 9, NOT, NOT, NOT, NOT, NOT, NOT,

    NOT, 10, 11, 12, 13, 14, 15, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, 10, 11, 12, 13, 14, 15, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,

    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,

    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
    NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT, NOT,
};

pub fn charToDigitOptimized(c: u8, radix: u8) (error{InvalidCharacter}!u6) {
    // These are manual optimizations for the common cases.
    // They only require a single branch instruction.

// ... rest of code

$ zig run original.zig 
optimized: 53.7MiB/sec, checksum: 962597485
simple: 48.7MiB/sec, checksum: 962597485
optimized: 2.0MiB/sec, checksum: 921599919
simple: 46.8MiB/sec, checksum: 921599919

$ zig run original.zig --release-fast
optimized: 311.6MiB/sec, checksum: 962597485
simple: 177.1MiB/sec, checksum: 962597485
optimized: 275.6MiB/sec, checksum: 921599919
simple: 168.7MiB/sec, checksum: 921599919

$ zig run new.zig 
optimized: 54.3MiB/sec, checksum: 962597485
simple: 49.6MiB/sec, checksum: 962597485
optimized: 61.3MiB/sec, checksum: 921599919
simple: 47.7MiB/sec, checksum: 921599919

$ zig run new.zig --release-fast
optimized: 315.9MiB/sec, checksum: 962597485
simple: 177.5MiB/sec, checksum: 962597485
optimized: 280.3MiB/sec, checksum: 921599919
simple: 166.7MiB/sec, checksum: 921599919

It appeared that Zig wasn't eliding a memcpy otherwise in debug mode. Alternatively you can initialize it in a comptime block locally and it should work fine. Might be clearer if you did that and set the digit ranges selectively.

As for whether explicitly specifying if the argument radix was comptime, that didn't help here. I did try initially a variation with the radix explicitly comptime, but saw no difference as the main issue was the hidden memcpy.

Also, #868 is related a little to this issue.

It will be used by the zig native compiler to parse Char and String escape sequences.

…

do you really need to optimize charToDigit though? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#2128 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAfLcqEYMIfctZLp0fCTqAsBEDqxFA4oks5vbYYpgaJpZM4cRgnF> .

I think the optimization is find without checking for comprime, because having it not comptime is so unusual, however it would make it more correct.

optimize fmt.charToDigit

Related: #868

I don't think this is a good use case for this feature. It would be better to have the same code path for comptime and runtime.

Navigation Menu

Repository navigation

new intrinsic: @isComptime(value: var) bool #2128

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

new intrinsic: @isComptime(value: var) bool #2128

Description

Activity

andrewrk commented on Mar 29, 2019

shawnl commented on Mar 29, 2019

tiehuis commented on Mar 29, 2019

shawnl commented on Mar 29, 2019

shawnl commented on Mar 29, 2019

andrewrk commented on Oct 17, 2019

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions