recent performance regression in the x86 backend on print1mil benchmark

[andrewrk]

print1mil.zig:

const std = @import("std");

fn print() void {
    std.io.getStdOut().writeAll("hello, world!\n") catch return;
}

pub fn main() !void {
    print();
    print();
    // ...

zig build-exe print1mil.zig -fno-llvm -fno-lld -fstrip

0.12.0-dev.617+d1e39b691:   804ms
0.12.0-dev.694+937e8cb70:   919ms
0.12.0-dev.704+8e1421f19:   774ms
0.12.0-dev.706+3bd1b9e15:   855ms
0.12.0-dev.708+8470652f1:   853ms
0.12.0-dev.709+c3042cbe1:  2840ms
0.12.0-dev.714+0f1652dc6:  2670ms
0.12.0-dev.733+6734d2117:  3020ms
0.12.0-dev.772+9f0d2f941:  3030ms
0.12.0-dev.927+895c81ce7:  6770ms
0.12.0-dev.1237+b82459fa4: 6350ms

c3042cb is the first bad commit