Today i was investigating effects of MAP_POPULATE mmap flag, and noticed that Zig’s stdlib allocator intrefacte touches every byte of memory even when compiled in ReleaseFast.
This line:
is not eliminated by optimizations as of Zig 0.11.0.
Is this intentional? I would expect ReleaseFast to not incur extra memory scan
(not really related, but, for the curious, the benchmark for populate is here: https://github.com/matklad/benchmarks/tree/master/prefault)
Seems like a big regression, and pretty straightforward to fix. Likely this flew under the radar with the new compiler implementation. @memset to undefined is certainly supposed to be a no-op in ReleaseFast optimization mode. In theory this is LLVM’s fault since memset to undefined technically helps the optimizer more than a no-op does, but if it can’t handle it in practice then Zig should simply omit it in the LLVM backend.
I looked into this a bit today. I think you should open an issue for discussion, because at least in small examples I am observing status quo working as designed:
export fn foo(ptr: [*]u8, len: usize) void {
@memset(ptr[0..len], undefined);
}
After optimizations I get:
define dso_local void @foo(ptr nocapture nonnull align 1 %0, i64 %1) local_unnamed_addr #0 {
ret void
}
The relevant source code in the compiler:
if (elem_val.isUndefDeep(mod)) {
// Even if safety is disabled, we still emit a memset to undefined since it conveys
// extra information to LLVM. However, safety makes the difference between using
// 0xaa or actual undefined for the fill byte.
const fill_byte = if (safety)
try o.builder.intValue(.i8, 0xaa)
else
try o.builder.undefValue(.i8);
…and then it calls LLVM’s memset intrinsic.
I think if you found an issue here, it is an LLVM bug and should be reported there. I take back what I said about this being a straightforward fix.
Ok, upon double checking, it actually does work in 0.11.0 as expected, and only 0.10.0 is affected!
22:20:50|~/tmp
λ zig version && zig build-exe -O ReleaseFast ./main.zig && perf stat -- ./main
0.10.1
Performance counter stats for './main':
2,589.83 msec task-clock # 1.000 CPUs utilized
7 context-switches # 2.703 /sec
0 cpu-migrations # 0.000 /sec
2,621,447 page-faults # 1.012 M/sec
12,124,763,418 cpu_core/cycles/ # 4.682 G/sec
<not counted> cpu_atom/cycles/ (0.00%)
16,540,599,726 cpu_core/instructions/ # 6.387 G/sec
<not counted> cpu_atom/instructions/ (0.00%)
3,400,612,061 cpu_core/branches/ # 1.313 G/sec
<not counted> cpu_atom/branches/ (0.00%)
1,824,870 cpu_core/branch-misses/ # 704.628 K/sec
<not counted> cpu_atom/branch-misses/ (0.00%)
72,744,610,344 cpu_core/slots/ # 28.089 G/sec
19,398,562,758 cpu_core/topdown-retiring/ # 26.7% Retiring
1,996,910,872 cpu_core/topdown-bad-spec/ # 2.7% Bad Speculation
13,407,830,141 cpu_core/topdown-fe-bound/ # 18.4% Frontend Bound
37,941,306,571 cpu_core/topdown-be-bound/ # 52.2% Backend Bound
3,993,821,744 cpu_core/topdown-heavy-ops/ # 5.5% Heavy Operations # 21.2% Light Operations
570,545,963 cpu_core/topdown-br-mispredict/ # 0.8% Branch Mispredict # 2.0% Machine Clears
6,561,278,580 cpu_core/topdown-fetch-lat/ # 9.0% Fetch Latency # 9.4% Fetch Bandwidth
26,815,660,283 cpu_core/topdown-mem-bound/ # 36.9% Memory Bound # 15.3% Core Bound
2.590434744 seconds time elapsed
0.268046000 seconds user
2.322399000 seconds sys
22:21:22|~/tmp
λ nix shell nixpkgs#zig_0_11
22:21:25|~/tmp
λ zig version && zig build-exe -O ReleaseFast ./main.zig && perf stat -- ./main
0.11.0
Performance counter stats for './main':
0.06 msec task-clock # 0.132 CPUs utilized
0 context-switches # 0.000 /sec
0 cpu-migrations # 0.000 /sec
9 page-faults # 151.174 K/sec
225,243 cpu_core/cycles/ # 3.783 G/sec
<not counted> cpu_atom/cycles/ (0.00%)
246,533 cpu_core/instructions/ # 4.141 G/sec
<not counted> cpu_atom/instructions/ (0.00%)
50,879 cpu_core/branches/ # 854.621 M/sec
<not counted> cpu_atom/branches/ (0.00%)
1,466 cpu_core/branch-misses/ # 24.625 M/sec
<not counted> cpu_atom/branch-misses/ (0.00%)
1,351,458 cpu_core/slots/ # 22.701 G/sec
259,691 cpu_core/topdown-retiring/ # 19.3% Retiring
127,196 cpu_core/topdown-bad-spec/ # 9.4% Bad Speculation
577,682 cpu_core/topdown-fe-bound/ # 42.9% Frontend Bound
381,588 cpu_core/topdown-be-bound/ # 28.3% Backend Bound
42,398 cpu_core/topdown-heavy-ops/ # 3.1% Heavy Operations # 16.1% Light Operations
127,196 cpu_core/topdown-br-mispredict/ # 9.4% Branch Mispredict # 0.0% Machine Clears
413,387 cpu_core/topdown-fetch-lat/ # 30.7% Fetch Latency # 12.2% Fetch Bandwidth
222,593 cpu_core/topdown-mem-bound/ # 16.5% Memory Bound # 11.8% Core Bound
0.000449748 seconds time elapsed
0.000553000 seconds user
0.000000000 seconds sys
22:21:30|~/tmp
λ bat main.zig
pub fn main() !void {
_ = try @import("std").heap.page_allocator.alloc(u8, 10 * 1024 * 1024 * 1024);
}
I guess I messed up ReleaseFast flags when originally testing with Zig 0.11.0