lack of AVX/2 not detected correctly when generating the build executable

[essess]

Zig Version

0.14.0

Steps to Reproduce and Observed Behavior

The long read is here (with some troubleshooting steps performed along the way)

The short is that I am running zig on a clean debian install within a virtualbox vm. I think what's happening here is that virtualbox does not play well with hyper-v enabled (windows host) and disables AVX/2 instructions to the guest. I've followed all of the various random articles/etc to ensure that AVX/2 is exposed to the guest --- but have failed for various reasons outside of my control because this is a corporate owned/controlled equipment.

This results in the generated build executable raising a SIGILL for a VMOVD instruction used. The build can no longer continue.

Of maybe special note, 0.13 works out of the box for me in the guest. Zig works out of the box just fine in the windows host (no matter the version).

Expected Behavior

I expect that AVX/2 instructions are not generated when they're not available during the build process for the build executable.

Maybe the build executable is super dumb and just uses the bare-minimum instructions no matter what to get the process up to a point where better decisions can get made about what to actually target?

[alexrp]

Can you please include (from within the guest):

• cat /proc/cpuinfo
• zig build-exe --show-builtin (both for 0.13.0 and 0.14.0 would be good)

[essess]

cat /proc/cpuinfo -> cpuinfo.txt
zig build-exe --show-builtin -> 0.13.show-builtin.txt
zig build-exe --show-builtin -> 0.14.show-builtin.txt

[alexrp]

My best guess here is that we correctly detect your CPU as being Alder Lake, which means we make certain baseline assumptions from that model, such as avx2. However, when we then go to refine those results based on actual cpuid bits, we're failing to toggle off some features because we just aren't checking the CPUID bits for them.

I suspect it's this feature in particular that's the problem here:

zig/lib/std/Target/x86.zig

Line 1343 in 6e8493d

.avxvnni,

Note its dependencies:

zig/lib/std/Target/x86.zig

Lines 462 to 468 in 6e8493d

	result[@intFromEnum(Feature.avxvnni)] = .{
	.llvm_name = "avxvnni",
	.description = "Support AVX_VNNI encoding",
	.dependencies = featureSet(&[_]Feature{
	.avx2,
	}),
	};

You can see how that becomes a problem here:

zig/lib/std/zig/system/x86.zig

Lines 67 to 72 in 6e8493d

	// Add the CPU model's feature set into the working set, but then
	// override with actual detected features again.
	cpu.features.addFeatureSet(cpu.model.features);
	detectNativeFeatures(&cpu, os.tag);
	cpu.features.populateDependencies(cpu.arch.allFeaturesList());

zig/lib/std/zig/system/x86.zig

Lines 439 to 440 in 6e8493d

	// AVX2 is only supported if we have the OS save support from AVX.
	setFeature(cpu, .avx2, bit(leaf.ebx, 5) and has_avx_save);

We correctly disable avx2, but because avxvnni is still on, populateDependencies() will just toggle avx2 right back on again.

[alexrp]

Please give this patch a try:

diff --git a/lib/std/zig/system/x86.zig b/lib/std/zig/system/x86.zig
index 428561c371..7667b18e86 100644
--- a/lib/std/zig/system/x86.zig
+++ b/lib/std/zig/system/x86.zig
@@ -490,15 +490,12 @@ fn detectNativeFeatures(cpu: *Target.Cpu, os_tag: Target.Os.Tag) void {
         setFeature(cpu, .pconfig, bit(leaf.edx, 18));
         setFeature(cpu, .uintr, bit(leaf.edx, 5));
 
-        // TODO I feel unsure about this check.
-        //      It doesn't really seem to check for 7.1, just for 7.
-        //      Is this a sound assumption to make?
-        //      Note that this is what other implementations do, so I kind of trust it.
-        const has_leaf_7_1 = max_level >= 7;
-        if (has_leaf_7_1) {
+        if (leaf.eax >= 1) {
             leaf = cpuid(0x7, 0x1);
+            setFeature(cpu, .avxvnni, bit(leaf.eax, 4) and has_avx_save);
             setFeature(cpu, .avx512bf16, bit(leaf.eax, 5) and has_avx512_save);
         } else {
+            setFeature(cpu, .avxvnni, false);
             setFeature(cpu, .avx512bf16, false);
         }
     } else {

[essess]

No change.
A VMOVD is still being generated.

zig build-exe --show-builtin (for x86_64) -> 0.14.0.x86_64.show-builtin.txt
zig build-exe --show-builtin (for x86) -> 0.14.0.x86.show-builtin.txt

I'll have some time tomorrow morning to try and step through this and watch what's going on.
I can also revert and try anything else you'd like.

[alexrp]

Did you rebuild the compiler with the patch? Just applying it to the standard library is not sufficient.

[essess]

Alright, I'll figure that out and report back.

EDIT:
This works.
I applied your patch to the 0.14.0 tag.

[essess]

Would you like me to handle the change/PR and send your way?
I'm not sure of the next step to take here.

[alexrp]

I'll do a PR for 0.14.1. I need to thoroughly audit our CPUID logic in that file anyway; we're missing lots of other feature bits.