std.Target.x86: purge avx10.n-256, rename avx10.n_512 to avx10.n, require evex512 for avx512f

Intel has abandoned AVX10.N/128,256; AVX10.N is now always 512-bit.
This commit is contained in:
Alex Rønne Petersen 2025-08-28 01:00:12 +02:00
parent 2c0cc81e74
commit 978555eea4
No known key found for this signature in database
3 changed files with 32 additions and 49 deletions

View File

@ -24,10 +24,8 @@ pub const Feature = enum {
amx_tile,
amx_transpose,
avx,
avx10_1_256,
avx10_1_512,
avx10_2_256,
avx10_2_512,
avx10_1,
avx10_2,
avx2,
avx512bf16,
avx512bitalg,
@ -340,9 +338,9 @@ pub const all_features = blk: {
.sse4_2,
}),
};
result[@intFromEnum(Feature.avx10_1_256)] = .{
.llvm_name = "avx10.1-256",
.description = "Support AVX10.1 up to 256-bit instruction",
result[@intFromEnum(Feature.avx10_1)] = .{
.llvm_name = "avx10.1-512",
.description = "Support AVX10.1 up to 512-bit instruction",
.dependencies = featureSet(&[_]Feature{
.avx512bf16,
.avx512bitalg,
@ -357,27 +355,11 @@ pub const all_features = blk: {
.avx512vpopcntdq,
}),
};
result[@intFromEnum(Feature.avx10_1_512)] = .{
.llvm_name = "avx10.1-512",
.description = "Support AVX10.1 up to 512-bit instruction",
.dependencies = featureSet(&[_]Feature{
.avx10_1_256,
.evex512,
}),
};
result[@intFromEnum(Feature.avx10_2_256)] = .{
.llvm_name = "avx10.2-256",
.description = "Support AVX10.2 up to 256-bit instruction",
.dependencies = featureSet(&[_]Feature{
.avx10_1_256,
}),
};
result[@intFromEnum(Feature.avx10_2_512)] = .{
result[@intFromEnum(Feature.avx10_2)] = .{
.llvm_name = "avx10.2-512",
.description = "Support AVX10.2 up to 512-bit instruction",
.dependencies = featureSet(&[_]Feature{
.avx10_1_512,
.avx10_2_256,
.avx10_1,
}),
};
result[@intFromEnum(Feature.avx2)] = .{
@ -434,6 +416,7 @@ pub const all_features = blk: {
.description = "Enable AVX-512 instructions",
.dependencies = featureSet(&[_]Feature{
.avx2,
.evex512,
.f16c,
.fma,
}),
@ -2192,7 +2175,6 @@ pub const cpu = struct {
.cmov,
.cx16,
.ermsb,
.evex512,
.fast_15bytenop,
.fast_gather,
.fast_scalar_fsqrt,
@ -2252,7 +2234,6 @@ pub const cpu = struct {
.cmov,
.cx16,
.ermsb,
.evex512,
.false_deps_popcnt,
.fast_15bytenop,
.fast_gather,
@ -2392,7 +2373,6 @@ pub const cpu = struct {
.cmov,
.cx16,
.ermsb,
.evex512,
.false_deps_popcnt,
.fast_15bytenop,
.fast_gather,
@ -2485,7 +2465,7 @@ pub const cpu = struct {
.amx_movrs,
.amx_tf32,
.amx_transpose,
.avx10_2_512,
.avx10_2,
.avxifma,
.avxneconvert,
.avxvnni,
@ -2601,7 +2581,6 @@ pub const cpu = struct {
.cx16,
.enqcmd,
.ermsb,
.evex512,
.false_deps_getmant,
.false_deps_mulc,
.false_deps_mullq,
@ -2914,7 +2893,6 @@ pub const cpu = struct {
.cx16,
.enqcmd,
.ermsb,
.evex512,
.false_deps_getmant,
.false_deps_mulc,
.false_deps_mullq,
@ -3004,7 +2982,6 @@ pub const cpu = struct {
.cx16,
.enqcmd,
.ermsb,
.evex512,
.false_deps_getmant,
.false_deps_mulc,
.false_deps_mullq,
@ -3167,7 +3144,6 @@ pub const cpu = struct {
.cmov,
.cx16,
.ermsb,
.evex512,
.fast_15bytenop,
.fast_gather,
.fast_scalar_fsqrt,
@ -3231,7 +3207,6 @@ pub const cpu = struct {
.cmov,
.cx16,
.ermsb,
.evex512,
.fast_15bytenop,
.fast_gather,
.fast_scalar_fsqrt,
@ -3394,7 +3369,6 @@ pub const cpu = struct {
.bmi2,
.cmov,
.cx16,
.evex512,
.fast_gather,
.fast_imm16,
.fast_movbe,
@ -3436,7 +3410,6 @@ pub const cpu = struct {
.bmi2,
.cmov,
.cx16,
.evex512,
.fast_gather,
.fast_imm16,
.fast_movbe,
@ -4003,7 +3976,6 @@ pub const cpu = struct {
.cmov,
.cx16,
.ermsb,
.evex512,
.fast_15bytenop,
.fast_gather,
.fast_scalar_fsqrt,
@ -4104,7 +4076,6 @@ pub const cpu = struct {
.cx16,
.enqcmd,
.ermsb,
.evex512,
.false_deps_getmant,
.false_deps_mulc,
.false_deps_mullq,
@ -4277,7 +4248,6 @@ pub const cpu = struct {
.cmov,
.cx16,
.ermsb,
.evex512,
.false_deps_popcnt,
.fast_15bytenop,
.fast_gather,
@ -4390,7 +4360,6 @@ pub const cpu = struct {
.cmov,
.cx16,
.ermsb,
.evex512,
.false_deps_popcnt,
.fast_15bytenop,
.fast_gather,
@ -4485,7 +4454,6 @@ pub const cpu = struct {
.cmov,
.cx16,
.ermsb,
.evex512,
.fast_15bytenop,
.fast_gather,
.fast_scalar_fsqrt,
@ -4706,7 +4674,6 @@ pub const cpu = struct {
.bmi2,
.cmov,
.cx16,
.evex512,
.false_deps_popcnt,
.fast_15bytenop,
.fast_gather,
@ -4943,7 +4910,6 @@ pub const cpu = struct {
.clzero,
.cmov,
.cx16,
.evex512,
.fast_15bytenop,
.fast_bextr,
.fast_dpwssd,
@ -5017,7 +4983,6 @@ pub const cpu = struct {
.clzero,
.cmov,
.cx16,
.evex512,
.fast_15bytenop,
.fast_bextr,
.fast_dpwssd,

View File

@ -558,7 +558,6 @@ fn detectNativeFeatures(cpu: *Target.Cpu, os_tag: Target.Os.Tag) void {
setFeature(cpu, .avxvnniint16, bit(leaf.edx, 10) and has_avx_save);
setFeature(cpu, .prefetchi, bit(leaf.edx, 14));
setFeature(cpu, .usermsr, bit(leaf.edx, 15));
setFeature(cpu, .avx10_1_256, bit(leaf.edx, 19));
// APX
setFeature(cpu, .egpr, bit(leaf.edx, 21));
setFeature(cpu, .push2pop2, bit(leaf.edx, 21));
@ -585,7 +584,6 @@ fn detectNativeFeatures(cpu: *Target.Cpu, os_tag: Target.Os.Tag) void {
.avxvnniint16,
.prefetchi,
.usermsr,
.avx10_1_256,
.egpr,
.push2pop2,
.ppx,
@ -668,7 +666,6 @@ fn detectNativeFeatures(cpu: *Target.Cpu, os_tag: Target.Os.Tag) void {
.avxvnniint16,
.prefetchi,
.usermsr,
.avx10_1_256,
.egpr,
.push2pop2,
.ppx,
@ -724,10 +721,10 @@ fn detectNativeFeatures(cpu: *Target.Cpu, os_tag: Target.Os.Tag) void {
if (max_level >= 0x24) {
leaf = cpuid(0x24, 0);
setFeature(cpu, .avx10_1_512, bit(leaf.ebx, 18));
setFeature(cpu, .avx10_1, bit(leaf.ebx, 18));
} else {
for ([_]Target.x86.Feature{
.avx10_1_512,
.avx10_1,
}) |feat| {
setFeature(cpu, feat, false);
}

View File

@ -1229,6 +1229,27 @@ const targets = [_]ArchTarget{
.llvm_name = "64bit-mode",
.omit = true,
},
// Remove these when LLVM removes AVX10.N-256 support.
.{
.llvm_name = "avx10.1-256",
.flatten = true,
},
.{
.llvm_name = "avx10.2-256",
.flatten = true,
},
.{
.llvm_name = "avx10.1-512",
.zig_name = "avx10_1",
},
.{
.llvm_name = "avx10.2-512",
.zig_name = "avx10_2",
},
.{
.llvm_name = "avx512f",
.extra_deps = &.{"evex512"},
},
.{
.llvm_name = "alderlake",
.extra_deps = &.{ "smap", "smep" },