LATX: expose implemented F16C and LZCNT features

target/i386/cpu.c

@@ -1915,7 +1915,8 @@ static X86CPUDefinition builtin_x86_defs[] = {
             CPUID_PSE36,
         .features[FEAT_1_ECX] =
 #ifdef CONFIG_LATX_AVX_OPT
-            CPUID_EXT_AVX | CPUID_EXT_FMA | CPUID_EXT_XSAVE |
+            CPUID_EXT_AVX | CPUID_EXT_FMA | CPUID_EXT_F16C |
+            CPUID_EXT_XSAVE |
 #endif/*CONFIG_LATX_AVX_OPT*/
 #ifdef CONFIG_LATX_SSSE3_SSE4
             CPUID_EXT_AES | CPUID_EXT_POPCNT | CPUID_EXT_SSE42 |
@@ -1926,7 +1927,7 @@ static X86CPUDefinition builtin_x86_defs[] = {
         .features[FEAT_8000_0001_EDX] =
             CPUID_EXT2_LM | CPUID_EXT2_SYSCALL | CPUID_EXT2_NX,
         .features[FEAT_8000_0001_ECX] =
-            CPUID_EXT3_LAHF_LM | CPUID_EXT3_SVM,
+            CPUID_EXT3_LAHF_LM | CPUID_EXT3_SVM | CPUID_EXT3_ABM,
 #ifdef CONFIG_LATX_AVX_OPT
         .features[FEAT_7_0_EBX] =
         CPUID_7_0_EBX_HLE |
@@ -2065,7 +2066,8 @@ static X86CPUDefinition builtin_x86_defs[] = {
             PPRO_FEATURES,
         .features[FEAT_1_ECX] =
 #ifdef CONFIG_LATX_AVX_OPT
-            CPUID_EXT_AVX | CPUID_EXT_FMA | CPUID_EXT_XSAVE |
+            CPUID_EXT_AVX | CPUID_EXT_FMA | CPUID_EXT_F16C |
+            CPUID_EXT_XSAVE |
 #endif/*CONFIG_LATX_AVX_OPT*/
 #ifdef CONFIG_LATX_SSSE3_SSE4
             CPUID_EXT_AES | CPUID_EXT_POPCNT | CPUID_EXT_SSE42 |
@@ -2074,7 +2076,7 @@ static X86CPUDefinition builtin_x86_defs[] = {
 #endif
             CPUID_EXT_SSE3,
         .features[FEAT_8000_0001_ECX] =
-            CPUID_EXT3_LAHF_LM | CPUID_EXT3_SVM,
+            CPUID_EXT3_LAHF_LM | CPUID_EXT3_SVM | CPUID_EXT3_ABM,
 #ifdef CONFIG_LATX_AVX_OPT
         .features[FEAT_7_0_EBX] =
         CPUID_7_0_EBX_HLE |
@@ -7581,6 +7583,7 @@ static void x86_cpu_register_types(void)
     if(!option_avx_cpuid){
         uint64_t feat_1_ecx_mask = ~(CPUID_EXT_AVX |
                 CPUID_EXT_FMA |
+                CPUID_EXT_F16C |
                 CPUID_EXT_XSAVE);
         uint64_t feat_7_0_ebx_mask = ~(CPUID_7_0_EBX_HLE |
                 CPUID_7_0_EBX_BMI1 |

tests/integration/avx-cpu-features.S

@@ -0,0 +1,84 @@
+.intel_syntax noprefix
+.section .text
+.global _start
+
+_start:
+    mov eax, 1
+    cpuid
+    bt ecx, 29
+    jnc .Lfail_f16c
+
+    mov eax, 0x80000001
+    cpuid
+    bt ecx, 5
+    jnc .Lfail_lzcnt
+
+    xor eax, eax
+    lzcnt eax, eax
+    jnc .Lfail_lzcnt_flags
+    jz .Lfail_lzcnt_flags
+    cmp eax, 32
+    jne .Lfail_lzcnt_value
+
+    mov eax, 0x80000000
+    lzcnt eax, eax
+    jc .Lfail_lzcnt_flags
+    jnz .Lfail_lzcnt_flags
+    test eax, eax
+    jne .Lfail_lzcnt_value
+
+    mov eax, 0x100
+    lzcnt ax, ax
+    cmp ax, 7
+    jne .Lfail_lzcnt_value
+
+    mov rax, 1
+    lzcnt rax, rax
+    cmp rax, 63
+    jne .Lfail_lzcnt_value
+
+    vmovq xmm0, qword ptr [rip + .Lhalf]
+    vcvtph2ps xmm1, xmm0
+    vmovdqu xmm2, xmmword ptr [rip + .Lfloat]
+    vpcmpeqd xmm1, xmm1, xmm2
+    vpmovmskb eax, xmm1
+    cmp eax, 0xffff
+    jne .Lfail_f16c_value
+
+    vcvtps2ph xmm3, xmm2, 0
+    vmovq rax, xmm3
+    cmp rax, qword ptr [rip + .Lhalf]
+    jne .Lfail_f16c_roundtrip
+
+    xor edi, edi
+    jmp .Lexit
+
+.Lfail_f16c:
+    mov edi, 1
+    jmp .Lexit
+.Lfail_lzcnt:
+    mov edi, 2
+    jmp .Lexit
+.Lfail_lzcnt_value:
+    mov edi, 3
+    jmp .Lexit
+.Lfail_f16c_value:
+    mov edi, 4
+    jmp .Lexit
+.Lfail_f16c_roundtrip:
+    mov edi, 5
+    jmp .Lexit
+.Lfail_lzcnt_flags:
+    mov edi, 6
+.Lexit:
+    mov eax, 60
+    syscall
+
+.section .rodata
+.p2align 4
+.Lhalf:
+    .short 0x3c00, 0xc000, 0x3800, 0x0000
+.Lfloat:
+    .long 0x3f800000, 0xc0000000, 0x3f000000, 0x00000000
+
+.section .note.GNU-stack,"",@progbits

tests/integration/meson.build

@@ -1,5 +1,7 @@
 latx_integration_tests = []
 
+subdir('registrations/cpu')
+
 if host_machine.cpu_family() == 'loongarch64'
   subdir('registrations/x11-kzt')
   subdir('registrations/sandbox')

tests/integration/registrations/cpu/meson.build

@@ -0,0 +1,12 @@
+if 'x86_64-linux-user' in target_dirs and \
+   'CONFIG_LATX_AVX_OPT' in config_host
+  latx_integration_tests += [{
+    'name': 'test-avx-cpu-features',
+    'runner': find_program('../../test-avx-cpu-features.sh'),
+    'args': [
+      emulators['latx-x86_64'],
+      files('../../avx-cpu-features.S'),
+    ],
+    'timeout': 120,
+  }]
+endif

tests/integration/test-avx-cpu-features.sh

@@ -0,0 +1,43 @@
+#!/bin/sh
+set -eu
+
+emulator=$1
+source_file=$2
+workdir=$(mktemp -d)
+trap 'rm -rf "$workdir"' EXIT HUP INT TERM
+
+if command -v clang-19 >/dev/null 2>&1; then
+    clang=clang-19
+elif command -v clang >/dev/null 2>&1; then
+    clang=clang
+else
+    echo "SKIP: clang is required to build the x86_64 guest"
+    exit 77
+fi
+
+"$clang" --target=x86_64-linux-gnu -fuse-ld=lld -nostdlib -static \
+    -Wl,--build-id=none "$source_file" -o "$workdir/avx-cpu-features"
+
+LATX_AOT=0 LATX_TU=0 "$emulator" "$workdir/avx-cpu-features"
+
+aot_home="$workdir/aot"
+mkdir -p "$aot_home"
+HOME="$aot_home" LATX_AOT=1 LATX_TU=1 \
+    "$emulator" "$workdir/avx-cpu-features"
+
+aot_file=
+for _ in $(seq 1 100); do
+    aot_file=$(find "$aot_home/.cache/latx" -type f -name '*.aot2' \
+        -size +0c -print -quit 2>/dev/null || true)
+    [ -n "$aot_file" ] && break
+    sleep 0.1
+done
+if [ -z "$aot_file" ]; then
+    echo "FAIL: no non-empty AOT file generated" >&2
+    exit 1
+fi
+
+HOME="$aot_home" LATX_AOT=1 LATX_TU=1 \
+    "$emulator" "$workdir/avx-cpu-features"
+
+echo "PASS: F16C/LZCNT CPUID and instructions JIT/cold-AOT/hot-AOT"