Skip to content

macOS: detect AVX at runtime instead of assuming every Mac has it #292

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
tobocop2 wants to merge 1 commit into
base: main
Choose a base branch
from
+16 −22
Open

macOS: detect AVX at runtime instead of assuming every Mac has it #292

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 16 additions & 22 deletions Source/JavaScriptCore/assembler/MacroAssemblerX86_64.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,10 +40,8 @@ namespace JSC {

JSC_DECLARE_NOEXCEPT_JIT_OPERATION(ctiMasmProbeTrampoline, void, ());
JSC_ANNOTATE_JIT_OPERATION_PROBE(ctiMasmProbeTrampoline);
#if !OS(DARWIN)
JSC_DECLARE_NOEXCEPT_JIT_OPERATION(ctiMasmProbeTrampolineAVX, void, ());
JSC_ANNOTATE_JIT_OPERATION_PROBE(ctiMasmProbeTrampolineAVX);
#endif

// The following are offsets for Probe::State fields accessed by the ctiMasmProbeTrampoline stub.

Expand Down Expand Up @@ -193,7 +191,6 @@ static_assert(sizeof(Probe::State) == PROBE_SIZE, "Probe::State::size's matches
"vmovaps " STRINGIZE_VALUE_OF(PROBE_CPU_XMM14_VECTOR_OFFSET) "(%rbp), %xmm14" "\n" \
"vmovaps " STRINGIZE_VALUE_OF(PROBE_CPU_XMM15_VECTOR_OFFSET) "(%rbp), %xmm15" "\n"

#if !OS(DARWIN)
#define PROBE_XMM_SAVE_MOVAPS \
"movaps %xmm0, " STRINGIZE_VALUE_OF(PROBE_CPU_XMM0_VECTOR_OFFSET) "(%rsp)" "\n" \
"movaps %xmm1, " STRINGIZE_VALUE_OF(PROBE_CPU_XMM1_VECTOR_OFFSET) "(%rsp)" "\n" \
Expand Down Expand Up @@ -229,7 +226,6 @@ static_assert(sizeof(Probe::State) == PROBE_SIZE, "Probe::State::size's matches
"movaps " STRINGIZE_VALUE_OF(PROBE_CPU_XMM13_VECTOR_OFFSET) "(%rbp), %xmm13" "\n" \
"movaps " STRINGIZE_VALUE_OF(PROBE_CPU_XMM14_VECTOR_OFFSET) "(%rbp), %xmm14" "\n" \
"movaps " STRINGIZE_VALUE_OF(PROBE_CPU_XMM15_VECTOR_OFFSET) "(%rbp), %xmm15" "\n"
#endif // !OS(DARWIN)

#if COMPILER(MSVC)
#define ASM_PREVIOUS_SECTION
Expand Down Expand Up @@ -404,19 +400,13 @@ __asm__( \
ASM_PREVIOUS_SECTION \
);

#if OS(DARWIN)
// On macOS, all x86_64 CPUs support AVX. Use vmovaps unconditionally.
DEFINE_PROBE_TRAMPOLINE(ctiMasmProbeTrampoline, PROBE_XMM_SAVE_VMOVAPS, PROBE_XMM_RESTORE_VMOVAPS,
ctiMasmProbeTrampolineCopyLoop, ctiMasmProbeTrampolineProbeStateIsSafe, ctiMasmProbeTrampolineRestoreRegisters)
#else
// On Linux/Windows, AVX may not be available. Provide two trampolines:
// AVX may not be available. Provide two trampolines:
// - ctiMasmProbeTrampoline: uses movaps (SSE, always available on x86_64)
// - ctiMasmProbeTrampolineAVX: uses vmovaps (requires AVX)
DEFINE_PROBE_TRAMPOLINE(ctiMasmProbeTrampoline, PROBE_XMM_SAVE_MOVAPS, PROBE_XMM_RESTORE_MOVAPS,
ctiMasmProbeTrampolineCopyLoop, ctiMasmProbeTrampolineProbeStateIsSafe, ctiMasmProbeTrampolineRestoreRegisters)
DEFINE_PROBE_TRAMPOLINE(ctiMasmProbeTrampolineAVX, PROBE_XMM_SAVE_VMOVAPS, PROBE_XMM_RESTORE_VMOVAPS,
ctiMasmProbeTrampolineAVXCopyLoop, ctiMasmProbeTrampolineAVXProbeStateIsSafe, ctiMasmProbeTrampolineAVXRestoreRegisters)
#endif

// What code is emitted for the probe?
// ==================================
Expand Down Expand Up @@ -464,14 +454,10 @@ void MacroAssembler::probe(Probe::Function function, void* arg)
push(RegisterID::eax);
push(RegisterID::eax);

#if OS(DARWIN)
move(TrustedImmPtr(reinterpret_cast<void*>(ctiMasmProbeTrampoline)), RegisterID::eax);
#else
if (supportsAVX())
move(TrustedImmPtr(reinterpret_cast<void*>(ctiMasmProbeTrampolineAVX)), RegisterID::eax);
else
move(TrustedImmPtr(reinterpret_cast<void*>(ctiMasmProbeTrampoline)), RegisterID::eax);
#endif

push(RegisterID::edx);
move(TrustedImmPtr(reinterpret_cast<void*>(function)), RegisterID::edx);
Expand All @@ -496,14 +482,12 @@ MacroAssemblerX86_64::CPUID MacroAssemblerX86_64::getCPUIDEx(unsigned level, uns
return result;
}

#if !OS(DARWIN)
static inline uint64_t readXCR0()
{
uint32_t lo, hi;
__asm__("xgetbv" : "=a"(lo), "=d"(hi) : "c"(0));
return static_cast<uint64_t>(lo) | (static_cast<uint64_t>(hi) << 32);
}
#endif

void MacroAssemblerX86_64::collectCPUFeatures()
{
Expand All @@ -516,9 +500,6 @@ void MacroAssemblerX86_64::collectCPUFeatures()
s_sse4_1CheckState = (cpuid[2] & (1 << 19)) ? CPUIDCheckState::Set : CPUIDCheckState::Clear;
s_sse4_2CheckState = (cpuid[2] & (1 << 20)) ? CPUIDCheckState::Set : CPUIDCheckState::Clear;
s_popcntCheckState = (cpuid[2] & (1 << 23)) ? CPUIDCheckState::Set : CPUIDCheckState::Clear;
#if OS(DARWIN)
s_avxCheckState = CPUIDCheckState::Set;
#else
// Per the Intel SDM, AVX is usable only when:
// 1. CPUID.1:ECX.AVX[bit 28] = 1 (CPU supports AVX)
// 2. CPUID.1:ECX.OSXSAVE[bit 27] = 1 (OS has enabled XGETBV)
Expand All @@ -527,13 +508,20 @@ void MacroAssemblerX86_64::collectCPUFeatures()
// (Xen, KVM, Hyper-V, etc.) the AVX feature bit can be passed
// through from the host CPU while XSAVE/YMM state is not enabled
// in the guest, causing every VEX-encoded instruction to #UD.
//
// Darwin needs the same runtime check. Every Mac Apple shipped with
// a supported macOS has AVX (Sandy Bridge, 2011, and later), so
// assuming it held for a long time. It does not hold on an older Mac
// running a current macOS via OpenCore, nor under Rosetta 2, which
// reports no AVX. Assuming AVX there makes the JIT emit VEX-encoded
// instructions the CPU cannot decode, and the process dies with
// SIGILL once any code tiers up.
s_avxCheckState = CPUIDCheckState::Clear;
if ((cpuid[2] & (1 << 28)) && (cpuid[2] & (1 << 27))) {
constexpr uint64_t ymmStateMask = 0x6; // XCR0.SSE | XCR0.AVX
if ((readXCR0() & ymmStateMask) == ymmStateMask)
s_avxCheckState = CPUIDCheckState::Set;
}
Comment thread
coderabbitai[bot] marked this conversation as resolved.
Comment thread
coderabbitai[bot] marked this conversation as resolved.
#endif
}
#if OS(DARWIN)
{
Expand All @@ -542,8 +530,14 @@ void MacroAssemblerX86_64::collectCPUFeatures()
int rc = sysctlbyname("hw.optional.bmi1", &val, &valSize, nullptr, 0);
s_bmi1CheckState = (rc >= 0 && val) ? CPUIDCheckState::Set : CPUIDCheckState::Clear;

// AVX2 is VEX-encoded and therefore also requires the OS to have
// enabled YMM state; gate it on the AVX result computed above, the
// same way the non-Darwin path does. The sysctl and the CPUID/XCR0
// check agree on real hardware, but if they ever disagree, emitting
// AVX2 while AVX is unavailable would #UD.
rc = sysctlbyname("hw.optional.avx2_0", &val, &valSize, nullptr, 0);
s_avx2CheckState = (rc >= 0 && val) ? CPUIDCheckState::Set : CPUIDCheckState::Clear;
bool hasAVX2Feature = rc >= 0 && val;
s_avx2CheckState = (hasAVX2Feature && s_avxCheckState == CPUIDCheckState::Set) ? CPUIDCheckState::Set : CPUIDCheckState::Clear;
}
#else
{
Expand Down