mirror/wiicompiled
1
0
Fork 0
mirror of https://github.com/patchzyy/wiicompiled synced 2026-10-14 20:07:24 -04:00
Code Issues Packages Projects Releases Wiki Activity

Add experimental x86-64-v2 CPU support for older PCs and Intel Macs (#258)

Browse Source 
* runtime: add experimental x86-64-v2 profile

* fix review findings in network and translator

* build: include mbedtls in macos compile audit

* review fixes

* cmake: restore macOS deployment target floor before project()

* fix: Preserve explicit package versions and gate v3 tests on host support

* fix: Prevent TLS session reuse after failed writes

* fix: Gate cross-compiled profile tests on emulator availability

---------

Co-authored-by: patchzyy <64382339+patchzyy@users.noreply.github.com>
This commit is contained in:
Michael G
2026-10-09 16:46:11 -04:00
committed by GitHub
parent f3df63b529
commit 54c608ef9c
21 changed files with 539 additions and 219 deletions
Download Patch File Download Diff File Expand all files Collapse all files
runtime/include/host_context.h
+1 -1
View File
@@ -4,7 +4,7 @@
// HostContext is the deliberately small boundary between the guest scheduler
// and the host's cooperative-context facility. Windows uses native Fibers;
// Linux and Intel macOS use libco's System V x86-64 backend. macOS AArch64 uses
// Linux uses libco's host backend; Intel macOS uses its System V x86-64 backend. macOS AArch64 uses
// the local assembly backend because it must preserve Darwin's platform-reserved
// x18 register, which libco's AArch64 backend does not save. Its handles are
// only valid on the thread that initialized the scheduler.
runtime/include/isa/ppc_isa_float.h
+20
View File
@@ -607,7 +607,18 @@ inline double PPC_PsMulNoNiInline(double lhs, double rhs)
inline PpcPairVec PpcFmaddPairInline(PpcPairVec multiplicand, PpcPairVec multiplier, PpcPairVec addend)
{
#if defined(__x86_64__)
#if defined(__FMA__)
return _mm_fmadd_ps(multiplicand, multiplier, addend);
#else
// x86-64-v2 has SSE4.2 but no FMA. Preserve the one rounding point per
// lane through the scalar C++ FMA rather than decomposing into mul+add.
const double a = PpcM128ToPsInline(multiplicand);
const double c = PpcM128ToPsInline(multiplier);
const double b = PpcM128ToPsInline(addend);
return PpcPsToM128Inline(PpcPackPairedInline(
PpcAccuratePsMaddLaneNoNiInline<false>(PpcGetPs0Inline(a), PpcGetPs0Inline(c), PpcGetPs0Inline(b)),
PpcAccuratePsMaddLaneNoNiInline<false>(PpcGetPs1Inline(a), PpcGetPs1Inline(c), PpcGetPs1Inline(b))));
#endif
#elif defined(__aarch64__)
const PpcPairVec result = vfma_f32(addend, multiplicand, multiplier);
if (PpcPairNanLaneBitsInline(result) != 0) [[unlikely]]
@@ -619,7 +630,16 @@ inline PpcPairVec PpcFmaddPairInline(PpcPairVec multiplicand, PpcPairVec multipl
inline PpcPairVec PpcFmsubPairInline(PpcPairVec multiplicand, PpcPairVec multiplier, PpcPairVec subtractor)
{
#if defined(__x86_64__)
#if defined(__FMA__)
return _mm_fmsub_ps(multiplicand, multiplier, subtractor);
#else
const double a = PpcM128ToPsInline(multiplicand);
const double c = PpcM128ToPsInline(multiplier);
const double b = PpcM128ToPsInline(subtractor);
return PpcPsToM128Inline(PpcPackPairedInline(
PpcAccuratePsMaddLaneNoNiInline<true>(PpcGetPs0Inline(a), PpcGetPs0Inline(c), PpcGetPs0Inline(b)),
PpcAccuratePsMaddLaneNoNiInline<true>(PpcGetPs1Inline(a), PpcGetPs1Inline(c), PpcGetPs1Inline(b))));
#endif
#elif defined(__aarch64__)
const PpcPairVec result = vfma_f32(vneg_f32(subtractor), multiplicand, multiplier);
if (PpcPairNanLaneBitsInline(result) != 0) [[unlikely]]