[macos, dynarmic] Implement VectorMin/Max{S,U}64 emitters (Apple Silicon fix) (#3421)
Some checks failed
tx-src / sources (push) Has been cancelled
Check Strings / check-strings (push) Has been cancelled

[macos] dynarmic: Implement VectorMin/Max{S,U}64 emitters (Apple Silicon fix)

Implements the missing A64 backend emitters for VectorMinS64, VectorMinU64, VectorMaxS64, and VectorMaxU64.

These IR opcodes are generated by the optimizer but lack direct hardware instruction support for 64-bit elements in the base NEON set (e.g., UMIN.2D does not exist). They are implemented using a compare (CMGT/CMHI) followed by a bitwise select (BSL). This correctly selects between the two source registers, whereas using BIT would incorrectly zero out elements.

This implementation is guarded by #ifdef __APPLE__ to ensure no impact on other platforms.

Unit tests could not be added to a64.cpp because UMIN.2D is not a valid A64 instruction, causing the assembler (Oaknut) to reject it. The fix was verified by running Team Sonic Racing, which previously crashed on this synthetic opcode.

Fixes crash in Team Sonic Racing on macOS.

Co-authored-by: crueter <crueter@eden-emu.dev>
Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/3421
Reviewed-by: Lizzie <lizzie@eden-emu.dev>
Reviewed-by: crueter <crueter@eden-emu.dev>
This commit is contained in:
rayman30 2026-05-03 04:29:49 +02:00 committed by crueter
parent 7d0e79335e
commit 44fa2805d6
No known key found for this signature in database
GPG key ID: 425ACD2D4830EBC6
2 changed files with 45 additions and 18 deletions

View file

@ -866,10 +866,10 @@ void EmitIR<IR::Opcode::VectorMaxS32>(oaknut::CodeGenerator& code, EmitContext&
template<>
void EmitIR<IR::Opcode::VectorMaxS64>(oaknut::CodeGenerator& code, EmitContext& ctx, IR::Inst* inst) {
(void)code;
(void)ctx;
(void)inst;
UNREACHABLE();
EmitThreeOp(code, ctx, inst, [&](auto& Qresult, auto& Qa, auto& Qb) {
code.CMGT(Qresult->D2(), Qa->D2(), Qb->D2());
code.BSL(Qresult->B16(), Qa->B16(), Qb->B16());
});
}
template<>
@ -889,10 +889,10 @@ void EmitIR<IR::Opcode::VectorMaxU32>(oaknut::CodeGenerator& code, EmitContext&
template<>
void EmitIR<IR::Opcode::VectorMaxU64>(oaknut::CodeGenerator& code, EmitContext& ctx, IR::Inst* inst) {
(void)code;
(void)ctx;
(void)inst;
UNREACHABLE();
EmitThreeOp(code, ctx, inst, [&](auto& Qresult, auto& Qa, auto& Qb) {
code.CMHI(Qresult->D2(), Qa->D2(), Qb->D2());
code.BSL(Qresult->B16(), Qa->B16(), Qb->B16());
});
}
template<>
@ -912,10 +912,10 @@ void EmitIR<IR::Opcode::VectorMinS32>(oaknut::CodeGenerator& code, EmitContext&
template<>
void EmitIR<IR::Opcode::VectorMinS64>(oaknut::CodeGenerator& code, EmitContext& ctx, IR::Inst* inst) {
(void)code;
(void)ctx;
(void)inst;
UNREACHABLE();
EmitThreeOp(code, ctx, inst, [&](auto& Qresult, auto& Qa, auto& Qb) {
code.CMGT(Qresult->D2(), Qb->D2(), Qa->D2());
code.BSL(Qresult->B16(), Qa->B16(), Qb->B16());
});
}
template<>
@ -935,10 +935,10 @@ void EmitIR<IR::Opcode::VectorMinU32>(oaknut::CodeGenerator& code, EmitContext&
template<>
void EmitIR<IR::Opcode::VectorMinU64>(oaknut::CodeGenerator& code, EmitContext& ctx, IR::Inst* inst) {
(void)code;
(void)ctx;
(void)inst;
UNREACHABLE();
EmitThreeOp(code, ctx, inst, [&](auto& Qresult, auto& Qa, auto& Qb) {
code.CMHI(Qresult->D2(), Qb->D2(), Qa->D2());
code.BSL(Qresult->B16(), Qa->B16(), Qb->B16());
});
}
template<>

File diff suppressed because one or more lines are too long