backend/x64: Reduce conversions required for cpsr_nzcv

The guest program often accesses the NZCV flags directly much less
often than we need to use them for jumps and other such uses.

Therefore, we store our flags in cpsr_nzcv in a x64-friendly format.

This allows for a reduction in conditional jump related code.
This commit is contained in:
MerryMage 2020-05-06 22:08:38 +01:00
parent f4922a97f6
commit 8b3bc92bce
7 changed files with 213 additions and 189 deletions

View file

@ -174,115 +174,82 @@ void EmitX64::EmitAddCycles(size_t cycles) {
}
Xbyak::Label EmitX64::EmitCond(IR::Cond cond) {
Xbyak::Label label;
Xbyak::Label pass;
const Xbyak::Reg32 cpsr = eax;
code.mov(cpsr, dword[r15 + code.GetJitStateInfo().offsetof_cpsr_nzcv]);
code.mov(eax, dword[r15 + code.GetJitStateInfo().offsetof_cpsr_nzcv]);
constexpr size_t n_shift = 31;
constexpr size_t z_shift = 30;
constexpr size_t c_shift = 29;
constexpr size_t v_shift = 28;
constexpr u32 n_mask = 1u << n_shift;
constexpr u32 z_mask = 1u << z_shift;
constexpr u32 c_mask = 1u << c_shift;
constexpr u32 v_mask = 1u << v_shift;
// sahf restores SF, ZF, CF
// add al, 0x7F restores OF
switch (cond) {
case IR::Cond::EQ: //z
code.test(cpsr, z_mask);
code.jnz(label);
code.sahf();
code.jz(pass);
break;
case IR::Cond::NE: //!z
code.test(cpsr, z_mask);
code.jz(label);
code.sahf();
code.jnz(pass);
break;
case IR::Cond::CS: //c
code.test(cpsr, c_mask);
code.jnz(label);
code.sahf();
code.jc(pass);
break;
case IR::Cond::CC: //!c
code.test(cpsr, c_mask);
code.jz(label);
code.sahf();
code.jnc(pass);
break;
case IR::Cond::MI: //n
code.test(cpsr, n_mask);
code.jnz(label);
code.sahf();
code.js(pass);
break;
case IR::Cond::PL: //!n
code.test(cpsr, n_mask);
code.jz(label);
code.sahf();
code.jns(pass);
break;
case IR::Cond::VS: //v
code.test(cpsr, v_mask);
code.jnz(label);
code.add(al, 0x7F);
code.jo(pass);
break;
case IR::Cond::VC: //!v
code.test(cpsr, v_mask);
code.jz(label);
code.add(al, 0x7F);
code.jno(pass);
break;
case IR::Cond::HI: { //c & !z
code.and_(cpsr, z_mask | c_mask);
code.cmp(cpsr, c_mask);
code.je(label);
case IR::Cond::HI: //c & !z
code.sahf();
code.cmc();
code.ja(pass);
break;
}
case IR::Cond::LS: { //!c | z
code.and_(cpsr, z_mask | c_mask);
code.cmp(cpsr, c_mask);
code.jne(label);
case IR::Cond::LS: //!c | z
code.sahf();
code.cmc();
code.jna(pass);
break;
}
case IR::Cond::GE: { // n == v
code.and_(cpsr, n_mask | v_mask);
code.jz(label);
code.cmp(cpsr, n_mask | v_mask);
code.je(label);
case IR::Cond::GE: // n == v
code.add(al, 0x7F);
code.sahf();
code.jge(pass);
break;
}
case IR::Cond::LT: { // n != v
Xbyak::Label fail;
code.and_(cpsr, n_mask | v_mask);
code.jz(fail);
code.cmp(cpsr, n_mask | v_mask);
code.jne(label);
code.L(fail);
case IR::Cond::LT: // n != v
code.add(al, 0x7F);
code.sahf();
code.jl(pass);
break;
}
case IR::Cond::GT: { // !z & (n == v)
const Xbyak::Reg32 tmp1 = ebx;
const Xbyak::Reg32 tmp2 = esi;
code.mov(tmp1, cpsr);
code.mov(tmp2, cpsr);
code.shr(tmp1, n_shift);
code.shr(tmp2, v_shift);
code.shr(cpsr, z_shift);
code.xor_(tmp1, tmp2);
code.or_(tmp1, cpsr);
code.test(tmp1, 1);
code.jz(label);
case IR::Cond::GT: // !z & (n == v)
code.add(al, 0x7F);
code.sahf();
code.jg(pass);
break;
}
case IR::Cond::LE: { // z | (n != v)
const Xbyak::Reg32 tmp1 = ebx;
const Xbyak::Reg32 tmp2 = esi;
code.mov(tmp1, cpsr);
code.mov(tmp2, cpsr);
code.shr(tmp1, n_shift);
code.shr(tmp2, v_shift);
code.shr(cpsr, z_shift);
code.xor_(tmp1, tmp2);
code.or_(tmp1, cpsr);
code.test(tmp1, 1);
code.jnz(label);
case IR::Cond::LE: // z | (n != v)
code.add(al, 0x7F);
code.sahf();
code.jle(pass);
break;
}
default:
ASSERT_MSG(false, "Unknown cond {}", static_cast<size_t>(cond));
break;
}
return label;
return pass;
}
EmitX64::BlockDescriptor EmitX64::RegisterBlock(const IR::LocationDescriptor& descriptor, CodePtr entrypoint, size_t size) {