fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Some binary xmm ops.

Browse Source
This commit is contained in:
Ben Vanik
2014-01-27 22:18:44 -08:00
parent da36baba8d
commit d67f786af8
2 changed files with 186 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

125
src/alloy/backend/x64/lowering/op_utils.inl
View File

@@ -715,17 +715,15 @@ void IntTernaryOp(X64Emitter& e, Instr*& i, vvv_fn vvv_fn, vvc_fn vvc_fn, vcv_fn
// Since alot of SSE ops can take dest + src, just do that.
// Worst case the callee can dedupe.
typedef void(xmm_v_fn)(X64Emitter& e, Instr& i, const Xmm& dest, const Xmm& src);
template<typename T>
void XmmUnaryOpV(X64Emitter& e, Instr*& i, xmm_v_fn v_fn,
T& dest, T& src1) {
Xmm& dest, Xmm& src1) {
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src1, 0);
v_fn(e, *i, dest, src1);
e.EndOp(dest, src1);
}
template<typename T>
void XmmUnaryOpC(X64Emitter& e, Instr*& i, xmm_v_fn v_fn,
T& dest, Value* src1) {
Xmm& dest, Value* src1) {
e.BeginOp(i->dest, dest, REG_DEST);
if (src1->type == FLOAT32_TYPE) {
e.mov(e.eax, (uint32_t)src1->constant.i32);
@@ -771,6 +769,125 @@ void XmmUnaryOp(X64Emitter& e, Instr*& i, uint32_t flags, xmm_v_fn v_fn) {
}
};
// TODO(benvanik): allow a vvv form for dest = src1 + src2 that new SSE
// ops support.
typedef void(xmm_vv_fn)(X64Emitter& e, Instr& i, const Xmm& dest_src, const Xmm& src);
void XmmBinaryOpVV(X64Emitter& e, Instr*& i, xmm_vv_fn vv_fn,
Xmm& dest, Xmm& src1, Xmm& src2) {
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src1, 0,
i->src2.value, src2, 0);
if (dest == src1) {
vv_fn(e, *i, dest, src2);
} else if (dest == src2) {
if (i->opcode->flags & OPCODE_FLAG_COMMUNATIVE) {
vv_fn(e, *i, dest, src1);
} else {
// Eww.
e.movaps(e.xmm0, src1);
vv_fn(e, *i, e.xmm0, src2);
e.movaps(dest, e.xmm0);
}
} else {
e.movaps(dest, src1);
vv_fn(e, *i, dest, src2);
}
e.EndOp(dest, src1, src2);
}
void XmmBinaryOpVC(X64Emitter& e, Instr*& i, xmm_vv_fn vv_fn,
Xmm& dest, Xmm& src1, Value* src2) {
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src1, 0);
if (i->opcode->flags & OPCODE_FLAG_COMMUNATIVE) {
if (src2->type == FLOAT32_TYPE) {
e.mov(e.eax, (uint32_t)src2->constant.i32);
e.movss(dest, e.eax);
} else if (src2->type == FLOAT64_TYPE) {
e.mov(e.rax, (uint64_t)src2->constant.i64);
e.movsd(dest, e.rax);
} else {
UNIMPLEMENTED_SEQ();
}
vv_fn(e, *i, dest, src1);
} else {
if (dest != src1) {
e.movaps(dest, src1);
}
if (src2->type == FLOAT32_TYPE) {
e.mov(e.eax, (uint32_t)src2->constant.i32);
e.movss(e.xmm0, e.eax);
} else if (src2->type == FLOAT64_TYPE) {
e.mov(e.rax, (uint64_t)src2->constant.i64);
e.movsd(e.xmm0, e.rax);
} else {
UNIMPLEMENTED_SEQ();
}
vv_fn(e, *i, dest, e.xmm0);
}
e.EndOp(dest, src1);
}
void XmmBinaryOpCV(X64Emitter& e, Instr*& i, xmm_vv_fn vv_fn,
Xmm& dest, Value* src1, Xmm& src2) {
e.BeginOp(i->dest, dest, REG_DEST,
i->src2.value, src2, 0);
if (i->opcode->flags & OPCODE_FLAG_COMMUNATIVE) {
if (src1->type == FLOAT32_TYPE) {
e.mov(e.eax, (uint32_t)src1->constant.i32);
e.movss(dest, e.eax);
} else if (src1->type == FLOAT64_TYPE) {
e.mov(e.rax, (uint64_t)src1->constant.i64);
e.movsd(dest, e.rax);
} else {
UNIMPLEMENTED_SEQ();
}
vv_fn(e, *i, dest, src2);
} else {
auto real_src2 = src2;
if (dest == src2) {
e.movaps(e.xmm0, src2);
real_src2 = e.xmm0;
}
if (src1->type == FLOAT32_TYPE) {
e.mov(e.eax, (uint32_t)src1->constant.i32);
e.movss(dest, e.eax);
} else if (src1->type == FLOAT64_TYPE) {
e.mov(e.rax, (uint64_t)src1->constant.i64);
e.movsd(dest, e.rax);
} else {
UNIMPLEMENTED_SEQ();
}
vv_fn(e, *i, dest, real_src2);
}
e.EndOp(dest, src2);
}
void XmmBinaryOp(X64Emitter& e, Instr*& i, uint32_t flags, xmm_vv_fn vv_fn) {
// TODO(benvanik): table lookup. This linear scan is slow.
XEASSERT(i->dest->type == i->src1.value->type);
if (i->Match(SIG_TYPE_IGNORE, SIG_TYPE_F32, SIG_TYPE_F32) ||
i->Match(SIG_TYPE_IGNORE, SIG_TYPE_F64, SIG_TYPE_F64) ||
i->Match(SIG_TYPE_IGNORE, SIG_TYPE_V128, SIG_TYPE_V128)) {
Xmm dest, src1, src2;
XmmBinaryOpVV(e, i, vv_fn, dest, src1, src2);
} else if (i->Match(SIG_TYPE_IGNORE, SIG_TYPE_F32, SIG_TYPE_F32C) ||
i->Match(SIG_TYPE_IGNORE, SIG_TYPE_F64, SIG_TYPE_F64C) ||
i->Match(SIG_TYPE_IGNORE, SIG_TYPE_V128, SIG_TYPE_V128C)) {
Xmm dest, src1;
XmmBinaryOpVC(e, i, vv_fn, dest, src1, i->src2.value);
} else if (i->Match(SIG_TYPE_IGNORE, SIG_TYPE_F32C, SIG_TYPE_F32) ||
i->Match(SIG_TYPE_IGNORE, SIG_TYPE_F64C, SIG_TYPE_F64) ||
i->Match(SIG_TYPE_IGNORE, SIG_TYPE_V128C, SIG_TYPE_V128)) {
Xmm dest, src2;
XmmBinaryOpCV(e, i, vv_fn, dest, i->src1.value, src2);
} else {
ASSERT_INVALID_TYPE();
}
if (flags & ARITHMETIC_SET_CARRY) {
// EFLAGS should have CA set?
// (so long as we don't fuck with it)
// UNIMPLEMENTED_SEQ();
}
};
} // namespace
#endif // ALLOY_BACKEND_X64_X64_LOWERING_OP_UTILS_INL_