ShuriZma
/
xenia
mirror of https://github.com/xenia-project/xenia.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Hacking.

This commit is contained in:
Ben Vanik 2014-01-28 00:19:05 -08:00
parent 5421108b9e
commit 465c3a41dd
4 changed files with 213 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

205
src/alloy/backend/x64/lowering/lowering_sequences.cc
View File

@ -44,6 +44,11 @@ namespace {
// Basically, this identity must hold: // Basically, this identity must hold:
// shuffle(vec, b00011011) -> {x,y,z,w} => {x,y,z,w} // shuffle(vec, b00011011) -> {x,y,z,w} => {x,y,z,w}
// All indices and operations must respect that. // All indices and operations must respect that.
//
// Memory (big endian):
// [00 01 02 03] [04 05 06 07] [08 09 0A 0B] [0C 0D 0E 0F] (x, y, z, w)
// load into xmm register:
// [0F 0E 0D 0C] [0B 0A 09 08] [07 06 05 04] [03 02 01 00] (w, z, y, x)
void Dummy() { void Dummy() {
// //
@ -498,7 +503,63 @@ table->AddSequence(OPCODE_TRUNCATE, [](X64Emitter& e, Instr*& i) {
}); });
table->AddSequence(OPCODE_CONVERT, [](X64Emitter& e, Instr*& i) { table->AddSequence(OPCODE_CONVERT, [](X64Emitter& e, Instr*& i) {
UNIMPLEMENTED_SEQ(); if (i->Match(SIG_TYPE_I32, SIG_TYPE_F32)) {
Reg32 dest;
Xmm src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
// TODO(benvanik): additional checks for saturation/etc? cvtt* (trunc?)
e.cvtss2si(dest, src);
e.EndOp(dest, src);
} else if (i->Match(SIG_TYPE_I32, SIG_TYPE_F64)) {
Reg32 dest;
Xmm src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
// TODO(benvanik): additional checks for saturation/etc? cvtt* (trunc?)
e.cvtsd2ss(e.xmm0, src);
e.cvtss2si(dest, e.xmm0);
e.EndOp(dest, src);
} else if (i->Match(SIG_TYPE_I64, SIG_TYPE_F64)) {
Reg64 dest;
Xmm src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
// TODO(benvanik): additional checks for saturation/etc? cvtt* (trunc?)
e.cvtsd2si(dest, src);
e.EndOp(dest, src);
} else if (i->Match(SIG_TYPE_F32, SIG_TYPE_I32)) {
Xmm dest;
Reg32 src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
// TODO(benvanik): additional checks for saturation/etc?
e.cvtsi2ss(dest, src);
e.EndOp(dest, src);
} else if (i->Match(SIG_TYPE_F32, SIG_TYPE_F64)) {
Xmm dest, src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
// TODO(benvanik): additional checks for saturation/etc?
e.cvtsd2ss(dest, src);
e.EndOp(dest, src);
} else if (i->Match(SIG_TYPE_F64, SIG_TYPE_I64)) {
Xmm dest;
Reg64 src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
// TODO(benvanik): additional checks for saturation/etc?
e.cvtsi2sd(dest, src);
e.EndOp(dest, src);
} else if (i->Match(SIG_TYPE_F64, SIG_TYPE_F32)) {
Xmm dest, src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
e.cvtss2sd(dest, src);
e.EndOp(dest, src);
} else {
UNIMPLEMENTED_SEQ();
}
i = e.Advance(i); i = e.Advance(i);
return true; return true;
}); });
@ -506,9 +567,56 @@ table->AddSequence(OPCODE_CONVERT, [](X64Emitter& e, Instr*& i) {
table->AddSequence(OPCODE_ROUND, [](X64Emitter& e, Instr*& i) { table->AddSequence(OPCODE_ROUND, [](X64Emitter& e, Instr*& i) {
// flags = ROUND_TO_* // flags = ROUND_TO_*
if (IsFloatType(i->dest->type)) { if (IsFloatType(i->dest->type)) {
UNIMPLEMENTED_SEQ(); XmmUnaryOp(e, i, 0, [](X64Emitter& e, Instr& i, const Xmm& dest, const Xmm& src) {
if (i.src1.value->type == FLOAT32_TYPE) {
switch (i.flags) {
case ROUND_TO_ZERO:
e.roundss(dest, src, B00000011);
break;
case ROUND_TO_NEAREST:
e.roundss(dest, src, B00000000);
break;
case ROUND_TO_MINUS_INFINITY:
e.roundss(dest, src, B00000001);
break;
case ROUND_TO_POSITIVE_INFINITY:
e.roundss(dest, src, B00000010);
break;
}
} else {
switch (i.flags) {
case ROUND_TO_ZERO:
e.roundsd(dest, src, B00000011);
break;
case ROUND_TO_NEAREST:
e.roundsd(dest, src, B00000000);
break;
case ROUND_TO_MINUS_INFINITY:
e.roundsd(dest, src, B00000001);
break;
case ROUND_TO_POSITIVE_INFINITY:
e.roundsd(dest, src, B00000010);
break;
}
}
});
} else if (IsVecType(i->dest->type)) { } else if (IsVecType(i->dest->type)) {
UNIMPLEMENTED_SEQ(); XmmUnaryOp(e, i, 0, [](X64Emitter& e, Instr& i, const Xmm& dest, const Xmm& src) {
switch (i.flags) {
case ROUND_TO_ZERO:
e.roundps(dest, src, B00000011);
break;
case ROUND_TO_NEAREST:
e.roundps(dest, src, B00000000);
break;
case ROUND_TO_MINUS_INFINITY:
e.roundps(dest, src, B00000001);
break;
case ROUND_TO_POSITIVE_INFINITY:
e.roundps(dest, src, B00000010);
break;
}
});
} else { } else {
ASSERT_INVALID_TYPE(); ASSERT_INVALID_TYPE();
} }
@ -634,7 +742,7 @@ table->AddSequence(OPCODE_LOAD_CONTEXT, [](X64Emitter& e, Instr*& i) {
e.EndOp(dest); e.EndOp(dest);
#if DTRACE #if DTRACE
e.mov(e.rdx, i->src1.offset); e.mov(e.rdx, i->src1.offset);
e.movaps(e.xmm0, dest); e.lea(e.r8, Stash(e, dest));
CallNative(e, TraceContextLoadV128); CallNative(e, TraceContextLoadV128);
#endif // DTRACE #endif // DTRACE
} else { } else {
@ -755,7 +863,7 @@ table->AddSequence(OPCODE_STORE_CONTEXT, [](X64Emitter& e, Instr*& i) {
e.EndOp(src); e.EndOp(src);
#if DTRACE #if DTRACE
e.mov(e.rdx, i->src1.offset); e.mov(e.rdx, i->src1.offset);
e.movaps(e.xmm0, src); e.lea(e.r8, Stash(e, src));
CallNative(e, TraceContextStoreV128); CallNative(e, TraceContextStoreV128);
#endif // DTRACE #endif // DTRACE
} else if (i->Match(SIG_TYPE_X, SIG_TYPE_IGNORE, SIG_TYPE_V128C)) { } else if (i->Match(SIG_TYPE_X, SIG_TYPE_IGNORE, SIG_TYPE_V128C)) {
@ -765,7 +873,7 @@ table->AddSequence(OPCODE_STORE_CONTEXT, [](X64Emitter& e, Instr*& i) {
MovMem64(e, e.rcx + i->src1.offset + 8, i->src2.value->constant.v128.high); MovMem64(e, e.rcx + i->src1.offset + 8, i->src2.value->constant.v128.high);
#if DTRACE #if DTRACE
e.mov(e.rdx, i->src1.offset); e.mov(e.rdx, i->src1.offset);
e.movups(e.xmm0, e.ptr[e.rcx + i->src1.offset]); e.lea(e.r8, e.ptr[e.rcx + i->src1.offset]);
CallNative(e, TraceContextStoreV128); CallNative(e, TraceContextStoreV128);
#endif // DTRACE #endif // DTRACE
} else { } else {
@ -886,7 +994,7 @@ table->AddSequence(OPCODE_LOAD, [](X64Emitter& e, Instr*& i) {
e.db(0xCC); e.db(0xCC);
#if DTRACE #if DTRACE
e.lea(e.rdx, e.ptr[addr]); e.lea(e.rdx, e.ptr[addr]);
e.movaps(e.xmm0, dest); e.lea(e.r8, Stash(e, dest));
CallNative(e, TraceMemoryLoadV128); CallNative(e, TraceMemoryLoadV128);
#endif // DTRACE #endif // DTRACE
} else { } else {
@ -1063,7 +1171,7 @@ table->AddSequence(OPCODE_STORE, [](X64Emitter& e, Instr*& i) {
e.db(0xCC); e.db(0xCC);
#if DTRACE #if DTRACE
e.lea(e.rdx, e.ptr[addr]); e.lea(e.rdx, e.ptr[addr]);
e.movaps(e.xmm0, src); e.lea(e.r8, Stash(e, src));
CallNative(e, TraceMemoryStoreV128); CallNative(e, TraceMemoryStoreV128);
#endif // DTRACE #endif // DTRACE
} else if (i->Match(SIG_TYPE_X, SIG_TYPE_IGNORE, SIG_TYPE_V128C)) { } else if (i->Match(SIG_TYPE_X, SIG_TYPE_IGNORE, SIG_TYPE_V128C)) {
@ -1073,7 +1181,7 @@ table->AddSequence(OPCODE_STORE, [](X64Emitter& e, Instr*& i) {
MovMem64(e, addr + 8, i->src2.value->constant.v128.high); MovMem64(e, addr + 8, i->src2.value->constant.v128.high);
#if DTRACE #if DTRACE
e.lea(e.rdx, e.ptr[addr]); e.lea(e.rdx, e.ptr[addr]);
e.movups(e.xmm0, e.ptr[addr]); e.lea(e.r8, e.ptr[addr]);
CallNative(e, TraceMemoryStoreV128); CallNative(e, TraceMemoryStoreV128);
#endif // DTRACE #endif // DTRACE
} else { } else {
@ -2107,14 +2215,57 @@ table->AddSequence(OPCODE_INSERT, [](X64Emitter& e, Instr*& i) {
return true; return true;
}); });
// TODO(benvanik): sequence extract/splat:
// v0.i32 = extract v0.v128, 0
// v0.v128 = splat v0.i32
// This can be a single broadcast.
table->AddSequence(OPCODE_EXTRACT, [](X64Emitter& e, Instr*& i) { table->AddSequence(OPCODE_EXTRACT, [](X64Emitter& e, Instr*& i) {
if (IsVecType(i->src1.value->type)) { if (IsVecType(i->src1.value->type)) {
if (i->dest->type == INT8_TYPE) { if (i->dest->type == INT8_TYPE) {
UNIMPLEMENTED_SEQ(); Reg8 dest;
Xmm src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
if (i->src2.value->IsConstant()) {
e.pextrb(dest, src, i->src2.value->constant.i8);
} else {
UNIMPLEMENTED_SEQ();
}
e.EndOp(dest, src);
} else if (i->dest->type == INT16_TYPE) { } else if (i->dest->type == INT16_TYPE) {
UNIMPLEMENTED_SEQ(); Reg16 dest;
Xmm src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
if (i->src2.value->IsConstant()) {
e.pextrw(dest, src, i->src2.value->constant.i8);
} else {
UNIMPLEMENTED_SEQ();
}
e.EndOp(dest, src);
} else if (i->dest->type == INT32_TYPE) { } else if (i->dest->type == INT32_TYPE) {
UNIMPLEMENTED_SEQ(); Reg32 dest;
Xmm src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
if (i->src2.value->IsConstant()) {
e.pextrd(dest, src, i->src2.value->constant.i8);
} else {
UNIMPLEMENTED_SEQ();
}
e.EndOp(dest, src);
} else if (i->dest->type == FLOAT32_TYPE) {
Reg32 dest;
Xmm src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
if (i->src2.value->IsConstant()) {
e.extractps(dest, src, i->src2.value->constant.i8);
} else {
UNIMPLEMENTED_SEQ();
}
e.EndOp(dest, src);
} else { } else {
ASSERT_INVALID_TYPE(); ASSERT_INVALID_TYPE();
} }
@ -2128,13 +2279,35 @@ table->AddSequence(OPCODE_EXTRACT, [](X64Emitter& e, Instr*& i) {
table->AddSequence(OPCODE_SPLAT, [](X64Emitter& e, Instr*& i) { table->AddSequence(OPCODE_SPLAT, [](X64Emitter& e, Instr*& i) {
if (IsVecType(i->dest->type)) { if (IsVecType(i->dest->type)) {
if (i->src1.value->type == INT8_TYPE) { if (i->src1.value->type == INT8_TYPE) {
UNIMPLEMENTED_SEQ(); Xmm dest;
Reg8 src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
e.pinsrb(e.xmm0, src, 0);
e.vpbroadcastb(dest, e.xmm0);
e.EndOp(dest, src);
} else if (i->src1.value->type == INT16_TYPE) { } else if (i->src1.value->type == INT16_TYPE) {
UNIMPLEMENTED_SEQ(); Xmm dest;
Reg16 src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
e.pinsrw(e.xmm0, src, 0);
e.vpbroadcastw(dest, e.xmm0);
e.EndOp(dest, src);
} else if (i->src1.value->type == INT32_TYPE) { } else if (i->src1.value->type == INT32_TYPE) {
UNIMPLEMENTED_SEQ(); Xmm dest;
Reg32 src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
e.pinsrd(e.xmm0, src, 0);
e.vpbroadcastd(dest, e.xmm0);
e.EndOp(dest, src);
} else if (i->src1.value->type == FLOAT32_TYPE) { } else if (i->src1.value->type == FLOAT32_TYPE) {
UNIMPLEMENTED_SEQ(); Xmm dest, src;
e.BeginOp(i->dest, dest, REG_DEST,
i->src1.value, src, 0);
e.vbroadcastss(dest, src);
e.EndOp(dest, src);
} else { } else {
ASSERT_INVALID_TYPE(); ASSERT_INVALID_TYPE();
} }

18
src/alloy/backend/x64/lowering/op_utils.inl
View File

@ -17,6 +17,12 @@ namespace {
#define LIKE_REG(dest, like) Reg(dest.getIdx(), dest.getKind(), like.getBit(), false) #define LIKE_REG(dest, like) Reg(dest.getIdx(), dest.getKind(), like.getBit(), false)
#define NAX_LIKE(like) Reg(e.rax.getIdx(), e.rax.getKind(), like.getBit(), false) #define NAX_LIKE(like) Reg(e.rax.getIdx(), e.rax.getKind(), like.getBit(), false)
Address Stash(X64Emitter& e, const Xmm& r) {
auto addr = e.ptr[e.rsp + 40];
e.movaps(addr, r);
return addr;
}
// Moves a 64bit immediate into memory. // Moves a 64bit immediate into memory.
void MovMem64(X64Emitter& e, RegExp& addr, uint64_t v) { void MovMem64(X64Emitter& e, RegExp& addr, uint64_t v) {
if ((v & ~0x7FFFFFFF) == 0) { if ((v & ~0x7FFFFFFF) == 0) {
@ -869,19 +875,13 @@ void XmmBinaryOpCV(X64Emitter& e, Instr*& i, xmm_vv_fn vv_fn,
} }
void XmmBinaryOp(X64Emitter& e, Instr*& i, uint32_t flags, xmm_vv_fn vv_fn) { void XmmBinaryOp(X64Emitter& e, Instr*& i, uint32_t flags, xmm_vv_fn vv_fn) {
// TODO(benvanik): table lookup. This linear scan is slow. // TODO(benvanik): table lookup. This linear scan is slow.
if (i->Match(SIG_TYPE_IGNORE, SIG_TYPE_F32, SIG_TYPE_F32) || if (!i->src1.value->IsConstant() && !i->src2.value->IsConstant()) {
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; Xmm dest, src1, src2;
XmmBinaryOpVV(e, i, vv_fn, dest, src1, src2); XmmBinaryOpVV(e, i, vv_fn, dest, src1, src2);
} else if (i->Match(SIG_TYPE_IGNORE, SIG_TYPE_F32, SIG_TYPE_F32C) || } else if (!i->src1.value->IsConstant() && i->src2.value->IsConstant()) {
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; Xmm dest, src1;
XmmBinaryOpVC(e, i, vv_fn, dest, src1, i->src2.value); XmmBinaryOpVC(e, i, vv_fn, dest, src1, i->src2.value);
} else if (i->Match(SIG_TYPE_IGNORE, SIG_TYPE_F32C, SIG_TYPE_F32) || } else if (i->src1.value->IsConstant() && !i->src2.value->IsConstant()) {
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; Xmm dest, src2;
XmmBinaryOpCV(e, i, vv_fn, dest, i->src1.value, src2); XmmBinaryOpCV(e, i, vv_fn, dest, i->src1.value, src2);
} else { } else {

5
src/alloy/backend/x64/lowering/tracers.cc
View File

@ -77,9 +77,10 @@ void TraceContextLoadF64(void* raw_context, uint64_t offset, double value) {
} }
void TraceContextLoadV128(void* raw_context, uint64_t offset, __m128 value) { void TraceContextLoadV128(void* raw_context, uint64_t offset, __m128 value) {
auto thread_state = *((ThreadState**)raw_context); auto thread_state = *((ThreadState**)raw_context);
DPRINT("[%e, %e, %e, %e] [%.8X, %.8X, %.8X, %.8X] = ctx v128 +%d\n", offset, DPRINT("[%e, %e, %e, %e] [%.8X, %.8X, %.8X, %.8X] = ctx v128 +%d\n",
value.m128_f32[0], value.m128_f32[1], value.m128_f32[2], value.m128_f32[3], value.m128_f32[0], value.m128_f32[1], value.m128_f32[2], value.m128_f32[3],
value.m128_i32[0], value.m128_i32[1], value.m128_i32[2], value.m128_i32[3]); value.m128_i32[0], value.m128_i32[1], value.m128_i32[2], value.m128_i32[3],
offset);
} }
void TraceContextStoreI8(void* raw_context, uint64_t offset, uint8_t value) { void TraceContextStoreI8(void* raw_context, uint64_t offset, uint8_t value) {

18
src/alloy/compiler/passes/value_reduction_pass.cc
View File

@ -74,34 +74,40 @@ int ValueReductionPass::Run(HIRBuilder* builder) {
OpcodeSignatureType src1_type = GET_OPCODE_SIG_TYPE_SRC1(info->signature); OpcodeSignatureType src1_type = GET_OPCODE_SIG_TYPE_SRC1(info->signature);
OpcodeSignatureType src2_type = GET_OPCODE_SIG_TYPE_SRC2(info->signature); OpcodeSignatureType src2_type = GET_OPCODE_SIG_TYPE_SRC2(info->signature);
OpcodeSignatureType src3_type = GET_OPCODE_SIG_TYPE_SRC3(info->signature); OpcodeSignatureType src3_type = GET_OPCODE_SIG_TYPE_SRC3(info->signature);
if (src1_type == OPCODE_SIG_TYPE_V && !instr->src1.value->IsConstant()) { if (src1_type == OPCODE_SIG_TYPE_V) {
auto v = instr->src1.value; auto v = instr->src1.value;
if (!v->last_use) { if (!v->last_use) {
ComputeLastUse(v); ComputeLastUse(v);
} }
if (v->last_use == instr) { if (v->last_use == instr) {
// Available. // Available.
ordinals.set(v->ordinal, false); if (!instr->src1.value->IsConstant()) {
ordinals.set(v->ordinal, false);
}
} }
} }
if (src2_type == OPCODE_SIG_TYPE_V && !instr->src2.value->IsConstant()) { if (src2_type == OPCODE_SIG_TYPE_V) {
auto v = instr->src2.value; auto v = instr->src2.value;
if (!v->last_use) { if (!v->last_use) {
ComputeLastUse(v); ComputeLastUse(v);
} }
if (v->last_use == instr) { if (v->last_use == instr) {
// Available. // Available.
ordinals.set(v->ordinal, false); if (!instr->src2.value->IsConstant()) {
ordinals.set(v->ordinal, false);
}
} }
} }
if (src3_type == OPCODE_SIG_TYPE_V && !instr->src3.value->IsConstant()) { if (src3_type == OPCODE_SIG_TYPE_V) {
auto v = instr->src3.value; auto v = instr->src3.value;
if (!v->last_use) { if (!v->last_use) {
ComputeLastUse(v); ComputeLastUse(v);
} }
if (v->last_use == instr) { if (v->last_use == instr) {
// Available. // Available.
ordinals.set(v->ordinal, false); if (!instr->src3.value->IsConstant()) {
ordinals.set(v->ordinal, false);
}
} }
} }
if (dest_type == OPCODE_SIG_TYPE_V) { if (dest_type == OPCODE_SIG_TYPE_V) {