Hello_User/XenonRecomp
1
0
Fork 0
mirror of https://github.com/hedge-dev/XenonRecomp.git synced 2025-11-04 06:47:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Compare commits

base: Hello_User:865319a39cec873370500d26ce775959a4c5e784
Branches Tags
Hello_User:main Hello_User:BBB_MoreInstructions
...
compare: Hello_User:BBB_MoreInstructions
Branches Tags
Hello_User:main Hello_User:BBB_MoreInstructions

8 Commits
865319a39c ... BBB_MoreIn

Author SHA1 Message Date
DeaTh-G
9ff80d8321 make store instructions check for mmio 2025-07-07 20:35:10 +02:00
DeaTh-G
830be1f69a fix vaddsws implementation 2025-07-07 20:35:10 +02:00
DeaTh-G
a5d6382975 add remaining altivec instructions 2025-07-07 20:35:08 +02:00
DeaTh-G
1d452c60a8 add vpkuhus implementation 2025-07-07 20:33:33 +02:00
DeaTh-G
cea0b2fc38 Fix instruction implementations based on unit tests 2025-07-07 20:33:30 +02:00
DeaTh-G
f6193ebe43 add more basic instructions 2025-07-07 20:31:58 +02:00
DeaTh-G
f23d22bc7f Fix indexing on certain instructions 2025-07-07 20:31:57 +02:00
DeaTh-G
847b750786 Add more instructions regarding Bakugan Battle Brawlers 2025-07-07 20:31:57 +02:00
2 changed files with 331 additions and 7 deletions
Expand all files Collapse all files

325
XenonRecomp/recompiler.cpp
View File

@@ -531,6 +531,13 @@ bool Recompiler::Recompile(
println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
break;
case PPC_INST_ADDC:
println("\t{}.ca = {}.u32 > ~{}.u32;", xer(), r(insn.operands[2]), r(insn.operands[1]));
println("\t{}.u64 = {}.u64 + {}.u64;", r(insn.operands[0]), r(insn.operands[1]), r(insn.operands[2]));
if (strchr(insn.opcode->name, '.'))
println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
break;
case PPC_INST_ADDE:
println("\t{}.u8 = ({}.u32 + {}.u32 < {}.u32) | ({}.u32 + {}.u32 + {}.ca < {}.ca);", temp(), r(insn.operands[1]), r(insn.operands[2]), r(insn.operands[1]), r(insn.operands[1]), r(insn.operands[2]), xer(), xer());
println("\t{}.u64 = {}.u64 + {}.u64 + {}.ca;", r(insn.operands[0]), r(insn.operands[1]), r(insn.operands[2]), xer());
@@ -539,6 +546,14 @@ bool Recompiler::Recompile(
println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
break;
case PPC_INST_ADDME:
println("\t{}.u8 = ({}.u32 - 1 < {}.u32) | ({}.u32 - 1 + {}.ca < {}.ca);", temp(), r(insn.operands[1]), r(insn.operands[1]), r(insn.operands[1]), xer(), xer());
println("\t{}.u64 = {}.u64 - 1 + {}.ca;", r(insn.operands[0]), r(insn.operands[1]), xer());
println("\t{}.ca = {}.u8;", xer(), temp());
if (strchr(insn.opcode->name, '.'))
println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
break;
case PPC_INST_ADDI:
print("\t{}.s64 = ", r(insn.operands[0]));
if (insn.operands[1] != 0)
@@ -652,6 +667,14 @@ bool Recompiler::Recompile(
println("\tif ({}.u32 == 0) goto loc_{:X};", ctr(), insn.operands[0]);
break;
case PPC_INST_BDZF:
{
constexpr std::string_view fields[] = { "lt", "gt", "eq", "so" };
println("\t--{}.u64;", ctr());
println("\tif ({}.u32 == 0 && !{}.{}) goto loc_{:X};", ctr(), cr(insn.operands[0] / 4), fields[insn.operands[0] % 4], insn.operands[1]);
break;
}
case PPC_INST_BDZLR:
println("\t--{}.u64;", ctr());
println("\tif ({}.u32 == 0) return;", ctr(), insn.operands[0]);
@@ -663,10 +686,20 @@ bool Recompiler::Recompile(
break;
case PPC_INST_BDNZF:
// NOTE: assuming eq here as a shortcut because all the instructions in the game do that
{
constexpr std::string_view fields[] = { "lt", "gt", "eq", "so" };
println("\t--{}.u64;", ctr());
println("\tif ({}.u32 != 0 && !{}.eq) goto loc_{:X};", ctr(), cr(insn.operands[0] / 4), insn.operands[1]);
println("\tif ({}.u32 != 0 && !{}.{}) goto loc_{:X};", ctr(), cr(insn.operands[0] / 4), fields[insn.operands[0] % 4], insn.operands[1]);
break;
}
case PPC_INST_BDNZT:
{
constexpr std::string_view fields[] = { "lt", "gt", "eq", "so" };
println("\t--{}.u64;", ctr());
println("\tif ({}.u32 != 0 && {}.{}) goto loc_{:X};", ctr(), cr(insn.operands[0] / 4), fields[insn.operands[0] % 4], insn.operands[1]);
break;
}
case PPC_INST_BEQ:
printConditionalBranch(false, "eq");
@@ -796,6 +829,20 @@ bool Recompiler::Recompile(
println("\t{0}.u64 = {1}.u32 == 0 ? 32 : __builtin_clz({1}.u32);", r(insn.operands[0]), r(insn.operands[1]));
break;
case PPC_INST_CROR:
{
constexpr std::string_view fields[] = { "lt", "gt", "eq", "so" };
println("\t{}.{} = {}.{} | {}.{};", cr(insn.operands[0] / 4), fields[insn.operands[0] % 4], cr(insn.operands[1] / 4), fields[insn.operands[1] % 4], cr(insn.operands[2] / 4), fields[insn.operands[2] % 4]);
break;
}
case PPC_INST_CRORC:
{
constexpr std::string_view fields[] = { "lt", "gt", "eq", "so" };
println("\t{}.{} = {}.{} | (~{}.{} & 1);", cr(insn.operands[0] / 4), fields[insn.operands[0] % 4], cr(insn.operands[1] / 4), fields[insn.operands[1] % 4], cr(insn.operands[2] / 4), fields[insn.operands[2] % 4]);
break;
}
case PPC_INST_DB16CYC:
// no op
break;
@@ -808,6 +855,10 @@ bool Recompiler::Recompile(
// no op
break;
case PPC_INST_DCBST:
// no op
break;
case PPC_INST_DCBTST:
// no op
break;
@@ -852,6 +903,12 @@ bool Recompiler::Recompile(
// no op
break;
case PPC_INST_EQV:
println("\t{}.u64 = ~({}.u64 ^ {}.u64);", r(insn.operands[0]), r(insn.operands[1]), r(insn.operands[2]));
if (strchr(insn.opcode->name, '.'))
println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
break;
case PPC_INST_EXTSB:
println("\t{}.s64 = {}.s8;", r(insn.operands[0]), r(insn.operands[1]));
if (strchr(insn.opcode->name, '.'))
@@ -1035,6 +1092,12 @@ bool Recompiler::Recompile(
println("{}.u32);", r(insn.operands[2]));
break;
case PPC_INST_LBZUX:
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\t{}.u64 = PPC_LOAD_U8({});", r(insn.operands[0]), ea());
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
break;
case PPC_INST_LD:
print("\t{}.u64 = PPC_LOAD_U64(", r(insn.operands[0]));
if (insn.operands[2] != 0)
@@ -1063,6 +1126,12 @@ bool Recompiler::Recompile(
println("{}.u32);", r(insn.operands[2]));
break;
case PPC_INST_LDUX:
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\t{}.u64 = PPC_LOAD_U64({});", r(insn.operands[0]), ea());
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
break;
case PPC_INST_LFD:
printSetFlushMode(false);
print("\t{}.u64 = PPC_LOAD_U64(", f(insn.operands[0]));
@@ -1071,6 +1140,13 @@ bool Recompiler::Recompile(
println("{});", int32_t(insn.operands[1]));
break;
case PPC_INST_LFDU:
printSetFlushMode(false);
println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
println("\t{}.u64 = PPC_LOAD_U64({});", r(insn.operands[0]), ea());
println("\t{}.u32 = {};", r(insn.operands[2]), ea());
break;
case PPC_INST_LFDX:
printSetFlushMode(false);
print("\t{}.u64 = PPC_LOAD_U64(", f(insn.operands[0]));
@@ -1079,6 +1155,13 @@ bool Recompiler::Recompile(
println("{}.u32);", r(insn.operands[2]));
break;
case PPC_INST_LFDUX:
printSetFlushMode(false);
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\t{}.u64 = PPC_LOAD_U64({});", r(insn.operands[0]), ea());
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
break;
case PPC_INST_LFS:
printSetFlushMode(false);
print("\t{}.u32 = PPC_LOAD_U32(", temp());
@@ -1088,6 +1171,14 @@ bool Recompiler::Recompile(
println("\t{}.f64 = double({}.f32);", f(insn.operands[0]), temp());
break;
case PPC_INST_LFSU:
printSetFlushMode(false);
println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
println("\t{}.u32 = PPC_LOAD_U32({});", temp(), ea());
println("\t{}.u32 = {};", r(insn.operands[2]), ea());
println("\t{}.f64 = double({}.f32);", f(insn.operands[0]), temp());
break;
case PPC_INST_LFSX:
printSetFlushMode(false);
print("\t{}.u32 = PPC_LOAD_U32(", temp());
@@ -1097,6 +1188,14 @@ bool Recompiler::Recompile(
println("\t{}.f64 = double({}.f32);", f(insn.operands[0]), temp());
break;
case PPC_INST_LFSUX:
printSetFlushMode(false);
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\t{}.u32 = PPC_LOAD_U32({});", temp(), ea());
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
println("\t{}.f64 = double({}.f32);", f(insn.operands[0]), temp());
break;
case PPC_INST_LHA:
print("\t{}.s64 = int16_t(PPC_LOAD_U16(", r(insn.operands[0]));
if (insn.operands[2] != 0)
@@ -1104,6 +1203,12 @@ bool Recompiler::Recompile(
println("{}));", int32_t(insn.operands[1]));
break;
case PPC_INST_LHAU:
print("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
print("\t{}.s64 = int16_t(PPC_LOAD_U16({}));", r(insn.operands[0]), ea());
print("\t{}.u32 = {};", r(insn.operands[2]), ea());
break;
case PPC_INST_LHAX:
print("\t{}.s64 = int16_t(PPC_LOAD_U16(", r(insn.operands[0]));
if (insn.operands[1] != 0)
@@ -1118,6 +1223,12 @@ bool Recompiler::Recompile(
println("{});", int32_t(insn.operands[1]));
break;
case PPC_INST_LHZU:
println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
println("\t{}.u64 = PPC_LOAD_U16({});", r(insn.operands[0]), ea());
println("\t{}.u32 = {};", r(insn.operands[2]), ea());
break;
case PPC_INST_LHZX:
print("\t{}.u64 = PPC_LOAD_U16(", r(insn.operands[0]));
if (insn.operands[1] != 0)
@@ -1125,6 +1236,12 @@ bool Recompiler::Recompile(
println("{}.u32);", r(insn.operands[2]));
break;
case PPC_INST_LHZUX:
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\t{}.u64 = PPC_LOAD_U16({});", r(insn.operands[0]), ea());
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
break;
case PPC_INST_LI:
println("\t{}.s64 = {};", r(insn.operands[0]), int32_t(insn.operands[1]));
break;
@@ -1137,6 +1254,7 @@ bool Recompiler::Recompile(
case PPC_INST_LVEWX128:
case PPC_INST_LVX:
case PPC_INST_LVX128:
case PPC_INST_LVEHX:
// NOTE: for endian swapping, we reverse the whole vector instead of individual elements.
// this is accounted for in every instruction (eg. dp3 sums yzw instead of xyz)
print("\t_mm_store_si128((__m128i*){}.u8, _mm_shuffle_epi8(_mm_load_si128((__m128i*)(base + ((", v(insn.operands[0]));
@@ -1232,6 +1350,12 @@ bool Recompiler::Recompile(
println("{}.u32);", r(insn.operands[2]));
break;
case PPC_INST_LWZUX:
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\t{}.u64 = PPC_LOAD_U32({});", r(insn.operands[0]), ea());
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
break;
case PPC_INST_MFCR:
for (size_t i = 0; i < 32; i++)
{
@@ -1482,7 +1606,7 @@ bool Recompiler::Recompile(
case PPC_INST_STBU:
println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
println("\tPPC_STORE_U8({}, {}.u8);", ea(), r(insn.operands[0]));
println("\t{}{}, {}.u8);", mmioStore() ? "PPC_MM_STORE_U8(" : "PPC_STORE_U8(", ea(), r(insn.operands[0]));
println("\t{}.u32 = {};", r(insn.operands[2]), ea());
break;
@@ -1493,6 +1617,12 @@ bool Recompiler::Recompile(
println("{}.u32, {}.u8);", r(insn.operands[2]), r(insn.operands[0]));
break;
case PPC_INST_STBUX:
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\t{}{}, {}.u8);", mmioStore() ? "PPC_MM_STORE_U8(" : "PPC_STORE_U8(", ea(), r(insn.operands[0]));
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
break;
case PPC_INST_STD:
print("{}", mmioStore() ? "\tPPC_MM_STORE_U64(" : "\tPPC_STORE_U64(");
if (insn.operands[2] != 0)
@@ -1512,7 +1642,7 @@ bool Recompiler::Recompile(
case PPC_INST_STDU:
println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
println("\tPPC_STORE_U64({}, {}.u64);", ea(), r(insn.operands[0]));
println("\t{}{}, {}.u64);", mmioStore() ? "PPC_MM_STORE_U64(" : "PPC_STORE_U64(", ea(), r(insn.operands[0]));
println("\t{}.u32 = {};", r(insn.operands[2]), ea());
break;
@@ -1523,6 +1653,12 @@ bool Recompiler::Recompile(
println("{}.u32, {}.u64);", r(insn.operands[2]), r(insn.operands[0]));
break;
case PPC_INST_STDUX:
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\t{}{}, {}.u64);", mmioStore() ? "PPC_MM_STORE_U64(" : "PPC_STORE_U64(", ea(), r(insn.operands[0]));
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
break;
case PPC_INST_STFD:
printSetFlushMode(false);
print("{}", mmioStore() ? "\tPPC_MM_STORE_U64(" : "\tPPC_STORE_U64(");
@@ -1531,6 +1667,13 @@ bool Recompiler::Recompile(
println("{}, {}.u64);", int32_t(insn.operands[1]), f(insn.operands[0]));
break;
case PPC_INST_STFDU:
printSetFlushMode(false);
println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
println("\t{}{}, {}.u64);", mmioStore() ? "PPC_MM_STORE_U64(" : "PPC_STORE_U64(", ea(), r(insn.operands[0]));
println("\t{}.u32 = {};", r(insn.operands[2]), ea());
break;
case PPC_INST_STFDX:
printSetFlushMode(false);
print("{}", mmioStore() ? "\tPPC_MM_STORE_U64(" : "\tPPC_STORE_U64(");
@@ -1556,6 +1699,14 @@ bool Recompiler::Recompile(
println("{}, {}.u32);", int32_t(insn.operands[1]), temp());
break;
case PPC_INST_STFSU:
printSetFlushMode(false);
println("\t{}.f32 = float({}.f64);", temp(), f(insn.operands[0]));
println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
println("\t{}{}, {}.u32);", mmioStore() ? "PPC_MM_STORE_U32(" : "PPC_STORE_U32(", ea(), temp());
println("\t{}.u32 = {};", r(insn.operands[2]), ea());
break;
case PPC_INST_STFSX:
printSetFlushMode(false);
println("\t{}.f32 = float({}.f64);", temp(), f(insn.operands[0]));
@@ -1565,6 +1716,14 @@ bool Recompiler::Recompile(
println("{}.u32, {}.u32);", r(insn.operands[2]), temp());
break;
case PPC_INST_STFSUX:
printSetFlushMode(false);
println("\t{}.f32 = float({}.f64);", temp(), f(insn.operands[0]));
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\t{}{}, {}.u32);", mmioStore() ? "PPC_MM_STORE_U32(" : "PPC_STORE_U32(", ea(), temp());
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
break;
case PPC_INST_STH:
print("{}", mmioStore() ? "\tPPC_MM_STORE_U16(" : "\tPPC_STORE_U16(");
if (insn.operands[2] != 0)
@@ -1572,6 +1731,18 @@ bool Recompiler::Recompile(
println("{}, {}.u16);", int32_t(insn.operands[1]), r(insn.operands[0]));
break;
case PPC_INST_STHU:
println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
println("\t{}{}, {}.u16);", mmioStore() ? "PPC_MM_STORE_U16(" : "PPC_STORE_U16(", ea(), r(insn.operands[0]));
println("\t{}.u32 = {};", r(insn.operands[2]), ea());
break;
case PPC_INST_STHUX:
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\t{}{}, {}.u16);", mmioStore() ? "PPC_MM_STORE_U16(" : "PPC_STORE_U16(", ea(), r(insn.operands[0]));
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
break;
case PPC_INST_STHBRX:
print("{}", mmioStore() ? "\tPPC_MM_STORE_U16(" : "\tPPC_STORE_U16(");
if (insn.operands[1] != 0)
@@ -1667,13 +1838,13 @@ bool Recompiler::Recompile(
case PPC_INST_STWU:
println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
println("\tPPC_STORE_U32({}, {}.u32);", ea(), r(insn.operands[0]));
println("\t{}{}, {}.u32);", mmioStore() ? "PPC_MM_STORE_U32(" : "PPC_STORE_U32(", ea(), r(insn.operands[0]));
println("\t{}.u32 = {};", r(insn.operands[2]), ea());
break;
case PPC_INST_STWUX:
println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
println("\tPPC_STORE_U32({}, {}.u32);", ea(), r(insn.operands[0]));
println("\t{}{}, {}.u32);", mmioStore() ? "PPC_MM_STORE_U32(" : "PPC_STORE_U32(", ea(), r(insn.operands[0]));
println("\t{}.u32 = {};", r(insn.operands[1]), ea());
break;
@@ -1705,6 +1876,14 @@ bool Recompiler::Recompile(
println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
break;
case PPC_INST_SUBFZE:
println("\t{}.u8 = (~{}.u32 < ~{}.u32) | (~{}.u32 + {}.ca < {}.ca);", temp(), r(insn.operands[1]), r(insn.operands[1]), r(insn.operands[1]), xer(), xer());
println("\t{}.u64 = ~{}.u64 + {}.ca;", r(insn.operands[0]), r(insn.operands[1]), xer());
println("\t{}.ca = {}.u8;", xer(), temp());
if (strchr(insn.opcode->name, '.'))
println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
break;
case PPC_INST_SUBFIC:
println("\t{}.ca = {}.u32 <= {};", xer(), r(insn.operands[1]), insn.operands[2]);
println("\t{}.s64 = {} - {}.s64;", r(insn.operands[0]), int32_t(insn.operands[2]), r(insn.operands[1]));
@@ -1740,10 +1919,23 @@ bool Recompiler::Recompile(
println("\t_mm_store_ps({}.f32, _mm_add_ps(_mm_load_ps({}.f32), _mm_load_ps({}.f32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VADDSBS:
println("\t_mm_store_si128((__m128i*){}.s8, _mm_adds_epi8(_mm_load_si128((__m128i*){}.s8), _mm_load_si128((__m128i*){}.s8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VADDSHS:
println("\t_mm_store_si128((__m128i*){}.s16, _mm_adds_epi16(_mm_load_si128((__m128i*){}.s16), _mm_load_si128((__m128i*){}.s16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VADDSWS:
// TODO: vectorize
for (size_t i = 0; i < 4; i++)
{
println("\t{}.s64 = int64_t({}.s32[{}]) + int64_t({}.s32[{}]);", temp(), v(insn.operands[1]), i, v(insn.operands[2]), i);
println("\t{}.s32[{}] = {}.s64 > INT_MAX ? INT_MAX : {}.s64 < INT_MIN ? INT_MIN : {}.s64;", v(insn.operands[0]), i, temp(), temp(), temp());
}
break;
case PPC_INST_VADDUBM:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_add_epi8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
@@ -1785,6 +1977,10 @@ bool Recompiler::Recompile(
println("\t_mm_store_si128((__m128i*){}.u8, _mm_avg_epu8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VAVGUH:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_avg_epu16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VCTSXS:
case PPC_INST_VCFPSXWS128:
printSetFlushMode(true);
@@ -1795,6 +1991,16 @@ bool Recompiler::Recompile(
println("_mm_load_ps({}.f32)));", v(insn.operands[1]));
break;
case PPC_INST_VCTUXS:
case PPC_INST_VCFPUXWS128:
printSetFlushMode(true);
print("\t_mm_store_si128((__m128i*){}.u32, _mm_vctuxs(", v(insn.operands[0]));
if (insn.operands[2] != 0)
println("_mm_mul_ps(_mm_load_ps({}.f32), _mm_set1_ps({}))));", v(insn.operands[1]), 1u << insn.operands[2]);
else
println("_mm_load_ps({}.f32)));", v(insn.operands[1]));
break;
case PPC_INST_VCFSX:
case PPC_INST_VCSXWFP128:
{
@@ -1848,6 +2054,12 @@ bool Recompiler::Recompile(
println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.u8), 0xFFFF);", cr(6), v(insn.operands[0]));
break;
case PPC_INST_VCMPEQUH:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_cmpeq_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
if (strchr(insn.opcode->name, '.'))
println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.u16), 0xFFFF);", cr(6), v(insn.operands[0]));
break;
case PPC_INST_VCMPEQUW:
case PPC_INST_VCMPEQUW128:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_cmpeq_epi32(_mm_load_si128((__m128i*){}.u32), _mm_load_si128((__m128i*){}.u32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
@@ -1873,10 +2085,26 @@ bool Recompiler::Recompile(
case PPC_INST_VCMPGTUB:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_cmpgt_epu8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
if (strchr(insn.opcode->name, '.'))
println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.u8), 0xFFFF);", cr(6), v(insn.operands[0]));
break;
case PPC_INST_VCMPGTUH:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_cmpgt_epu16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
if (strchr(insn.opcode->name, '.'))
println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.u16), 0xFFFF);", cr(6), v(insn.operands[0]));
break;
case PPC_INST_VCMPGTSH:
println("\t_mm_store_si128((__m128i*){}.s8, _mm_cmpgt_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
if (strchr(insn.opcode->name, '.'))
println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.s16), 0xFFFF);", cr(6), v(insn.operands[0]));
break;
case PPC_INST_VCMPGTSW:
println("\t_mm_store_si128((__m128i*){}.s8, _mm_cmpgt_epi32(_mm_load_si128((__m128i*){}.u32), _mm_load_si128((__m128i*){}.u32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
if (strchr(insn.opcode->name, '.'))
println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.s32), 0xFFFF);", cr(6), v(insn.operands[0]));
break;
case PPC_INST_VEXPTEFP:
@@ -1908,10 +2136,18 @@ bool Recompiler::Recompile(
println("\t_mm_store_ps({}.f32, _mm_max_ps(_mm_load_ps({}.f32), _mm_load_ps({}.f32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VMAXSH:
println("\t_mm_store_si128((__m128i*){}.u16, _mm_max_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VMAXSW:
println("\t_mm_store_si128((__m128i*){}.u32, _mm_max_epi32(_mm_load_si128((__m128i*){}.u32), _mm_load_si128((__m128i*){}.u32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VMINSH:
println("\t_mm_store_si128((__m128i*){}.u16, _mm_min_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VMINFP:
case PPC_INST_VMINFP128:
printSetFlushMode(true);
@@ -2042,11 +2278,36 @@ bool Recompiler::Recompile(
}
break;
case PPC_INST_VPKSHSS:
case PPC_INST_VPKSHSS128:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_packs_epi16(_mm_load_si128((__m128i*){}.s16), _mm_load_si128((__m128i*){}.s16)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
break;
case PPC_INST_VPKSWSS:
case PPC_INST_VPKSWSS128:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_packs_epi32(_mm_load_si128((__m128i*){}.s32), _mm_load_si128((__m128i*){}.s32)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
break;
case PPC_INST_VPKSHUS:
case PPC_INST_VPKSHUS128:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_packus_epi16(_mm_load_si128((__m128i*){}.s16), _mm_load_si128((__m128i*){}.s16)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
break;
case PPC_INST_VPKSWUS:
case PPC_INST_VPKSWUS128:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_packus_epi32(_mm_load_si128((__m128i*){}.s32), _mm_load_si128((__m128i*){}.s32)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
break;
case PPC_INST_VPKUHUS:
case PPC_INST_VPKUHUS128:
for (size_t i = 0; i < 8; i++)
{
println("\t{0}.u8[{1}] = {2}.u16[{1}] > UCHAR_MAX ? UCHAR_MAX : {2}.u16[{1}];", vTemp(), i, v(insn.operands[2]));
println("\t{0}.u8[{1}] = {2}.u16[{3}] > UCHAR_MAX ? UCHAR_MAX : {2}.u16[{3}];", vTemp(), i + 8, v(insn.operands[1]), i);
}
println("{} = {};", v(insn.operands[0]), vTemp());
break;
case PPC_INST_VREFP:
case PPC_INST_VREFP128:
// TODO: see if we can use rcp safely
@@ -2079,6 +2340,14 @@ bool Recompiler::Recompile(
break;
}
case PPC_INST_VRLH:
for (size_t i = 0; i < 8; i++)
{
println("\t{0}.u16[{1}] = ({2}.u16[{1}] << ({3}.u16[{1}] & 0xF)) | ({2}.u16[{1}] >> (16 - ({3}.u16[{1}] & 0xF)));", vTemp(), i, v(insn.operands[1]), v(insn.operands[2]));
}
println("{} = {};", v(insn.operands[0]), vTemp());
break;
case PPC_INST_VRSQRTEFP:
case PPC_INST_VRSQRTEFP128:
// TODO: see if we can use rsqrt safely
@@ -2088,6 +2357,7 @@ bool Recompiler::Recompile(
break;
case PPC_INST_VSEL:
case PPC_INST_VSEL128:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_or_si128(_mm_andnot_si128(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)), _mm_and_si128(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8))));", v(insn.operands[0]), v(insn.operands[3]), v(insn.operands[1]), v(insn.operands[3]), v(insn.operands[2]));
break;
@@ -2097,6 +2367,12 @@ bool Recompiler::Recompile(
println("\t{}.u8[{}] = {}.u8[{}] << ({}.u8[{}] & 0x7);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i);
break;
case PPC_INST_VSLH:
// TODO: vectorize
for (size_t i = 0; i < 8; i++)
println("\t{}.u16[{}] = {}.u16[{}] << ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 2);
break;
case PPC_INST_VSLDOI:
case PPC_INST_VSLDOI128:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_alignr_epi8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8), {}));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]), 16 - insn.operands[3]);
@@ -2130,6 +2406,10 @@ bool Recompiler::Recompile(
println("\t_mm_store_si128((__m128i*){}.u8, _mm_set1_epi8(char(0x{:X})));", v(insn.operands[0]), insn.operands[1]);
break;
case PPC_INST_VSPLTISH:
println("\t_mm_store_si128((__m128i*){}.u16, _mm_set1_epi16(int(0x{:X})));", v(insn.operands[0]), insn.operands[1]);
break;
case PPC_INST_VSPLTISW:
case PPC_INST_VSPLTISW128:
println("\t_mm_store_si128((__m128i*){}.u32, _mm_set1_epi32(int(0x{:X})));", v(insn.operands[0]), insn.operands[1]);
@@ -2149,6 +2429,18 @@ bool Recompiler::Recompile(
println("\t_mm_store_si128((__m128i*){}.u8, _mm_vsr(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VSRAB:
// TODO: vectorize, ensure endianness is correct
for (size_t i = 0; i < 16; i++)
println("\t{}.s8[{}] = {}.s8[{}] >> ({}.u8[{}] & 0x7);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i);
break;
case PPC_INST_VSRAH:
// TODO: vectorize, ensure endianness is correct
for (size_t i = 0; i < 8; i++)
println("\t{}.s16[{}] = {}.s16[{}] >> ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 2);
break;
case PPC_INST_VSRAW:
case PPC_INST_VSRAW128:
// TODO: vectorize, ensure endianness is correct
@@ -2156,6 +2448,12 @@ bool Recompiler::Recompile(
println("\t{}.s32[{}] = {}.s32[{}] >> ({}.u8[{}] & 0x1F);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 4);
break;
case PPC_INST_VSRH:
// TODO: vectorize, ensure endianness is correct
for (size_t i = 0; i < 8; i++)
println("\t{}.u16[{}] = {}.u16[{}] >> ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 2);
break;
case PPC_INST_VSRW:
case PPC_INST_VSRW128:
// TODO: vectorize, ensure endianness is correct
@@ -2169,6 +2467,15 @@ bool Recompiler::Recompile(
println("\t_mm_store_ps({}.f32, _mm_sub_ps(_mm_load_ps({}.f32), _mm_load_ps({}.f32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VSUBSHS:
// TODO: vectorize
for (size_t i = 0; i < 8; i++)
{
println("\t{}.s64 = int64_t({}.s16[{}]) - int64_t({}.s16[{}]);", temp(), v(insn.operands[1]), i, v(insn.operands[2]), i);
println("\t{}.s16[{}] = {}.s64 > SHRT_MAX ? SHRT_MAX : {}.s64 < SHRT_MIN ? SHRT_MIN : {}.s64;", v(insn.operands[0]), i, temp(), temp(), temp());
}
break;
case PPC_INST_VSUBSWS:
// TODO: vectorize
for (size_t i = 0; i < 4; i++)
@@ -2182,8 +2489,12 @@ bool Recompiler::Recompile(
println("\t_mm_store_si128((__m128i*){}.u8, _mm_subs_epu8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VSUBUBM:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_sub_epi8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VSUBUHM:
println("\t_mm_store_si128((__m128i*){}.u8, _mm_sub_epi16(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
println("\t_mm_store_si128((__m128i*){}.u8, _mm_sub_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
break;
case PPC_INST_VUPKD3D128:

13
XenonUtils/ppc_context.h
View File

@@ -645,6 +645,19 @@ inline __m128i _mm_vctsxs(__m128 src1)
return _mm_andnot_si128(_mm_castps_si128(xmm2), _mm_castps_si128(dest));
}
inline __m128i _mm_vctuxs(__m128 src1)
{
__m128 xmm0 = _mm_max_ps(src1, _mm_set1_epi32(0));
__m128 xmm1 = _mm_cmpge_ps(xmm0, _mm_set1_ps((float)0x80000000));
__m128 xmm2 = _mm_sub_ps(xmm0, _mm_set1_ps((float)0x80000000));
xmm0 = _mm_blendv_ps(xmm0, xmm2, xmm1);
__m128i dest = _mm_cvttps_epi32(xmm0);
xmm0 = _mm_cmpeq_epi32(dest, _mm_set1_epi32(INT_MIN));
xmm1 = _mm_and_si128(xmm1, _mm_set1_epi32(INT_MIN));
dest = _mm_add_epi32(dest, xmm1);
return _mm_or_si128(dest, xmm0);
}
inline __m128i _mm_vsr(__m128i a, __m128i b)
{
b = _mm_srli_epi64(_mm_slli_epi64(b, 61), 61);