iris/kaizen
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

SUPER MARIO 64 BOOTS, works until peach letter

Browse Source
This commit is contained in:
Simone Coco
2022-10-05 11:22:04 +02:00
parent 1f0448fed5
commit 80c8718fef
4 changed files with 328 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/frontend/imgui/Window.hpp
View File

@@ -15,7 +15,7 @@ struct Window {
[[nodiscard]] bool gotClosed(SDL_Event event);
ImFont *uiFont, *codeFont;
u32 windowID;
float volumeL = 0.5, volumeR = 0.5;
float volumeL = 0.001, volumeR = 0.001;
void LoadROM(n64::Core& core, const std::string& path);
private:
bool lockVolume = true;

14
src/n64/core/RSP.hpp
View File

@@ -127,7 +127,7 @@ struct RSP {
VPR vpr[32]{};
s32 gpr[32]{};
VPR vce{};
u16 divIn{}, divOut{};
s16 divIn{}, divOut{};
bool divInLoaded = false;
struct {
@@ -340,11 +340,14 @@ struct RSP {
void lbu(u32 instr);
void lhu(u32 instr);
void lui(u32 instr);
void luv(u32 instr);
void ldv(u32 instr);
void lsv(u32 instr);
void llv(u32 instr);
void lrv(u32 instr);
void lqv(u32 instr);
void ltv(u32 instr);
void lpv(u32 instr);
void j(u32 instr);
void jal(u32 instr);
void jr(u32 instr);
@@ -360,9 +363,12 @@ struct RSP {
void sub(u32 instr);
void sbv(u32 instr);
void sdv(u32 instr);
void stv(u32 instr);
void sqv(u32 instr);
void ssv(u32 instr);
void suv(u32 instr);
void slv(u32 instr);
void spv(u32 instr);
void sllv(u32 instr);
void srlv(u32 instr);
void srav(u32 instr);
@@ -375,8 +381,10 @@ struct RSP {
void sltiu(u32 instr);
void vabs(u32 instr);
void vadd(u32 instr);
void vaddc(u32 instr);
void vand(u32 instr);
void vch(u32 instr);
void vcr(u32 instr);
void vcl(u32 instr);
void vmacf(u32 instr);
void vmadh(u32 instr);
@@ -393,12 +401,16 @@ struct RSP {
void veq(u32 instr);
void vne(u32 instr);
void vge(u32 instr);
void vrcp(u32 instr);
void vrcpl(u32 instr);
void vrsql(u32 instr);
void vrcph(u32 instr);
void vsar(u32 instr);
void vsub(u32 instr);
void vsubc(u32 instr);
void vxor(u32 instr);
void vxnor(u32 instr);
void vnor(u32 instr);
void vzero(u32 instr);
void mfc0(RDP& rdp, u32 instr);
void mtc0(Registers& regs, Mem& mem, u32 instr);

25
src/n64/core/rsp/decode.cpp
View File

@@ -62,6 +62,9 @@ inline void lwc2(RSP& rsp, u32 instr) {
case 0x03: rsp.ldv(instr); break;
case 0x04: rsp.lqv(instr); break;
case 0x05: rsp.lrv(instr); break;
case 0x06: rsp.lpv(instr); break;
case 0x07: rsp.luv(instr); break;
case 0x0B: rsp.ltv(instr); break;
default: util::panic("Unhandled RSP LWC2 {:05b}\n", mask);
}
}
@@ -74,6 +77,9 @@ inline void swc2(RSP& rsp, u32 instr) {
case 0x02: rsp.ssv(instr); break;
case 0x03: rsp.sdv(instr); break;
case 0x04: rsp.sqv(instr); break;
case 0x06: rsp.spv(instr); break;
case 0x07: rsp.suv(instr); break;
case 0x0B: rsp.stv(instr); break;
default: util::panic("Unhandled RSP SWC2 {:05b}\n", mask);
}
}
@@ -88,7 +94,8 @@ inline void cop2(RSP& rsp, u32 instr) {
case 0x02: rsp.cfc2(instr); break;
case 0x04: rsp.mtc2(instr); break;
case 0x06: rsp.ctc2(instr); break;
case 0x1E: rsp.vzero(instr); break;
case 0x10: case 0x1C: case 0x1E:
case 0x1F: case 0x14: break;
default: util::panic("Unhandled RSP COP2 sub ({:05b})\n", mask_sub);
}
break;
@@ -104,20 +111,28 @@ inline void cop2(RSP& rsp, u32 instr) {
case 0x10: rsp.vadd(instr); break;
case 0x11: rsp.vsub(instr); break;
case 0x13: rsp.vabs(instr); break;
case 0x14: rsp.vaddc(instr); break;
case 0x15: rsp.vsubc(instr); break;
case 0x1D: rsp.vsar(instr); break;
case 0x20: rsp.vlt(instr); break;
//case 0x21: rsp.veq(instr); break;
//case 0x22: rsp.vne(instr); break;
case 0x21: rsp.veq(instr); break;
case 0x22: rsp.vne(instr); break;
case 0x23: rsp.vge(instr); break;
case 0x24: rsp.vcl(instr); break;
case 0x25: rsp.vch(instr); break;
case 0x26: rsp.vcr(instr); break;
case 0x27: rsp.vmrg(instr); break;
case 0x28: rsp.vand(instr); break;
case 0x2A: rsp.vnor(instr); break;
case 0x2C: rsp.vxor(instr); break;
case 0x2D: rsp.vxnor(instr); break;
case 0x31: rsp.vrcpl(instr); break;
case 0x32: rsp.vrcph(instr); break;
//case 0x33: rsp.vmov(instr); break;
case 0x35: rsp.vrsql(instr); break;
case 0x32: case 0x36:
rsp.vrcph(instr);
break;
case 0x30: rsp.vrcp(instr); break;
case 0x33: rsp.vmov(instr); break;
default: util::panic("Unhandled RSP COP2 ({:06b})\n", mask);
}
}

302
src/n64/core/rsp/instructions.cpp
View File

@@ -229,6 +229,54 @@ void RSP::lrv(u32 instr) {
}
}
void RSP::lpv(u32 instr) {
int e = E1(instr);
u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 3);
int addrOffset = addr & 7;
addr &= ~7;
for(int elem = 0; elem < 8; elem++) {
int elemOffset = (16 - e + (elem + addrOffset)) & 0xF;
u16 value = ReadByte(addr + elemOffset);
value <<= 8;
vpr[VT(instr)].element[ELEMENT_INDEX(elem)] = value;
}
}
void RSP::luv(u32 instr) {
u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 3);
int e = E1(instr);
int addrOffset = addr & 7;
addr &= ~7;
for (int elem = 0; elem < 8; elem++) {
int elemOffset = (16 - e + (elem + addrOffset)) & 0xF;
u16 value = ReadByte(addr + elemOffset);
value <<= 7;
vpr[VT(instr)].element[ELEMENT_INDEX(elem)] = value;
}
}
void RSP::suv(u32 instr) {
u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 3);
int start = E1(instr);
int end = start + 8;
for (int offset = start; offset < end; offset++) {
if((offset & 15) < 8) {
WriteByte(addr++, vpr[VT(instr)].element[ELEMENT_INDEX(offset & 7)] >> 7);
} else {
WriteByte(addr++, vpr[VT(instr)].byte[BYTE_INDEX((offset & 7) << 1)]);
}
}
}
void RSP::ldv(u32 instr) {
int e = E1(instr);
u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 3);
@@ -338,6 +386,22 @@ void RSP::sqv(u32 instr) {
}
}
void RSP::spv(u32 instr) {
int e = E1(instr);
u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 3);
int start = E1(instr);
int end = start + 8;
for(int offset = start; offset < end; offset++) {
if((offset & 15) < 8) {
WriteByte(addr++, vpr[VT(instr)].byte[BYTE_INDEX((offset & 7) << 1)]);
} else {
WriteByte(addr++, vpr[VT(instr)].byte[ELEMENT_INDEX(offset & 7)] >> 7);
}
}
}
void RSP::sbv(u32 instr) {
int e = E1(instr);
u32 addr = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 3);
@@ -374,6 +438,47 @@ void RSP::slv(u32 instr) {
}
}
void RSP::stv(u32 instr) {
u32 base = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 4);
u32 addrOffset = base & 0x7;
base &= ~0x7;
u8 e = E1(instr) >> 1;
for (int i = 0; i < 8; i++) {
u32 address = base;
u32 offset = (i << 1) + addrOffset;
int reg = (VT(instr) & 0x18) | ((i + e) & 0x7);
u16 val = vpr[reg].element[ELEMENT_INDEX(i & 0x7)];
u16 hi = (val >> 8) & 0xFF;
u16 lo = (val >> 0) & 0xFF;
WriteByte(address + ((offset + 0) & 0xF), hi);
WriteByte(address + ((offset + 1) & 0xF), lo);
}
}
void RSP::ltv(u32 instr) {
u32 base = gpr[BASE(instr)] + SignExt7bit(OFFSET(instr), 4);
base &= ~0x7;
u8 e = E1(instr);
for (int i = 0; i < 8; i++) {
u32 address = base;
u32 offset = (i << 1) + e + (base & 8);
u16 hi = ReadByte(address + (offset & 0xF));
u16 lo = ReadByte(address + ((offset + 1) & 0xF));
int reg = (VT(instr) & 0x18) | ((i + (e >> 1)) & 0x7);
vpr[reg].element[ELEMENT_INDEX(i & 0x7)] = (hi << 8) | lo;
}
}
void RSP::sllv(u32 instr) {
u8 sa = (gpr[RS(instr)]) & 0x1F;
u32 rt = gpr[RT(instr)];
@@ -464,9 +569,7 @@ void RSP::vadd(u32 instr) {
VPR vte = GetVTE(vpr[VT(instr)], E2(instr));
for(int i = 0; i < 8; i++) {
s16 vsE = vs.selement[i];
s16 vteE = vte.selement[i];
s32 result = vsE + vteE + (vco.l.element[i] != 0);
s32 result = vs.selement[i] + vte.selement[i] + (vco.l.element[i] != 0);
acc.l.element[i] = result;
vd.element[i] = clamp_signed(result);
vco.l.element[i] = 0;
@@ -474,6 +577,20 @@ void RSP::vadd(u32 instr) {
}
}
void RSP::vaddc(u32 instr) {
VPR& vs = vpr[VS(instr)];
VPR& vd = vpr[VD(instr)];
VPR vte = GetVTE(vpr[VT(instr)], E2(instr));
for(int i = 0; i < 8; i++) {
s32 result = vs.selement[i] + vte.selement[i];
acc.l.element[i] = result;
vd.element[i] = result;
vco.l.element[i] = (result >> 16) & 1 ? 0xffff : 0;
vco.h.element[i] = 0;
}
}
void RSP::vch(u32 instr) {
int e = E2(instr);
VPR& vs = vpr[VS(instr)];
@@ -491,15 +608,15 @@ void RSP::vch(u32 instr) {
vcc.l.element[i] = result <= 0 ? 0xffff : 0;
vcc.h.element[i] = vteElem < 0 ? 0xffff : 0;
vco.l.element[i] = 0xffff;
vco.h.element[i] = result != 0 && (u16)vsElem != ((u16)vteElem ^ 0xffff);
vce.element[i] = result == -1;
vco.h.element[i] = (result != 0 && (u16)vsElem != ((u16)vteElem ^ 0xffff)) ? 0xffff : 0;
vce.element[i] = result == -1 ? 0xffff : 0;
} else {
s16 result = vsElem - vteElem;
acc.l.selement[i] = (result >= 0) ? vteElem : vsElem;
vcc.l.element[i] = vteElem < 0 ? 0xffff : 0;
vcc.h.element[i] = result >= 0 ? 0xffff : 0;
vco.l.element[i] = 0;
vco.h.element[i] = result != 0 && (u16)vsElem != ((u16)vteElem ^ 0xffff);
vco.h.element[i] = (result != 0 && (u16)vsElem != ((u16)vteElem ^ 0xffff)) ? 0xffff : 0;
vce.element[i] = 0;
}
@@ -507,6 +624,38 @@ void RSP::vch(u32 instr) {
}
}
void RSP::vcr(u32 instr) {
int e = E2(instr);
VPR& vs = vpr[VS(instr)];
VPR& vd = vpr[VD(instr)];
VPR vte = GetVTE(vpr[VT(instr)], e);
for(int i = 0; i < 8; i++) {
u16 vsE = vs.element[i];
u16 vteE = vte.element[i];
bool signDiff = (0x8000 & (vsE ^ vteE)) == 0x8000;
u16 vtAbs = signDiff ? ~vteE : vteE;
bool gte = s16(vteE) <= s16(signDiff ? 0xffff : vsE);
bool lte = (((signDiff ? vsE : 0) + vteE) & 0x8000) == 0x8000;
bool check = signDiff ? lte : gte;
u16 result = check ? vtAbs : vsE;
acc.l.element[i] = result;
vd.element[i] = result;
vcc.h.element[i] = gte ? 0xffff : 0;
vcc.l.element[i] = lte ? 0xffff : 0;
vco.l.element[i] = 0;
vco.h.element[i] = 0;
vce.element[i] = 0;
}
}
void RSP::vcl(u32 instr) {
int e = E2(instr);
VPR& vs = vpr[VS(instr)];
@@ -546,7 +695,35 @@ void RSP::vcl(u32 instr) {
}
void RSP::vmov(u32 instr) {
int e = E2(instr);
VPR& vd = vpr[VD(instr)];
VPR vte = GetVTE(vpr[VT(instr)], e);
u32 se;
switch (e) {
case 0 ... 1:
se = VS(instr) & 7;
break;
case 2 ... 3:
se = (e & 1) | (VS(instr) & 6);
break;
case 4 ... 7:
se = (e & 3) | (VS(instr) & 4);
break;
case 8 ... 15:
se = e & 7;
break;
default:
util::panic("VMOV: This should be unreachable!\n");
}
u8 de = VS(instr) & 7;
vd.element[ELEMENT_INDEX(de)] = vte.element[ELEMENT_INDEX(se)];
for(int i = 0; i < 8; i++) {
acc.l.element[i] = vte.element[i];
}
}
inline bool IsSignExtension(s16 hi, s16 lo) {
@@ -738,11 +915,31 @@ void RSP::vmacf(u32 instr) {
}
void RSP::veq(u32 instr) {
int e = E2(instr);
VPR& vd = vpr[VD(instr)];
VPR& vs = vpr[VS(instr)];
VPR vte = GetVTE(vpr[VT(instr)], e);
for(int i = 0; i < 8; i++) {
vcc.l.element[i] = vco.h.element[i] || (vs.element[i] == vte.element[i]) ? 0xffff : 0;
acc.l.element[i] = vcc.l.element[i] ? vs.element[i] : vte.element[i];
vd.element[i] = acc.l.element[i];
vcc.h.element[i] = vco.h.element[i] = vco.l.element[i] = 0;
}
}
void RSP::vne(u32 instr) {
int e = E2(instr);
VPR& vd = vpr[VD(instr)];
VPR& vs = vpr[VS(instr)];
VPR vte = GetVTE(vpr[VT(instr)], e);
for(int i = 0; i < 8; i++) {
vcc.l.element[i] = vco.h.element[i] || (vs.element[i] != vte.element[i]) ? 0xffff : 0;
acc.l.element[i] = vcc.l.element[i] ? vs.element[i] : vte.element[i];
vd.element[i] = acc.l.element[i];
vcc.h.element[i] = vco.h.element[i] = vco.l.element[i] = 0;
}
}
void RSP::vge(u32 instr) {
@@ -754,7 +951,7 @@ void RSP::vge(u32 instr) {
for(int i = 0; i < 8; i++) {
bool eql = vs.element[i] == vte.element[i];
bool neg = !(vco.h.element[i] & vco.l.element[i]) & eql;
vcc.l.element[i] = (neg | (vs.element[i] > vte.element[i])) ? 0xffff : 0;
vcc.l.element[i] = (neg || (vs.element[i] > vte.element[i])) ? 0xffff : 0;
acc.l.element[i] = vcc.l.element[i] ? vs.element[i] : vte.element[i];
vd.element[i] = acc.l.element[i];
vcc.h.element[i] = vco.h.element[i] = vco.l.element[i] = 0;
@@ -770,7 +967,7 @@ void RSP::vlt(u32 instr) {
for(int i = 0; i < 8; i++) {
bool eql = vs.element[i] == vte.element[i];
bool neg = vco.h.element[i] & vco.l.element[i] & eql;
vcc.l.element[i] = (neg | (vs.element[i] < vte.element[i])) ? 0xffff : 0;
vcc.l.element[i] = (neg || (vs.element[i] < vte.element[i])) ? 0xffff : 0;
acc.l.element[i] = vcc.l.element[i] ? vs.element[i] : vte.element[i];
vd.element[i] = acc.l.element[i];
vcc.h.element[i] = vco.h.element[i] = vco.l.element[i] = 0;
@@ -799,6 +996,29 @@ inline u32 rcp(s32 sinput) {
return result;
}
inline u32 rsq(s32 input) {
if (input == 0) {
return 0x7FFFFFFF;
} else if (input == 0xFFFF8000) {
return 0xFFFF0000;
} else if (input > 0xFFFF8000) {
input--;
}
s32 sinput = input;
s32 mask = sinput >> 31;
input ^= mask;
int shift = __builtin_clz(input) + 1;
int index = (((input << shift) >> 24) | ((shift & 1) << 8));
u32 rom = (((u32)rsq_rom[index]) << 14);
int r_shift = ((32 - shift) >> 1);
u32 result = (0x40000000 | rom) >> r_shift;
return result ^ mask;
}
void RSP::vrcpl(u32 instr) {
VPR& vd = vpr[VD(instr)];
VPR& vt = vpr[VT(instr)];
@@ -825,6 +1045,48 @@ void RSP::vrcpl(u32 instr) {
vd.element[ELEMENT_INDEX(de)] = result;
}
void RSP::vrcp(u32 instr) {
VPR& vd = vpr[VD(instr)];
VPR& vt = vpr[VT(instr)];
VPR vte = GetVTE(vpr[VT(instr)], E2(instr));
int e = E2(instr) & 7;
int de = DE(instr) & 7;
s32 input = vt.selement[ELEMENT_INDEX(e)];
s32 result = rcp(input);
vd.element[ELEMENT_INDEX(de)] = result;
divOut = result >> 16;
divInLoaded = false;
for (int i = 0; i < 8; i++) {
acc.l.element[i] = vte.element[i];
}
}
void RSP::vrsql(u32 instr) {
VPR& vd = vpr[VD(instr)];
VPR& vt = vpr[VT(instr)];
VPR vte = GetVTE(vpr[VT(instr)], E2(instr));
int e = E2(instr) & 7;
int de = DE(instr) & 7;
s32 input;
if(divInLoaded) {
input = (divIn << 16) | vt.element[ELEMENT_INDEX(e)];
} else {
input = vt.selement[ELEMENT_INDEX(e)];
}
u32 result = rsq(input);
divOut = result >> 16;
divInLoaded = false;
for(int i = 0; i < 8; i++) {
acc.l.element[i] = vte.element[i];
}
vd.element[ELEMENT_INDEX(de)] = result;
}
void RSP::vrcph(u32 instr) {
int e = E2(instr);
int de = DE(instr);
@@ -923,6 +1185,18 @@ void RSP::vxor(u32 instr) {
}
}
void RSP::vxnor(u32 instr) {
int e = E2(instr);
VPR& vs = vpr[VS(instr)];
VPR& vd = vpr[VD(instr)];
VPR vte = GetVTE(vpr[VT(instr)], e);
for(int i = 0; i < 8; i++) {
acc.l.element[i] = ~(vte.element[i] ^ vs.element[i]);
vd.element[i] = acc.l.element[i];
}
}
void RSP::vand(u32 instr) {
int e = E2(instr);
VPR& vs = vpr[VS(instr)];
@@ -935,6 +1209,18 @@ void RSP::vand(u32 instr) {
}
}
void RSP::vnor(u32 instr) {
int e = E2(instr);
VPR& vs = vpr[VS(instr)];
VPR& vd = vpr[VD(instr)];
VPR vte = GetVTE(vpr[VT(instr)], e);
for(int i = 0; i < 8; i++) {
acc.l.element[i] = ~(vte.element[i] | vs.element[i]);
vd.element[i] = acc.l.element[i];
}
}
void RSP::vzero(u32 instr) {
VPR& vs = vpr[VS(instr)];
VPR& vd = vpr[VD(instr)];