optimizations, bugfixes, uae works (but with timing glitches?)
[picodrive.git] / cpu / Cyclone / Cyclone.txt
CommitLineData
cc68a136 1\r
2 _____ __ \r
3 / ___/__ __ ____ / /___ ___ ___ ___________________ \r
4 / /__ / // // __// // _ \ / _ \/ -_) ___________________ \r
5 \___/ \_, / \__//_/ \___//_//_/\__/ ___________________ \r
6 /___/ \r
7 ___________________ ____ ___ ___ ___ ___ \r
8 ___________________ / __// _ \ / _ \ / _ \ / _ \ \r
9 ___________________ / _ \/ _ // // // // // // / \r
10 \___/\___/ \___/ \___/ \___/ \r
11 \r
12___________________________________________________________________________\r
13\r
14 Cyclone 68000 (c) Copyright 2004 Dave. Free for non-commercial use\r
15\r
16 Homepage: http://www.finalburn.com/\r
03c5768c 17 Dave's e-mail: emudave(atsymbol)googlemail.com\r
cc68a136 18 Replace (atsymbol) with @\r
19\r
03c5768c 20 Additional coding and bugfixes done by notaz, 2005-2007\r
21 Homepage: http://mif.vu.lt/~grig2790/Cyclone/ , http://notaz.gp2x.de\r
cc68a136 22 e-mail: notasas(atsymbol)gmail.com\r
23___________________________________________________________________________\r
24\r
25\r
26What is it?\r
27-----------\r
28\r
29Cyclone 68000 is an emulator for the 68000 microprocessor, written in ARM 32-bit assembly.\r
30It is aimed at chips such as ARM7 and ARM9 cores, StrongARM and XScale, to interpret 68000\r
31code as fast as possible.\r
32\r
33Flags are mapped onto ARM flags whenever possible, which speeds up the processing of opcode.\r
34\r
35\r
36What's New\r
37----------\r
03c5768c 38v0.0087 notaz\r
39 - Reduced amount of code in opcode handlers by ~23% by doing the following:\r
40 - Removed duplicate opcode handlers\r
41 - Optimized code to use less ARM instructions\r
42 - Merged some duplicate handler endings\r
43 + Cyclone now does better job avoiding pipeline interlocks.\r
44 + Replaced incorrect handler of DBT with proper one.\r
1c88b865 45 + Changed "MOVEA (An)+ An" behaviour.\r
46 + Fixed flag behaviour of ROXR, ASL, LSR and NBCD in certain situations.\r
47 Hopefully got them right now.\r
ee5e024c 48 + Cyclone no longer sets most significant bits while pushing PC to stack.\r
49 Amiga Kickstart depends on this.\r
03c5768c 50 + Additional functionality added for MAME and other ports (see config.h).\r
51\r
cc68a136 52v0.0086 notaz\r
53 + Cyclone now can be customized to better suit your project, see config.h .\r
54 + Added an option to compress the jumptable at compile-time. Must call CycloneInit()\r
55 at runtime to decompress it if enabled (see config.h).\r
56 + Added missing CHK opcode handler (used by SeaQuest DSV).\r
57 + Added missing TAS opcode handler (Gargoyles,Bubba N Stix,...). As in real genesis,\r
58 memory write-back phase is ignored (but can be enabled in config.h if needed).\r
59 + Added missing NBCD and TRAPV opcode handlers.\r
60 + Added missing addressing mode for CMP/EOR.\r
61 + Added some minor optimizations.\r
62 - Removed 216 handlers for 2927 opcodes which were generated for invalid addressing modes.\r
63 + Fixed flags for ASL, NEG, NEGX, DIVU, ADDX, SUBX, ROXR.\r
64 + Bugs fixed in MOVEP, LINK, ADDQ, DIVS handlers.\r
65 * Undocumented flags for CHK, ABCD, SBCD and NBCD are now emulated the same way as in Musashi.\r
66 + Added Uninitialized Interrupt emulation.\r
67 + Altered timing for about half of opcodes to match Musashi's.\r
68\r
69v0.0082 Reesy\r
70 + Change cyclone to clear cycles before returning when halted\r
71 + Added Irq call back function. This allows emulators to be notified\r
72 when cyclone has taken an interrupt allowing them to set internal flags\r
73 which can help fix timing problems.\r
74\r
75v0.0081 notaz\r
76 + .asm version was broken and did not compile with armasm. Fixed.\r
77 + Finished implementing Stop opcode. Now it really stops the processor.\r
78\r
79v0.0080 notaz\r
80 + Added real cmpm opcode, it was using eor handler before this.\r
81 Fixes Dune and Sensible Soccer.\r
82\r
83v0.0078 notaz\r
84 note: these bugs were actually found Reesy, I reimplemented these by\r
85 using his changelog as a guide.\r
86 + Fixed a problem with divu which was using long divisor instead of word.\r
87 Fixes gear switching in Top Gear 2.\r
88 + Fixed btst opcode, The bit to test should shifted a max of 31 or 7\r
89 depending on if a register or memory location is being tested.\r
90 + Fixed abcd,sbcd. They did bad decimal correction on invalid BCD numbers\r
91 Score counters in Streets of Rage level end work now.\r
92 + Changed flag handling of abcd,sbcd,addx,subx,asl,lsl,...\r
93 Some ops did not have flag handling at all.\r
94 Some ops must not change Z flag when result is zero, but they did.\r
95 Shift ops must not change X if shift count is zero, but they did.\r
96 There are probably still some flag problems left.\r
97 + Patially implemented Stop and Reset opcodes - Fixes Thunderforce IV\r
98\r
99v0.0075 notaz\r
100 + Added missing displacement addressing mode for movem (Fantastic Dizzy)\r
101 + Added OSP <-> A7 swapping code in opcodes, which change privilege mode\r
102 + Implemented privilege violation, line emulator and divide by zero exceptions\r
103 + Added negx opcode (Shining Force works!)\r
104 + Added overflow detection for divs/divu\r
105\r
106v0.0072 notaz\r
107 note: I could only get v0.0069 cyclone, so I had to implement these myself using Dave's\r
108 changelog as a guide.\r
109 + Fixed a problem with divs - remainder should be negative when divident is negative\r
110 + Added movep opcode (Sonic 3 works)\r
111 + Fixed a problem with DBcc incorrectly decrementing if the condition is true (Shadow of the Beast)\r
112\r
113v0.0069\r
114 + Added SBCD and the flags for ABCD/SBCD. Score and time now works in games such as\r
115 Rolling Thunder 2, Ghouls 'N Ghosts\r
116 + Fixed a problem with addx and subx with 8-bit and 16-bit values.\r
117 Ghouls 'N' Ghosts now works!\r
118\r
119v0.0068\r
120 + Added ABCD opcode (Streets of Rage works now!)\r
121\r
122v0.0067\r
123 + Added dbCC (After Burner)\r
124 + Added asr EA (Sonic 1 Boss/Labyrinth Zone)\r
125 + Added andi/ori/eori ccr (Altered Beast)\r
126 + Added trap (After Burner)\r
127 + Added special case for move.b (a7)+ and -(a7), stepping by 2\r
128 After Burner is playable! Eternal Champions shows more\r
129 + Fixed lsr.b/w zero flag (Ghostbusters)\r
130 Rolling Thunder 2 now works!\r
131 + Fixed N flag for .b and .w arithmetic. Golden Axe works!\r
132\r
133v0.0066\r
134 + Fixed a stupid typo for exg (orr r10,r10, not orr r10,r8), which caused alignment\r
135 crashes on Strider\r
136\r
137v0.0065\r
138 + Fixed a problem with immediate values - they weren't being shifted up correctly for some\r
139 opcodes. Spiderman works, After Burner shows a bit of graphics.\r
140 + Fixed a problem with EA:"110nnn" extension word. 32-bit offsets were being decoded as 8-bit\r
141 offsets by mistake. Castlevania Bloodlines seems fine now.\r
142 + Added exg opcode\r
143 + Fixed asr opcode (Sonic jumping left is fixed)\r
144 + Fixed a problem with the carry bit in rol.b (Marble Madness)\r
145\r
146v0.0064\r
147 + Added rtr\r
148 + Fixed addq/subq.l (all An opcodes are 32-bit) (Road Rash)\r
149 + Fixed various little timings\r
150\r
151v0.0063\r
152 + Added link/unlk opcodes\r
153 + Fixed various little timings\r
154 + Fixed a problem with dbCC opcode being emitted at set opcodes\r
155 + Improved long register access, the EA fetch now does ldr r0,[r7,r0,lsl #2] whenever\r
156 possible, saving 1 or 2 cycles on many opcodes, which should give a nice speed up.\r
157 + May have fixed N flag on ext opcode?\r
158 + Added dasm for link opcode.\r
159\r
160v0.0062\r
161 * I was a bit too keen with the Arithmetic opcodes! Some of them should have been abcd,\r
162 exg and addx. Removed the incorrect opcodes, pending re-adding them as abcd, exg and addx.\r
163 + Changed unknown opcodes to act as nops.\r
164 Not very technical, but fun - a few more games show more graphics ;)\r
165\r
166v0.0060\r
167 + Fixed divu (EA intro)\r
168 + Added sf (set false) opcode - SOR2\r
169 * Todo: pea/link/unlk opcodes\r
170\r
171v0.0059: Added remainder to divide opcodes.\r
172\r
173\r
174The new stuff\r
175-------------\r
176\r
177Before using Cyclone, be sure to customize config.h to better suit your project. All options\r
178are documented inside that file.\r
179\r
180IrqCallback has been changed a bit, unlike in previous version, it should not return anything.\r
181If you need to change IRQ level, you can safely do that in your handler.\r
182\r
183Cyclone has changed quite a bit from the time when Dave stopped updating it, but the rest of\r
184documentation still applies, so read it if you haven't done that yet. If you have, check the\r
185"Accessing ..." parts.\r
186\r
187\r
188ARM Register Usage\r
189------------------\r
190\r
191See source code for up to date of register usage, however a summary is here:\r
192\r
193 r0-3: Temporary registers\r
194 r4 : Current PC + Memory Base (i.e. pointer to next opcode)\r
195 r5 : Cycles remaining\r
196 r6 : Pointer to Opcode Jump table\r
197 r7 : Pointer to Cpu Context\r
198 r8 : Current Opcode\r
199 r9 : Flags (NZCV) in highest four bits\r
200 (r10 : Temporary source value or Memory Base)\r
201 (r11 : Temporary register)\r
202\r
203\r
204How to Compile\r
205--------------\r
206\r
207Like Starscream and A68K, Cyclone uses a 'Core Creator' program which calculates and outputs\r
208all possible 68000 Opcodes and a jump table into files called Cyclone.s and .asm\r
209It then assembles these files into Cyclone.o and .obj\r
210\r
211Cyclone.o is the GCC assembled version and Cyclone.obj is the Microsoft assembled version.\r
212\r
213First unzip "Cyclone.zip" into a "Cyclone" directory.\r
214If you are compiling for Windows CE, find ARMASM.EXE (the Microsoft ARM assembler) and\r
215put it in the directory as well or put it on your path.\r
216\r
217Open up Cyclone.dsw in Visual Studio 6.0, compile and run the project.\r
218Cyclone.obj and Cyclone.o will be created.\r
219\r
220\r
221Compiling without Visual C++\r
222----------------------------\r
223If you aren't using Visual C++, it still shouldn't be too hard to compile, just get a C compiler,\r
224compile all the CPPs and C file, link them into an EXE, and run the exe.\r
225\r
226 e.g. gcc Main.cpp OpAny.cpp OpArith.cpp OpBranch.cpp OpLogic.cpp OpMove.cpp Disa.c\r
227 Main.exe\r
228\r
229\r
230Adding to your project\r
231----------------------\r
232\r
233To add Cyclone to you project, add Cyclone.o or obj, and include Cyclone.h\r
234There is one structure: 'struct Cyclone', and one function: CycloneRun\r
235\r
236Don't worry if this seem very minimal - its all you need to run as many 68000s as you want.\r
237It works with both C and C++.\r
238\r
239Byteswapped Memory\r
240------------------\r
241\r
242If you have used Starscream, A68K or Turbo68K or similar emulators you'll be familiar with this!\r
243\r
244Any memory which the 68000 can access directly must be have every two bytes swapped around.\r
245This is to speed up 16-bit memory accesses, because the 68000 has Big-Endian memory\r
246and ARM has Little-Endian memory.\r
247\r
248Now you may think you only technically have to byteswap ROM, not RAM, because\r
24916-bit RAM reads go through a memory handler and you could just return (mem[a]<<8) | mem[a+1].\r
250\r
251This would work, but remember some systems can execute code from RAM as well as ROM, and\r
252that would fail.\r
253So it's best to use byteswapped ROM and RAM if the 68000 can access it directly.\r
254It's also faster for the memory handlers, because you can do this:\r
255 \r
256 return *(unsigned short *)(mem+a)\r
257\r
258\r
259Declaring Memory handlers\r
260-------------------------\r
261\r
262Before you can reset or execute 68000 opcodes you must first set up a set of memory handlers.\r
263There are 7 functions you have to set up per CPU, like this:\r
264\r
265 static unsigned int MyCheckPc(unsigned int pc)\r
266 static unsigned char MyRead8 (unsigned int a)\r
267 static unsigned short MyRead16 (unsigned int a)\r
268 static unsigned int MyRead32 (unsigned int a)\r
269 static void MyWrite8 (unsigned int a,unsigned char d)\r
270 static void MyWrite16(unsigned int a,unsigned short d)\r
271 static void MyWrite32(unsigned int a,unsigned int d)\r
272\r
273You can think of these functions representing the 68000's memory bus.\r
274The Read and Write functions are called whenever the 68000 reads or writes memory.\r
275For example you might set MyRead8 like this:\r
276\r
277 unsigned char MyRead8(unsigned int a)\r
278 {\r
279 a&=0xffffff; // Clip address to 24-bits\r
280\r
281 if (a<RomLength) return RomData[a^1]; // ^1 because the memory is byteswapped\r
282 if (a>=0xe00000) return RamData[(a^1)&0xffff];\r
283 return 0xff; // Out of range memory access\r
284 }\r
285\r
286The other 5 read/write functions are similar. I'll describe the CheckPc function later on.\r
287\r
288Declaring a CPU Context\r
289-----------------------\r
290\r
291To declare a CPU simple declare a struct Cyclone in your code. For example to declare\r
292two 68000s:\r
293\r
294 struct Cyclone MyCpu;\r
295 struct Cyclone MyCpu2;\r
296\r
297It's probably a good idea to initialise the memory to zero:\r
298\r
299 memset(&MyCpu, 0,sizeof(MyCpu));\r
300 memset(&MyCpu2,0,sizeof(MyCpu2));\r
301\r
302Next point to your memory handlers:\r
303\r
304 MyCpu.checkpc=MyCheckPc;\r
305 MyCpu.read8 =MyRead8;\r
306 MyCpu.read16 =MyRead16;\r
307 MyCpu.read32 =MyRead32;\r
308 MyCpu.write8 =MyWrite8;\r
309 MyCpu.write16=MyWrite16;\r
310 MyCpu.write32=MyWrite32;\r
311\r
312You also need to point the fetch handlers - for most systems out there you can just\r
313point them at the read handlers:\r
314 MyCpu.fetch8 =MyRead8;\r
315 MyCpu.fetch16 =MyRead16;\r
316 MyCpu.fetch32 =MyRead32;\r
317\r
318( Why a different set of function pointers for fetch?\r
319 Well there are some systems, the main one being CPS2, which return different data\r
320 depending on whether the 'fetch' line on the 68000 bus is high or low.\r
321 If this is the case, you can set up different functions for fetch reads.\r
322 Generally though you don't need to. )\r
323\r
324Now you are nearly ready to reset the 68000, except you need one more function: checkpc().\r
325\r
326The checkpc() function\r
327----------------------\r
328\r
329When Cyclone reads opcodes, it doesn't use a memory handler every time, this would be\r
330far too slow, instead it uses a direct pointer to ARM memory.\r
331For example if your Rom image was at 0x3000000 and the program counter was $206,\r
332Cyclone's program counter would be 0x3000206.\r
333\r
334The difference between an ARM address and a 68000 address is also stored in a variable called\r
335'membase'. In the above example it's 0x3000000. To retrieve the real PC, Cyclone just\r
336subtracts 'membase'.\r
337\r
338When a long jump happens, Cyclone calls checkpc(). If the PC is in a different bank,\r
339for example Ram instead of Rom, change 'membase', recalculate the new PC and return it:\r
340\r
341static int MyCheckPc(unsigned int pc)\r
342{\r
343 pc-=MyCpu.membase; // Get the real program counter\r
344\r
345 if (pc<RomLength) MyCpu.membase=(int)RomMem; // Jump to Rom\r
346 if (pc>=0xff0000) MyCpu.membase=(int)RamMem-0xff0000; // Jump to Ram\r
347\r
348 return MyCpu.membase+pc; // New program counter\r
349}\r
350\r
351Notice that the membase is always ARM address minus 68000 address.\r
352\r
353The above example doesn't consider mirrored ram, but for an example of what to do see\r
354PicoDrive (in Memory.cpp).\r
355\r
356\r
357Almost there - Reset the 68000!\r
358-------------------------------\r
359\r
360Next we need to Reset the 68000 to get the initial Program Counter and Stack Pointer. This\r
361is obtained from addresses 000000 and 000004.\r
362\r
363Here is code which resets the 68000 (using your memory handlers):\r
364\r
365 MyCpu.srh=0x27; // Set supervisor mode\r
366 MyCpu.a[7]=MyCpu.read32(0); // Get Stack Pointer\r
367 MyCpu.membase=0;\r
368 MyCpu.pc=MyCpu.checkpc(MyCpu.read32(4)); // Get Program Counter\r
369\r
370And that's ready to go.\r
371\r
372\r
373Executing the 68000\r
374-------------------\r
375\r
376To execute the 68000, set the 'cycles' variable to the number of cycles you wish to execute,\r
377and then call CycloneRun with a pointer to the Cyclone structure.\r
378\r
379e.g.:\r
380 // Execute 1000 cycles on the 68000:\r
381 MyCpu.cycles=1000; CycloneRun(&MyCpu);\r
382\r
383For each opcode, the number of cycles it took is subtracted and the function returns when\r
384it reaches 0.\r
385\r
386e.g.\r
387 // Execute one instruction on the 68000:\r
388 MyCpu.cycles=0; CycloneRun(&MyCpu);\r
389 printf(" The opcode took %d cycles\n", -MyCpu.cycles);\r
390\r
391You should try to execute as many cycles as you can for maximum speed.\r
392The number actually executed may be slightly more than requested, i.e. cycles may come\r
393out with a small negative value:\r
394\r
395e.g.\r
396 int todo=12000000/60; // 12Mhz, for one 60hz frame\r
397 MyCpu.cycles=todo; CycloneRun(&MyCpu);\r
398 printf(" Actually executed %d cycles\n", todo-MyCpu.cycles);\r
399\r
400To calculate the number of cycles executed, use this formula:\r
401 Number of cycles requested - Cycle counter at the end\r
402\r
403\r
404Interrupts\r
405----------\r
406\r
407Causing an interrupt is very simple, simply set the irq variable in the Cyclone structure\r
408to the IRQ number.\r
409To lower the IRQ line, set it to zero.\r
410\r
411e.g:\r
412 MyCpu.irq=6; // Interrupt level 6\r
413 MyCpu.cycles=20000; CycloneRun(&MyCpu);\r
414\r
415Note that the interrupt is not actually processed until the next call to CycloneRun,\r
416and the interrupt may not be taken until the 68000 interrupt mask is changed to allow it.\r
417\r
418( The IRQ isn't checked on exiting from a memory handler: I don't think this will cause\r
419 me any trouble because I've never needed to trigger an interrupt from a memory handler,\r
420 but if someone needs to, let me know...)\r
421\r
422\r
423Accessing Cycle Counter\r
424-----------------------\r
425\r
426The cycle counter in the Cyclone structure is not, by default, updated before\r
427calling a memory handler, only at the end of an execution.\r
428\r
429*update*\r
430Now this is configurable in config.h, there is no 'debug' variable.\r
431\r
432\r
433Accessing Program Counter and registers\r
434---------------------------------------\r
435\r
436You can read Cyclone's registers directly from the structure at any time (as far as I know).\r
437\r
438The Program Counter, should you need to read or write it, is stored with membase\r
439added on. So use this formula to calculate the real 68000 program counter:\r
440\r
441 pc = MyCpu.pc - MyCpu.membase;\r
442\r
443The program counter is stored in r4 during execution, and isn't written back to the\r
444structure until the end of execution, which means you can't read normally real it from\r
445a memory handler.\r
446\r
447*update*\r
448Now this is configurable in config.h, there is no 'debug' variable. You can even enable\r
449access to SR if you need. However changing PC in memhandlers is still not safe, you should\r
450better clear cycles, wait untill CycloneRun() returns and then do whatever you need.\r
451\r
452\r
453Emulating more than one CPU\r
454---------------------------\r
455\r
456Since everything is based on the structures, emulating more than one cpu at the same time\r
457is just a matter of declaring more than one structures and timeslicing. You can emulate\r
458as many 68000s as you want.\r
459Just set up the memory handlers for each cpu and run each cpu for a certain number of cycles.\r
460\r
461e.g.\r
462 // Execute 1000 cycles on 68000 #1:\r
463 MyCpu.cycles=1000; CycloneRun(&MyCpu);\r
464\r
465 // Execute 1000 cycles on 68000 #2:\r
466 MyCpu2.cycles=1000; CycloneRun(&MyCpu2);\r
467\r
468\r
469Thanks to...\r
470------------\r
471\r
472* All the previous code-generating assembler cpu core guys!\r
473 Who are iirc... Neill Corlett, Neil Bradley, Mike Coates, Darren Olafson\r
474 and Bart Trzynadlowski\r
475\r
476* Charles Macdonald, for researching just about every console ever\r
477* MameDev+FBA, for keeping on going and going and going\r
478\r
479\r
480-------------\r
481\r
482Dave - 17th April 2004\r
483notaz - 17th July 2006\r
484\r
485Homepage: http://www.finalburn.com/\r
486Dave's e-mail: dev(atsymbol)finalburn.com\r
487Replace (atsymbol) with @\r