diff --git a/src/Makefile.am b/src/Makefile.am index 1732e68..83236d0 100644 --- a/src/Makefile.am +++ b/src/Makefile.am @@ -30,7 +30,7 @@ codegen_ops_branch.c codegen_ops_fpu_arith.c codegen_ops_fpu_constant.c codegen_ codegen_ops_fpu_misc.c codegen_ops_mmx_arith.c codegen_ops_mmx_cmp.c codegen_ops_mmx_loadstore.c \ codegen_ops_mmx_logic.c codegen_ops_mmx_pack.c codegen_ops_mmx_shift.c codegen_ops_helpers.c codegen_ops_jump.c \ codegen_ops_logic.c codegen_ops_shift.c codegen_ops_misc.c codegen_ops_mov.c codegen_ops_stack.c codegen_reg.c \ -codegen_timing_486.c codegen_timing_686.c codegen_timing_common.c codegen_timing_k6.c codegen_timing_p6.c codegen_timing_pentium.c \ +codegen_timing_486.c codegen_timing_686.c codegen_timing_common.c codegen_timing_cyrixiii.c codegen_timing_k6.c codegen_timing_p6.c codegen_timing_pentium.c \ codegen_timing_winchip.c codegen_timing_winchip2.c compaq.c config.c cpu.c cpu_tables.c cs8230.c dells200.c device.c disc.c \ disc_fdi.c disc_img.c disc_sector.c dma.c esdi_at.c f82c710_upc.c fdc.c fdc37c665.c fdc37c93x.c fdd.c fdi2raw.c gameport.c hdd.c hdd_esdi.c \ hdd_file.c headland.c i430lx.c i430fx.c i430hx.c i430vx.c i440fx.c i440bx.c ide.c ide_atapi.c ide_sff8038i.c intel.c intel_flash.c io.c \ diff --git a/src/Makefile.mingw-wx-sdl2 b/src/Makefile.mingw-wx-sdl2 index d7ae669..4c13f55 100644 --- a/src/Makefile.mingw-wx-sdl2 +++ b/src/Makefile.mingw-wx-sdl2 @@ -15,7 +15,7 @@ OBJ = 386.o 386_common.o 386_dynarec.o 386_dynarec_ops.o 808x.o 82091aa.o acc203 codegen_ops_logic.o codegen_ops_misc.o codegen_ops_mmx_arith.o codegen_ops_mmx_cmp.o \ codegen_ops_mmx_loadstore.o codegen_ops_mmx_logic.o codegen_ops_mmx_pack.o codegen_ops_mmx_shift.o \ codegen_ops_mov.o codegen_ops_shift.o codegen_ops_stack.o codegen_reg.o codegen_timing_486.o \ - codegen_timing_686.o codegen_timing_common.o codegen_timing_k6.o codegen_timing_p6.o codegen_timing_pentium.o \ + codegen_timing_686.o codegen_timing_common.o codegen_timing_cyrixiii.o codegen_timing_k6.o codegen_timing_p6.o codegen_timing_pentium.o \ codegen_timing_winchip.o codegen_timing_winchip2.o compaq.o config.o cpu.o cpu_tables.o cs8230.o device.o \ dells200.o disc.o disc_fdi.o disc_img.o disc_sector.o dma.o esdi_at.o f82c710_upc.o fdc.o fdc37c665.o fdc37c93x.o fdd.o \ fdi2raw.o gameport.o hdd.o hdd_esdi.o hdd_file.o headland.o i430hx.o i430lx.o i430fx.o i430vx.o i440fx.o i440bx.o ide.o \ diff --git a/src/Makefile.mingw-wx-sdl2-network b/src/Makefile.mingw-wx-sdl2-network index 7a4bb46..e58e406 100644 --- a/src/Makefile.mingw-wx-sdl2-network +++ b/src/Makefile.mingw-wx-sdl2-network @@ -15,7 +15,7 @@ OBJ = 386.o 386_common.o 386_dynarec.o 386_dynarec_ops.o 808x.o 82091aa.o acc203 codegen_ops_logic.o codegen_ops_misc.o codegen_ops_mmx_arith.o codegen_ops_mmx_cmp.o \ codegen_ops_mmx_loadstore.o codegen_ops_mmx_logic.o codegen_ops_mmx_pack.o codegen_ops_mmx_shift.o \ codegen_ops_mov.o codegen_ops_shift.o codegen_ops_stack.o codegen_reg.o codegen_timing_486.o \ - codegen_timing_686.o codegen_timing_common.o codegen_timing_k6.o codegen_timing_p6.o codegen_timing_pentium.o \ + codegen_timing_686.o codegen_timing_common.o codegen_timing_cyrixiii.o codegen_timing_k6.o codegen_timing_p6.o codegen_timing_pentium.o \ codegen_timing_winchip.o codegen_timing_winchip2.o compaq.o config.o cpu.o cpu_tables.o cs8230.o device.o \ dells200.o disc.o disc_fdi.o disc_img.o disc_sector.o dma.o esdi_at.o f82c710_upc.o fdc.o fdc37c665.o fdc37c93x.o fdd.o \ fdi2raw.o gameport.o hdd.o hdd_esdi.o hdd_file.o headland.o i430hx.o i430lx.o i430fx.o i430vx.o i440fx.o i440bx.o ide.o \ diff --git a/src/codegen.h b/src/codegen.h index 0ca3df5..83f37ae 100644 --- a/src/codegen.h +++ b/src/codegen.h @@ -374,6 +374,7 @@ extern codegen_timing_t codegen_timing_686; extern codegen_timing_t codegen_timing_486; extern codegen_timing_t codegen_timing_winchip; extern codegen_timing_t codegen_timing_winchip2; +extern codegen_timing_t codegen_timing_cyrixiii; extern codegen_timing_t codegen_timing_k6; extern codegen_timing_t codegen_timing_p6; diff --git a/src/codegen_timing_cyrixiii.c b/src/codegen_timing_cyrixiii.c new file mode 100644 index 0000000..c89defc --- /dev/null +++ b/src/codegen_timing_cyrixiii.c @@ -0,0 +1,679 @@ +/*Since VIA didn't document cycle timings in the Cyrix III/C3 datasheets, and + I don't currently own either chip to test against, most of the timing here is + a guess. This code makes the current (probably wrong) assumptions, based on the + few details in the Cyrix III datasheet : + - Most common integer instructions are 1 cycle, including register-memory and + memory-register versions + - Basic FPU instructions are 2 cycles. The datasheet claims "a non-dependent + add or multiply can start every two clocks"; it's not clear if this is referring + to CPU or FPU clocks, since the FPU runs at 1/2 CPU speed on these chips. + - MMX instructions are 1 cycle + - No instruction pairing of any kind (unlike WinChip 2) + - Prefixes other than 0x0f cost 1 cycle each + - AGI stalls cost 1 cycle, except for implicit references to ESP*/ +#include "ibm.h" +#include "cpu.h" +#include "x86.h" +#include "x86_ops.h" +#include "x87.h" +#include "mem.h" +#include "codegen.h" +#include "codegen_ops.h" +#include "codegen_timing_common.h" + +/*Instruction has different execution time for 16 and 32 bit data. Does not pair */ +#define CYCLES_HAS_MULTI (1 << 31) + +#define CYCLES_FPU (1 << 30) + +#define CYCLES_IS_MMX_MUL (1 << 29) +#define CYCLES_IS_MMX_SHIFT (1 << 28) +#define CYCLES_IS_MMX_ANY (1 << 27) +#define CYCLES_IS_3DNOW (1 << 26) + +#define CYCLES_MMX_MUL(c) (CYCLES_IS_MMX_MUL | c) +#define CYCLES_MMX_SHIFT(c) (CYCLES_IS_MMX_SHIFT | c) +#define CYCLES_MMX_ANY(c) (CYCLES_IS_MMX_ANY | c) +#define CYCLES_3DNOW(c) (CYCLES_IS_3DNOW | c) + +#define CYCLES_IS_MMX (CYCLES_IS_MMX_MUL | CYCLES_IS_MMX_SHIFT | CYCLES_IS_MMX_ANY | CYCLES_IS_3DNOW) + +#define GET_CYCLES(c) (c & ~(CYCLES_HAS_MULTI | CYCLES_FPU | CYCLES_IS_MMX)) + +#define CYCLES(c) c +#define CYCLES2(c16, c32) (CYCLES_HAS_MULTI | c16 | (c32 << 8)) + +/*comp_time = cycles until instruction complete + i_overlap = cycles that overlap with integer + f_overlap = cycles that overlap with subsequent FPU*/ +#define FPU_CYCLES(comp_time, i_overlap, f_overlap) (comp_time) | (i_overlap << 8) | (f_overlap << 16) | CYCLES_FPU + +#define FPU_COMP_TIME(timing) (timing & 0xff) +#define FPU_I_OVERLAP(timing) ((timing >> 8) & 0xff) +#define FPU_F_OVERLAP(timing) ((timing >> 16) & 0xff) + +#define FPU_I_LATENCY(timing) (FPU_COMP_TIME(timing) - FPU_I_OVERLAP(timing)) + +#define FPU_F_LATENCY(timing) (FPU_I_OVERLAP(timing) - FPU_F_OVERLAP(timing)) + +#define FPU_RESULT_LATENCY(timing) ((timing >> 8) & 0xff) + +#define INVALID 0 + +static uint32_t opcode_timings[256] = +{ +/*00*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), CYCLES(3), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), INVALID, +/*10*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), CYCLES(3), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), CYCLES(3), +/*20*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(3), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(3), +/*30*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), + +/*40*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), +/*50*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), +/*60*/ CYCLES(11), CYCLES(9), CYCLES(7), CYCLES(9), CYCLES(4), CYCLES(4), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES2(17,25), CYCLES(1), CYCLES2(17,20), CYCLES(17), CYCLES(17), CYCLES(17), CYCLES(17), +/*70*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), + +/*80*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), CYCLES(2), CYCLES(5), CYCLES(5), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(3), CYCLES(1), CYCLES(5), CYCLES(6), +/*90*/ CYCLES(1), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(0), CYCLES(4), CYCLES(4), CYCLES(5), CYCLES(2), CYCLES(3), +/*a0*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(7), CYCLES(7), CYCLES(8), CYCLES(8), CYCLES(1), CYCLES(1), CYCLES(5), CYCLES(5), CYCLES(5), CYCLES(5), CYCLES(6), CYCLES(6), +/*b0*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), + +/*c0*/ CYCLES(1), CYCLES(1), CYCLES(5), CYCLES(5), CYCLES(6), CYCLES(6), CYCLES(1), CYCLES(1), CYCLES(14), CYCLES(5), CYCLES(0), CYCLES(0), CYCLES(0), CYCLES(0), CYCLES(3), CYCLES(0), +/*d0*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(15), CYCLES(14), CYCLES(2), CYCLES(4), INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, +/*e0*/ CYCLES(6), CYCLES(6), CYCLES(6), CYCLES(5), CYCLES(14), CYCLES(14), CYCLES(16), CYCLES(16), CYCLES(3), CYCLES(3), CYCLES(17), CYCLES(3), CYCLES(14), CYCLES(14), CYCLES(14), CYCLES(14), +/*f0*/ CYCLES(4), CYCLES(0), CYCLES(0), CYCLES(0), CYCLES(4), CYCLES(2), INVALID, INVALID, CYCLES(2), CYCLES(2), CYCLES(3), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(3), INVALID +}; + +static uint32_t opcode_timings_mod3[256] = +{ +/*00*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), CYCLES(3), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), INVALID, +/*10*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), CYCLES(3), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), CYCLES(3), +/*20*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(3), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(3), +/*30*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), + +/*40*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), +/*50*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), +/*60*/ CYCLES(11), CYCLES(9), CYCLES(7), CYCLES(9), CYCLES(4), CYCLES(4), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES2(14,25), CYCLES(1), CYCLES2(17,20), CYCLES(17), CYCLES(17), CYCLES(17), CYCLES(17), +/*70*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), + +/*80*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(5), CYCLES(5), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(2), CYCLES(1), CYCLES(2), CYCLES(1), +/*90*/ CYCLES(1), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(0), CYCLES(4), CYCLES(4), CYCLES(5), CYCLES(2), CYCLES(3), +/*a0*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(7), CYCLES(7), CYCLES(8), CYCLES(8), CYCLES(1), CYCLES(1), CYCLES(5), CYCLES(5), CYCLES(5), CYCLES(5), CYCLES(6), CYCLES(6), +/*b0*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), + +/*c0*/ CYCLES(1), CYCLES(1), CYCLES(5), CYCLES(5), CYCLES(6), CYCLES(6), CYCLES(1), CYCLES(1), CYCLES(14), CYCLES(5), CYCLES(0), CYCLES(0), CYCLES(0), CYCLES(0), CYCLES(3), CYCLES(0), +/*d0*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(15), CYCLES(14), CYCLES(2), CYCLES(4), INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, +/*e0*/ CYCLES(6), CYCLES(6), CYCLES(6), CYCLES(5), CYCLES(14), CYCLES(14), CYCLES(16), CYCLES(16), CYCLES(3), CYCLES(3), CYCLES(17), CYCLES(3), CYCLES(14), CYCLES(14), CYCLES(14), CYCLES(14), +/*f0*/ CYCLES(4), CYCLES(0), CYCLES(0), CYCLES(0), CYCLES(4), CYCLES(2), INVALID, INVALID, CYCLES(2), CYCLES(2), CYCLES(3), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(3), INVALID, +}; + +static uint32_t opcode_timings_0f[256] = +{ +/*00*/ CYCLES(20), CYCLES(11), CYCLES(11), CYCLES(10), INVALID, CYCLES(195), CYCLES(7), INVALID, CYCLES(1000), CYCLES(10000), INVALID, INVALID, INVALID, CYCLES_3DNOW(1), CYCLES(1), CYCLES_3DNOW(1), +/*10*/ INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, +/*20*/ CYCLES(6), CYCLES(6), CYCLES(6), CYCLES(6), CYCLES(6), CYCLES(6), INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, +/*30*/ CYCLES(9), CYCLES(1), CYCLES(9), INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, + +/*40*/ INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, +/*50*/ INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, +/*60*/ CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), INVALID, INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), +/*70*/ INVALID, CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES(100), INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), + +/*80*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), +/*90*/ CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), +/*a0*/ CYCLES(3), CYCLES(3), CYCLES(14), CYCLES(8), CYCLES(3), CYCLES(4), INVALID, INVALID, CYCLES(3), CYCLES(3), INVALID, CYCLES(13), CYCLES(3), CYCLES(3), INVALID, CYCLES2(18,30), +/*b0*/ CYCLES(10), CYCLES(10), CYCLES(6), CYCLES(13), CYCLES(6), CYCLES(6), CYCLES(3), CYCLES(3), INVALID, INVALID, CYCLES(6), CYCLES(13), CYCLES(7), CYCLES(7), CYCLES(3), CYCLES(3), + +/*c0*/ CYCLES(4), CYCLES(4), INVALID, INVALID, INVALID, INVALID, INVALID, CYCLES(3), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), +/*d0*/ INVALID, CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), INVALID, CYCLES_MMX_MUL(1), INVALID, INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, CYCLES_MMX_ANY(1), +/*e0*/ INVALID, CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), INVALID, INVALID, CYCLES_MMX_MUL(1), INVALID, INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, CYCLES_MMX_ANY(1), +/*f0*/ INVALID, CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), INVALID, CYCLES_MMX_MUL(1), INVALID, INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(2), INVALID, +}; +static uint32_t opcode_timings_0f_mod3[256] = +{ +/*00*/ CYCLES(20), CYCLES(11), CYCLES(11), CYCLES(10), INVALID, CYCLES(195), CYCLES(7), INVALID, CYCLES(1000), CYCLES(10000), INVALID, INVALID, INVALID, CYCLES_3DNOW(1), CYCLES(1), CYCLES_3DNOW(1), +/*10*/ INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, +/*20*/ CYCLES(6), CYCLES(6), CYCLES(6), CYCLES(6), CYCLES(6), CYCLES(6), INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, +/*30*/ CYCLES(9), CYCLES(1), CYCLES(9), INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, + +/*40*/ INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, +/*50*/ INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, +/*60*/ CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), INVALID, INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), +/*70*/ INVALID, CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES(100), INVALID, INVALID, INVALID, INVALID, INVALID, INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), + +/*80*/ CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), +/*90*/ CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3), +/*a0*/ CYCLES(3), CYCLES(3), CYCLES(14), CYCLES(8), CYCLES(3), CYCLES(4), INVALID, INVALID, CYCLES(3), CYCLES(3), INVALID, CYCLES(13), CYCLES(3), CYCLES(3), INVALID, CYCLES2(18,30), +/*b0*/ CYCLES(10), CYCLES(10), CYCLES(6), CYCLES(13), CYCLES(6), CYCLES(6), CYCLES(3), CYCLES(3), INVALID, INVALID, CYCLES(6), CYCLES(13), CYCLES(7), CYCLES(7), CYCLES(3), CYCLES(3), + +/*c0*/ CYCLES(4), CYCLES(4), INVALID, INVALID, INVALID, INVALID, INVALID, CYCLES(3), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), CYCLES(1), +/*d0*/ INVALID, CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), INVALID, CYCLES_MMX_MUL(1), INVALID, INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, CYCLES_MMX_ANY(1), +/*e0*/ INVALID, CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), INVALID, INVALID, CYCLES_MMX_MUL(1), INVALID, INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, CYCLES_MMX_ANY(1), +/*f0*/ INVALID, CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), CYCLES_MMX_SHIFT(1), INVALID, CYCLES_MMX_MUL(1), INVALID, INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), CYCLES_MMX_ANY(1), INVALID, +}; + +static uint32_t opcode_timings_shift[8] = +{ + CYCLES(7), CYCLES(7), CYCLES(10), CYCLES(10), CYCLES(7), CYCLES(7), CYCLES(7), CYCLES(7) +}; +static uint32_t opcode_timings_shift_mod3[8] = +{ + CYCLES(3), CYCLES(3), CYCLES(9), CYCLES(9), CYCLES(3), CYCLES(3), CYCLES(3), CYCLES(3) +}; + +static uint32_t opcode_timings_f6[8] = +{ + CYCLES(2), INVALID, CYCLES(2), CYCLES(2), CYCLES(13), CYCLES(14), CYCLES(16), CYCLES(19) +}; +static uint32_t opcode_timings_f6_mod3[8] = +{ + CYCLES(1), INVALID, CYCLES(1), CYCLES(1), CYCLES(13), CYCLES(14), CYCLES(16), CYCLES(19) +}; +static uint32_t opcode_timings_f7[8] = +{ + CYCLES(2), INVALID, CYCLES(2), CYCLES(2), CYCLES(21), CYCLES2(22,38), CYCLES2(24,40), CYCLES2(27,43) +}; +static uint32_t opcode_timings_f7_mod3[8] = +{ + CYCLES(1), INVALID, CYCLES(1), CYCLES(1), CYCLES(21), CYCLES2(22,38), CYCLES2(24,40), CYCLES2(27,43) +}; +static uint32_t opcode_timings_ff[8] = +{ + CYCLES(2), CYCLES(2), CYCLES(5), CYCLES(0), CYCLES(5), CYCLES(0), CYCLES(5), INVALID +}; +static uint32_t opcode_timings_ff_mod3[8] = +{ + CYCLES(1), CYCLES(1), CYCLES(5), CYCLES(0), CYCLES(5), CYCLES(0), CYCLES(5), INVALID +}; + +static uint32_t opcode_timings_d8[8] = +{ +/* FADDs FMULs FCOMs FCOMPs*/ + FPU_CYCLES(4,2,2), FPU_CYCLES(4,2,2), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), +/* FSUBs FSUBRs FDIVs FDIVRs*/ + FPU_CYCLES(4,2,2), FPU_CYCLES(4,2,2), FPU_CYCLES(39,38,2), FPU_CYCLES(39,38,2) +}; +static uint32_t opcode_timings_d8_mod3[8] = +{ +/* FADD FMUL FCOM FCOMP*/ + FPU_CYCLES(4,2,2), FPU_CYCLES(4,2,2), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), +/* FSUB FSUBR FDIV FDIVR*/ + FPU_CYCLES(4,2,2), FPU_CYCLES(4,2,2), FPU_CYCLES(39,38,2), FPU_CYCLES(39,38,2) +}; + +static uint32_t opcode_timings_d9[8] = +{ +/* FLDs FSTs FSTPs*/ + FPU_CYCLES(2,0,0), INVALID, FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), +/* FLDENV FLDCW FSTENV FSTCW*/ + FPU_CYCLES(32,0,0), FPU_CYCLES(8,0,0), FPU_CYCLES(48,0,0), FPU_CYCLES(2,0,0) +}; +static uint32_t opcode_timings_d9_mod3[64] = +{ + /*FLD*/ + FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), + FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), + /*FXCH*/ + FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), + FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), + /*FNOP*/ + FPU_CYCLES(3,0,0), INVALID, INVALID, INVALID, + INVALID, INVALID, INVALID, INVALID, + /*FSTP*/ + FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), + FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), +/* opFCHS opFABS*/ + FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), INVALID, INVALID, +/* opFTST opFXAM*/ + FPU_CYCLES(2,0,0), FPU_CYCLES(21,4,0), INVALID, INVALID, +/* opFLD1 opFLDL2T opFLDL2E opFLDPI*/ + FPU_CYCLES(2,0,0), FPU_CYCLES(5,2,2), FPU_CYCLES(5,2,2), FPU_CYCLES(5,2,2), +/* opFLDEG2 opFLDLN2 opFLDZ*/ + FPU_CYCLES(5,2,2), FPU_CYCLES(5,2,2), FPU_CYCLES(2,0,0), INVALID, +/* opF2XM1 opFYL2X opFPTAN opFPATAN*/ + FPU_CYCLES(53,2,2), FPU_CYCLES(103,2,2),FPU_CYCLES(120,36,0),FPU_CYCLES(112,2,2), +/* opFDECSTP opFINCSTP,*/ + INVALID, INVALID, FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), +/* opFPREM opFSQRT opFSINCOS*/ + FPU_CYCLES(64,2,2), INVALID, FPU_CYCLES(70,69,2),FPU_CYCLES(89,2,2), +/* opFRNDINT opFSCALE opFSIN opFCOS*/ + FPU_CYCLES(9,0,0), FPU_CYCLES(20,5,0), FPU_CYCLES(65,2,2), FPU_CYCLES(65,2,2) +}; + +static uint32_t opcode_timings_da[8] = +{ +/* FIADDl FIMULl FICOMl FICOMPl*/ + FPU_CYCLES(6,2,2), FPU_CYCLES(6,2,2), FPU_CYCLES(4,0,0), FPU_CYCLES(4,0,0), +/* FISUBl FISUBRl FIDIVl FIDIVRl*/ + FPU_CYCLES(6,2,2), FPU_CYCLES(6,2,2), FPU_CYCLES(42,38,2), FPU_CYCLES(42,38,2) +}; +static uint32_t opcode_timings_da_mod3[8] = +{ + INVALID, INVALID, INVALID, INVALID, +/* FCOMPP*/ + INVALID, FPU_CYCLES(2,0,0), INVALID, INVALID +}; + + +static uint32_t opcode_timings_db[8] = +{ +/* FLDil FSTil FSTPil*/ + FPU_CYCLES(3,2,2), INVALID, FPU_CYCLES(6,0,0), FPU_CYCLES(6,0,0), +/* FLDe FSTPe*/ + INVALID, FPU_CYCLES(3,0,0), INVALID, FPU_CYCLES(3,0,0) +}; +static uint32_t opcode_timings_db_mod3[64] = +{ + INVALID, INVALID, INVALID, INVALID, + INVALID, INVALID, INVALID, INVALID, + + INVALID, INVALID, INVALID, INVALID, + INVALID, INVALID, INVALID, INVALID, + + INVALID, INVALID, INVALID, INVALID, + INVALID, INVALID, INVALID, INVALID, + + INVALID, INVALID, INVALID, INVALID, + INVALID, INVALID, INVALID, INVALID, + +/* opFNOP opFCLEX opFINIT*/ + INVALID, FPU_CYCLES(2,0,0), FPU_CYCLES(7,0,0), FPU_CYCLES(17,0,0), +/* opFNOP opFNOP*/ + FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), INVALID, INVALID, + + INVALID, INVALID, INVALID, INVALID, + INVALID, INVALID, INVALID, INVALID, + + INVALID, INVALID, INVALID, INVALID, + INVALID, INVALID, INVALID, INVALID, + + INVALID, INVALID, INVALID, INVALID, + INVALID, INVALID, INVALID, INVALID, +}; + +static uint32_t opcode_timings_dc[8] = +{ +/* FADDd FMULd FCOMd FCOMPd*/ + FPU_CYCLES(4,2,2), FPU_CYCLES(4,2,2), FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), +/* FSUBd FSUBRd FDIVd FDIVRd*/ + FPU_CYCLES(4,2,2), FPU_CYCLES(4,2,2), FPU_CYCLES(39,38,2), FPU_CYCLES(39,38,2) +}; +static uint32_t opcode_timings_dc_mod3[8] = +{ +/* opFADDr opFMULr*/ + FPU_CYCLES(4,2,2), FPU_CYCLES(4,2,2),INVALID, INVALID, +/* opFSUBRr opFSUBr opFDIVRr opFDIVr*/ + FPU_CYCLES(4,2,2), FPU_CYCLES(4,2,2),FPU_CYCLES(39,38,2), FPU_CYCLES(39,38,2) +}; + +static uint32_t opcode_timings_dd[8] = +{ +/* FLDd FSTd FSTPd*/ + FPU_CYCLES(2,0,0), INVALID, FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), +/* FRSTOR FSAVE FSTSW*/ + FPU_CYCLES(70,0,0), INVALID, FPU_CYCLES(127,0,0), FPU_CYCLES(6,0,0) +}; +static uint32_t opcode_timings_dd_mod3[8] = +{ +/* FFFREE FST FSTP*/ + FPU_CYCLES(2,0,0), INVALID, FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0), +/* FUCOM FUCOMP*/ + FPU_CYCLES(2,0,0), FPU_CYCLES(2,0,0),INVALID, INVALID +}; + +static uint32_t opcode_timings_de[8] = +{ +/* FIADDw FIMULw FICOMw FICOMPw*/ + FPU_CYCLES(6,2,2), FPU_CYCLES(6,2,2), FPU_CYCLES(4,0,0), FPU_CYCLES(4,0,0), +/* FISUBw FISUBRw FIDIVw FIDIVRw*/ + FPU_CYCLES(6,2,2), FPU_CYCLES(6,2,2), FPU_CYCLES(42,38,2), FPU_CYCLES(42,38,2) +}; +static uint32_t opcode_timings_de_mod3[8] = +{ +/* FADDP FMULP FCOMPP*/ + FPU_CYCLES(4,2,2), FPU_CYCLES(4,2,2), INVALID, FPU_CYCLES(2,0,0), +/* FSUBP FSUBRP FDIVP FDIVRP*/ + FPU_CYCLES(4,2,2), FPU_CYCLES(4,2,2), FPU_CYCLES(39,38,2), FPU_CYCLES(39,38,2) +}; + +static uint32_t opcode_timings_df[8] = +{ +/* FILDiw FISTiw FISTPiw*/ + FPU_CYCLES(3,2,2), INVALID, FPU_CYCLES(6,0,0), FPU_CYCLES(6,0,0), +/* FILDiq FBSTP FISTPiq*/ + INVALID, FPU_CYCLES(3,2,2), FPU_CYCLES(148,0,0), FPU_CYCLES(6,0,0) +}; +static uint32_t opcode_timings_df_mod3[8] = +{ + INVALID, INVALID, INVALID, INVALID, +/* FSTSW AX*/ + FPU_CYCLES(6,0,0), INVALID, INVALID, INVALID +}; + +static uint32_t opcode_timings_8x[8] = +{ + CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2) +}; +static uint32_t opcode_timings_8x_mod3[8] = +{ + CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2) +}; +static uint32_t opcode_timings_81[8] = +{ + CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2) +}; +static uint32_t opcode_timings_81_mod3[8] = +{ + CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2), CYCLES(2) +}; + +static int timing_count; +static uint8_t last_prefix; +static uint32_t regmask_modified; +static int decode_delay, decode_delay_offset; +static int fpu_latency; +static int fpu_st_latency[8]; + + +static inline int COUNT(uint32_t c, int op_32) +{ + if (c & CYCLES_FPU) + return FPU_I_LATENCY(c); + if (c & CYCLES_HAS_MULTI) + { + if (op_32 & 0x100) + return (c >> 8) & 0xff; + return c & 0xff; + } + return GET_CYCLES(c); +} + +static int check_agi(uint64_t *deps, uint8_t opcode, uint32_t fetchdat, int op_32) +{ + uint32_t addr_regmask = get_addr_regmask(deps[opcode], fetchdat, op_32); + + /*Instructions that use ESP implicitly (eg PUSH, POP, CALL etc) do not + cause AGIs with each other, but do with instructions that use it explicitly*/ + if ((addr_regmask & REGMASK_IMPL_ESP) && (regmask_modified & (1 << REG_ESP)) && !(regmask_modified & REGMASK_IMPL_ESP)) + addr_regmask |= (1 << REG_ESP); + + return (regmask_modified & addr_regmask) & ~REGMASK_IMPL_ESP; +} + +static int codegen_fpu_latencies(uint64_t deps, int reg) +{ + int latency = fpu_latency; + + if ((deps & FPU_RW_ST0) && fpu_st_latency[0] && fpu_st_latency[0] > latency) + latency = fpu_st_latency[0]; + if ((deps & FPU_RW_ST1) && fpu_st_latency[1] && fpu_st_latency[1] > latency) + latency = fpu_st_latency[1]; + if ((deps & FPU_RW_STREG) && fpu_st_latency[reg] && fpu_st_latency[reg] > latency) + latency = fpu_st_latency[reg]; + + return latency; +} + +#define SUB_AND_CLAMP(latency, count) \ + latency -= count; \ + if (latency < 0) \ + latency = 0 + +static void codegen_fpu_latency_clock(int count) +{ + SUB_AND_CLAMP(fpu_latency, count); + SUB_AND_CLAMP(fpu_st_latency[0], count); + SUB_AND_CLAMP(fpu_st_latency[1], count); + SUB_AND_CLAMP(fpu_st_latency[2], count); + SUB_AND_CLAMP(fpu_st_latency[3], count); + SUB_AND_CLAMP(fpu_st_latency[4], count); + SUB_AND_CLAMP(fpu_st_latency[5], count); + SUB_AND_CLAMP(fpu_st_latency[6], count); + SUB_AND_CLAMP(fpu_st_latency[7], count); +} + +static void codegen_instruction(uint32_t *timings, uint64_t *deps, uint8_t opcode, uint32_t fetchdat, int decode_delay_offset, int op_32, int exec_delay) +{ + int instr_cycles, latency = 0; + + if ((timings[opcode] & CYCLES_FPU) && !(deps[opcode] & FPU_FXCH)) + instr_cycles = latency = codegen_fpu_latencies(deps[opcode], fetchdat & 7); + else + instr_cycles = 0; + + if ((decode_delay + decode_delay_offset) > 0) + codegen_fpu_latency_clock(decode_delay + decode_delay_offset + instr_cycles); + else + codegen_fpu_latency_clock(instr_cycles); + instr_cycles += COUNT(timings[opcode], op_32); + instr_cycles += exec_delay; + if ((decode_delay + decode_delay_offset) > 0) + codegen_block_cycles += instr_cycles + decode_delay + decode_delay_offset; + else + codegen_block_cycles += instr_cycles; + decode_delay = (-instr_cycles) + 1; + + if (deps[opcode] & FPU_POP) + { + int c; + + for (c = 0; c < 7; c++) + fpu_st_latency[c] = fpu_st_latency[c+1]; + fpu_st_latency[7] = 0; + } + if (deps[opcode] & FPU_POP2) + { + int c; + + for (c = 0; c < 6; c++) + fpu_st_latency[c] = fpu_st_latency[c+2]; + fpu_st_latency[6] = fpu_st_latency[7] = 0; + } + if (timings[opcode] & CYCLES_FPU) + { + /* if (fpu_latency) + fatal("Bad latency FPU\n");*/ + fpu_latency = FPU_F_LATENCY(timings[opcode]); + } + + if (deps[opcode] & FPU_PUSH) + { + int c; + + for (c = 0; c < 7; c++) + fpu_st_latency[c+1] = fpu_st_latency[c]; + fpu_st_latency[0] = 0; + } + if (deps[opcode] & FPU_WRITE_ST0) + { +/* if (fpu_st_latency[0]) + fatal("Bad latency ST0\n");*/ + fpu_st_latency[0] = FPU_RESULT_LATENCY(timings[opcode]); + } + if (deps[opcode] & FPU_WRITE_ST1) + { +/* if (fpu_st_latency[1]) + fatal("Bad latency ST1\n");*/ + fpu_st_latency[1] = FPU_RESULT_LATENCY(timings[opcode]); + } + if (deps[opcode] & FPU_WRITE_STREG) + { + int reg = fetchdat & 7; + if (deps[opcode] & FPU_POP) + reg--; + if (reg >= 0 && + !(reg == 0 && (deps[opcode] & FPU_WRITE_ST0)) && + !(reg == 1 && (deps[opcode] & FPU_WRITE_ST1))) + { +/* if (fpu_st_latency[reg]) + fatal("Bad latency STREG %i %08x %i %016llx %02x\n",fpu_st_latency[reg], fetchdat, reg, timings[opcode], opcode);*/ + fpu_st_latency[reg] = FPU_RESULT_LATENCY(timings[opcode]); + } + } +} + +static void codegen_timing_cyrixiii_block_start() +{ + regmask_modified = 0; + decode_delay = decode_delay_offset = 0; +} + +static void codegen_timing_cyrixiii_start() +{ + timing_count = 0; + last_prefix = 0; +} + +static void codegen_timing_cyrixiii_prefix(uint8_t prefix, uint32_t fetchdat) +{ + if (prefix == 0x0f) + { + /*0fh prefix is 'free'*/ + last_prefix = prefix; + return; + } + /*On C3 all prefixes take 1 cycle to decode. Decode may be shadowed + by execution of previous instructions*/ + decode_delay_offset++; + last_prefix = prefix; +} + +static void codegen_timing_cyrixiii_opcode(uint8_t opcode, uint32_t fetchdat, int op_32, uint32_t op_pc) +{ + uint32_t *timings; + uint64_t *deps; + int mod3 = ((fetchdat & 0xc0) == 0xc0); + int bit8 = !(opcode & 1); + int agi_stall = 0; + + switch (last_prefix) + { + case 0x0f: + timings = mod3 ? opcode_timings_0f_mod3 : opcode_timings_0f; + deps = mod3 ? opcode_deps_0f_mod3 : opcode_deps_0f; +// pclog("timings 0f\n"); + break; + + case 0xd8: + timings = mod3 ? opcode_timings_d8_mod3 : opcode_timings_d8; + deps = mod3 ? opcode_deps_d8_mod3 : opcode_deps_d8; + opcode = (opcode >> 3) & 7; +// pclog("timings d8\n"); + break; + case 0xd9: + timings = mod3 ? opcode_timings_d9_mod3 : opcode_timings_d9; + deps = mod3 ? opcode_deps_d9_mod3 : opcode_deps_d9; + opcode = mod3 ? opcode & 0x3f : (opcode >> 3) & 7; +// pclog("timings d9\n"); + break; + case 0xda: + timings = mod3 ? opcode_timings_da_mod3 : opcode_timings_da; + deps = mod3 ? opcode_deps_da_mod3 : opcode_deps_da; + opcode = (opcode >> 3) & 7; +// pclog("timings da\n"); + break; + case 0xdb: + timings = mod3 ? opcode_timings_db_mod3 : opcode_timings_db; + deps = mod3 ? opcode_deps_db_mod3 : opcode_deps_db; + opcode = mod3 ? opcode & 0x3f : (opcode >> 3) & 7; +// pclog("timings db\n"); + break; + case 0xdc: + timings = mod3 ? opcode_timings_dc_mod3 : opcode_timings_dc; + deps = mod3 ? opcode_deps_dc_mod3 : opcode_deps_dc; + opcode = (opcode >> 3) & 7; +// pclog("timings dc\n"); + break; + case 0xdd: + timings = mod3 ? opcode_timings_dd_mod3 : opcode_timings_dd; + deps = mod3 ? opcode_deps_dd_mod3 : opcode_deps_dd; + opcode = (opcode >> 3) & 7; +// pclog("timings dd\n"); + break; + case 0xde: + timings = mod3 ? opcode_timings_de_mod3 : opcode_timings_de; + deps = mod3 ? opcode_deps_de_mod3 : opcode_deps_de; + opcode = (opcode >> 3) & 7; +// pclog("timings de\n"); + break; + case 0xdf: + timings = mod3 ? opcode_timings_df_mod3 : opcode_timings_df; + deps = mod3 ? opcode_deps_df_mod3 : opcode_deps_df; + opcode = (opcode >> 3) & 7; +// pclog("timings df\n"); + break; + + default: + switch (opcode) + { + case 0x80: case 0x82: case 0x83: + timings = mod3 ? opcode_timings_8x_mod3 : opcode_timings_8x; + deps = mod3 ? opcode_deps_8x_mod3 : opcode_deps_8x_mod3; + opcode = (fetchdat >> 3) & 7; +// pclog("timings 80 %p %p %p\n", (void *)timings, (void *)opcode_timings_mod3, (void *)opcode_timings_8x); + break; + case 0x81: + timings = mod3 ? opcode_timings_81_mod3 : opcode_timings_81; + deps = mod3 ? opcode_deps_81_mod3 : opcode_deps_81; + opcode = (fetchdat >> 3) & 7; +// pclog("timings 80 %p %p %p\n", (void *)timings, (void *)opcode_timings_mod3, (void *)opcode_timings_8x); + break; + + case 0xc0: case 0xc1: case 0xd0: case 0xd1: case 0xd2: case 0xd3: + timings = mod3 ? opcode_timings_shift_mod3 : opcode_timings_shift; + deps = mod3 ? opcode_deps_shift_mod3 : opcode_deps_shift; + opcode = (fetchdat >> 3) & 7; +// pclog("timings c0\n"); + break; + + case 0xf6: + timings = mod3 ? opcode_timings_f6_mod3 : opcode_timings_f6; + deps = mod3 ? opcode_deps_f6_mod3 : opcode_deps_f6; + opcode = (fetchdat >> 3) & 7; +// pclog("timings f6\n"); + break; + case 0xf7: + timings = mod3 ? opcode_timings_f7_mod3 : opcode_timings_f7; + deps = mod3 ? opcode_deps_f7_mod3 : opcode_deps_f7; + opcode = (fetchdat >> 3) & 7; +// pclog("timings f7\n"); + break; + case 0xff: + timings = mod3 ? opcode_timings_ff_mod3 : opcode_timings_ff; + deps = mod3 ? opcode_deps_ff_mod3 : opcode_deps_ff; + opcode = (fetchdat >> 3) & 7; +// pclog("timings ff\n"); + break; + + default: + timings = mod3 ? opcode_timings_mod3 : opcode_timings; + deps = mod3 ? opcode_deps_mod3 : opcode_deps; +// pclog("timings normal\n"); + break; + } + } + + if (check_agi(deps, opcode, fetchdat, op_32)) + agi_stall = 1; + codegen_instruction(timings, deps, opcode, fetchdat, decode_delay_offset, op_32, agi_stall); + decode_delay_offset = 0; + regmask_modified = get_dstdep_mask(deps[opcode], fetchdat, bit8); +} + +static void codegen_timing_cyrixiii_block_end() +{ +} + +int codegen_timing_cyrixiii_jump_cycles() +{ + return 0; +} + +codegen_timing_t codegen_timing_cyrixiii = +{ + codegen_timing_cyrixiii_start, + codegen_timing_cyrixiii_prefix, + codegen_timing_cyrixiii_opcode, + codegen_timing_cyrixiii_block_start, + codegen_timing_cyrixiii_block_end, + codegen_timing_cyrixiii_jump_cycles +}; diff --git a/src/cpu.c b/src/cpu.c index 3335a3e..fb9a042 100644 --- a/src/cpu.c +++ b/src/cpu.c @@ -1106,6 +1106,44 @@ void cpu_set() codegen_timing_set(&codegen_timing_p6); break; + case CPU_CYRIX_III: + x86_setopcodes(ops_386, ops_winchip2_0f, dynarec_ops_386, dynarec_ops_winchip2_0f); + timing_rr = 1; /*register dest - register src*/ + timing_rm = 1; /*register dest - memory src*/ + timing_mr = 1; /*memory dest - register src*/ + timing_mm = 1; + timing_rml = 1; /*register dest - memory src long*/ + timing_mrl = 1; /*memory dest - register src long*/ + timing_mml = 1; + timing_bt = 1; /*branch taken*/ + timing_bnt = 1; /*branch not taken*/ + cpu_features = CPU_FEATURE_RDTSC | CPU_FEATURE_MMX | CPU_FEATURE_MSR | CPU_FEATURE_CR4 | CPU_FEATURE_3DNOW; + msr.fcr = (1 << 8) | (1 << 9) | (1 << 12) | (1 << 16) | (1 << 18) | (1 << 19) | (1 << 20) | (1 << 21); + cpu_CR4_mask = CR4_TSD | CR4_DE | CR4_MCE | CR4_PCE; + /*unknown*/ + timing_int_rm = 26; + timing_int_v86 = 82; + timing_int_pm = 44; + timing_int_pm_outer = 71; + timing_iret_rm = 7; + timing_iret_v86 = 26; + timing_iret_pm = 10; + timing_iret_pm_outer = 26; + timing_call_rm = 4; + timing_call_pm = 15; + timing_call_pm_gate = 26; + timing_call_pm_gate_inner = 35; + timing_retf_rm = 4; + timing_retf_pm = 7; + timing_retf_pm_outer = 23; + timing_jmp_rm = 5; + timing_jmp_pm = 7; + timing_jmp_pm_gate = 17; + timing_misaligned = 2; + cpu_cyrix_alignment = 1; + codegen_timing_set(&codegen_timing_cyrixiii); + break; + default: fatal("cpu_set : unknown CPU type %i\n", cpu_s->cpu_type); } @@ -1684,6 +1722,75 @@ void cpu_CPUID() else EAX = EBX = ECX = EDX = 0; break; + + case CPU_CYRIX_III: + switch (EAX) + { + case 0: + EAX = 1; + EBX = 0x746e6543; /*CentaurHauls*/ + ECX = 0x736c7561; + EDX = 0x48727561; + break; + case 1: + EAX = models[model].cpu[cpu_manufacturer].cpus[cpu].cpuid_model; + EBX = ECX = 0; + EDX = CPUID_FPU | CPUID_TSC | CPUID_MSR; + if (cpu_has_feature(CPU_FEATURE_CX8)) + EDX |= CPUID_CMPXCHG8B; + if (msr.fcr & (1 << 9)) + EDX |= CPUID_MMX; + break; + case 0x80000000: + EAX = 0x80000006; + break; + case 0x80000001: + EAX = models[model].cpu[cpu_manufacturer].cpus[cpu].edx_reset; + EDX = CPUID_FPU | CPUID_TSC | CPUID_MSR; + if (cpu_has_feature(CPU_FEATURE_CX8)) + EDX |= CPUID_CMPXCHG8B; + if (msr.fcr & (1 << 9)) + EDX |= CPUID_MMX; + if (cpu_has_feature(CPU_FEATURE_3DNOW)) + EDX |= CPUID_3DNOW; + break; + + case 0x80000002: /*Processor name string*/ + if (models[model].cpu[cpu_manufacturer].cpus[cpu].cpuid_model >= 0x670) + { + EAX = 0x20414956; /*VIA Samuel 2*/ + EBX = 0x756d6153; + ECX = 0x32206c65; + EDX = 0x00000000; + } + else + { + EAX = 0x20414956; /*VIA SamuelM*/ + EBX = 0x756d6153; + ECX = 0x004d6c65; + EDX = 0x00000000; + } + break; + + case 0x80000005: /*Cache information*/ + EBX = 0x08800880; /*TLBs*/ + ECX = 0x40040120; /*L1 data cache*/ + EDX = 0x40020120; /*L1 instruction cache*/ + break; + + case 0x80000006: /*L2 Cache information*/ + if (models[model].cpu[cpu_manufacturer].cpus[cpu].cpuid_model >= 0x670) + ECX = 0x40040120; /*L2 cache*/ + else + ECX = 0; + EAX = EBX = EDX = 0; + break; + + default: + EAX = EBX = ECX = EDX = 0; + break; + } + break; } } @@ -1693,6 +1800,7 @@ void cpu_RDMSR() { case CPU_WINCHIP: case CPU_WINCHIP2: + case CPU_CYRIX_III: EAX = EDX = 0; switch (ECX) { @@ -1794,6 +1902,7 @@ void cpu_WRMSR() { case CPU_WINCHIP: case CPU_WINCHIP2: + case CPU_CYRIX_III: switch (ECX) { case 0x02: diff --git a/src/cpu.h b/src/cpu.h index dad9d2b..ca8f4da 100644 --- a/src/cpu.h +++ b/src/cpu.h @@ -45,11 +45,13 @@ extern int fpu_type; #define CPU_PENTIUM_2 29 #define CPU_CELERON 30 #define CPU_CELERON_A 31 +#define CPU_CYRIX_III 32 #define MANU_INTEL 0 #define MANU_AMD 1 #define MANU_CYRIX 2 #define MANU_IDT 3 +#define MANU_VIA 4 extern int timing_rr; extern int timing_mr, timing_mrl; @@ -123,6 +125,7 @@ extern CPU cpus_K6_S7[]; extern CPU cpus_K6_SS7[]; extern CPU cpus_PentiumPro[]; extern CPU cpus_Slot1_100MHz[]; +extern CPU cpus_VIA_100MHz[]; extern CPU cpus_pcjr[]; extern CPU cpus_europc[]; diff --git a/src/cpu_tables.c b/src/cpu_tables.c index ac6f7be..c0bd920 100644 --- a/src/cpu_tables.c +++ b/src/cpu_tables.c @@ -535,3 +535,17 @@ CPU cpus_Slot1_100MHz[] = {"Celeron 533", CPU_CELERON_A, fpus_builtin, 20, 533333333, 3, 33333333, 0x660, 0x660, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC, 18,18,9,9, 68}, {"", -1, 0, 0, 0} }; + +CPU cpus_VIA_100MHz[] = +{ + /*VIA Cyrix III*/ + {"Cyrix III/500", CPU_CYRIX_III, fpus_builtin, 20, 500000000, 3, 33333333, 0x663, 0x663, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC, 18,18,9,9, 64}, +/*Below are disabled as Windows 98 won't boot. This is most likely a limitation of the timer system*/ +/* {"Cyrix III/550", CPU_CYRIX_III, fpus_builtin, 20, 550000000, 3, 33333333, 0x663, 0x663, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC, 18,18,9,9, 70}, + {"Cyrix III/600", CPU_CYRIX_III, fpus_builtin, 20, 600000000, 3, 33333333, 0x663, 0x663, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC, 18,18,9,9, 76}, + {"Cyrix III/650", CPU_CYRIX_III, fpus_builtin, 20, 650000000, 3, 33333333, 0x663, 0x663, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC, 18,18,9,9, 82}, + {"Cyrix III/700", CPU_CYRIX_III, fpus_builtin, 20, 700000000, 3, 33333333, 0x663, 0x663, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC, 18,18,9,9, 88}, + {"C3/750", CPU_CYRIX_III, fpus_builtin, 20, 750000000, 3, 33333333, 0x673, 0x673, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC, 18,18,9,9, 94}, + {"C3/800", CPU_CYRIX_III, fpus_builtin, 20, 800000000, 3, 33333333, 0x673, 0x673, 0, CPU_SUPPORTS_DYNAREC | CPU_REQUIRES_DYNAREC, 18,18,9,9, 100},*/ + {"", -1, 0, 0, 0} +}; diff --git a/src/model.c b/src/model.c index de53ef2..55e9177 100644 --- a/src/model.c +++ b/src/model.c @@ -243,7 +243,7 @@ MODEL models[] = {"[Socket 8] Intel VS440FX", ROM_VS440FX, "vs440fx", { {"Intel", cpus_PentiumPro}}, MODEL_GFX_NONE|MODEL_AT|MODEL_PCI|MODEL_PS2|MODEL_HAS_IDE, 8, 256, 8, at_vs440fx_init, NULL}, - {"[Slot 1] Gigabyte GA-686BX", ROM_GA686BX, "ga686bx", { {"Intel", cpus_Slot1_100MHz}}, MODEL_GFX_NONE|MODEL_AT|MODEL_PCI|MODEL_PS2|MODEL_HAS_IDE, 8, 512, 8, at_ga686bx_init, NULL}, + {"[Slot 1] Gigabyte GA-686BX", ROM_GA686BX, "ga686bx", { {"Intel", cpus_Slot1_100MHz}, {"VIA", cpus_VIA_100MHz}}, MODEL_GFX_NONE|MODEL_AT|MODEL_PCI|MODEL_PS2|MODEL_HAS_IDE, 8, 512, 8, at_ga686bx_init, NULL}, {"", -1, "", {{"", 0}, {"", 0}, {"", 0}}, 0,0,0, 0} };