claunia/86Box
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'upstream/master' into nec-v20

Browse Source
This commit is contained in:
Jasmine Iwanek
2022-10-19 16:24:10 -04:00
parent 6845e490be 005819c74f
commit 51f84ce0be
435 changed files with 83890 additions and 84115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

294
src/cpu/386.c
View File

@@ -6,7 +6,7 @@
#include <wchar.h>
#include <math.h>
#ifndef INFINITY
# define INFINITY (__builtin_inff())
# define INFINITY (__builtin_inff())
#endif
#define HAVE_STDARG_H
@@ -25,14 +25,12 @@
#include <86box/gdbstub.h>
#include "386_common.h"
#ifdef USE_NEW_DYNAREC
#include "codegen.h"
# include "codegen.h"
#endif
#undef CPU_BLOCK_END
#define CPU_BLOCK_END()
extern int codegen_flags_changed;
int tempc, oldcpl, optype, inttype, oddeven = 0;
@@ -41,225 +39,231 @@ int timetolive;
uint16_t oldcs;
uint32_t oldds, oldss, olddslimit, oldsslimit,
olddslimitw, oldsslimitw;
olddslimitw, oldsslimitw;
uint32_t oxpc;
uint32_t rmdat32;
uint32_t backupregs[16];
x86seg _oldds;
#ifdef ENABLE_386_LOG
int x386_do_log = ENABLE_386_LOG;
void
x386_log(const char *fmt, ...)
{
va_list ap;
if (x386_do_log) {
va_start(ap, fmt);
pclog_ex(fmt, ap);
va_end(ap);
va_start(ap, fmt);
pclog_ex(fmt, ap);
va_end(ap);
}
}
#else
#define x386_log(fmt, ...)
# define x386_log(fmt, ...)
#endif
#undef CPU_BLOCK_END
#define CPU_BLOCK_END()
#include "x86_flags.h"
#define getbytef() ((uint8_t)(fetchdat)); cpu_state.pc++
#define getwordf() ((uint16_t)(fetchdat)); cpu_state.pc+=2
#define getbyte2f() ((uint8_t)(fetchdat>>8)); cpu_state.pc++
#define getword2f() ((uint16_t)(fetchdat>>8)); cpu_state.pc+=2
/*
#define getbytef() \
((uint8_t) (fetchdat)); \
cpu_state.pc++
#define getwordf() \
((uint16_t) (fetchdat)); \
cpu_state.pc += 2
#define getbyte2f() \
((uint8_t) (fetchdat >> 8)); \
cpu_state.pc++
#define getword2f() \
((uint16_t) (fetchdat >> 8)); \
cpu_state.pc += 2
*/
#define OP_TABLE(name) ops_ ## name
#define OP_TABLE(name) ops_##name
#if 0
#define CLOCK_CYCLES(c) \
{\
if (fpu_cycles > 0) {\
fpu_cycles -= (c);\
if (fpu_cycles < 0) {\
cycles += fpu_cycles;\
}\
} else {\
cycles -= (c);\
}\
}
# define CLOCK_CYCLES(c) \
{ \
if (fpu_cycles > 0) { \
fpu_cycles -= (c); \
if (fpu_cycles < 0) { \
cycles += fpu_cycles; \
} \
} else { \
cycles -= (c); \
} \
}
#define CLOCK_CYCLES_FPU(c) cycles -= (c)
#define CONCURRENCY_CYCLES(c) fpu_cycles = (c)
# define CLOCK_CYCLES_FPU(c) cycles -= (c)
# define CONCURRENCY_CYCLES(c) fpu_cycles = (c)
#else
#define CLOCK_CYCLES(c) cycles -= (c)
#define CLOCK_CYCLES_FPU(c) cycles -= (c)
#define CONCURRENCY_CYCLES(c)
# define CLOCK_CYCLES(c) cycles -= (c)
# define CLOCK_CYCLES_FPU(c) cycles -= (c)
# define CONCURRENCY_CYCLES(c)
#endif
#define CLOCK_CYCLES_ALWAYS(c) cycles -= (c)
#include "x86_ops.h"
void
exec386(int cycs)
{
int vector, tempi, cycdiff, oldcyc;
int cycle_period, ins_cycles;
int vector, tempi, cycdiff, oldcyc;
int cycle_period, ins_cycles;
uint32_t addr;
cycles += cycs;
while (cycles > 0) {
cycle_period = (timer_target - (uint32_t)tsc) + 1;
cycle_period = (timer_target - (uint32_t) tsc) + 1;
x86_was_reset = 0;
cycdiff = 0;
oldcyc = cycles;
while (cycdiff < cycle_period) {
ins_cycles = cycles;
x86_was_reset = 0;
cycdiff = 0;
oldcyc = cycles;
while (cycdiff < cycle_period) {
ins_cycles = cycles;
#ifndef USE_NEW_DYNAREC
oldcs=CS;
oldcpl=CPL;
oldcs = CS;
oldcpl = CPL;
#endif
cpu_state.oldpc = cpu_state.pc;
cpu_state.op32 = use32;
cpu_state.oldpc = cpu_state.pc;
cpu_state.op32 = use32;
#ifndef USE_NEW_DYNAREC
x86_was_reset = 0;
x86_was_reset = 0;
#endif
cpu_state.ea_seg = &cpu_state.seg_ds;
cpu_state.ssegs = 0;
cpu_state.ea_seg = &cpu_state.seg_ds;
cpu_state.ssegs = 0;
fetchdat = fastreadl(cs + cpu_state.pc);
fetchdat = fastreadl(cs + cpu_state.pc);
if (!cpu_state.abrt) {
if (!cpu_state.abrt) {
#ifdef ENABLE_386_LOG
if (in_smm)
x386_log("[%04X:%08X] %08X\n", CS, cpu_state.pc, fetchdat);
if (in_smm)
x386_log("[%04X:%08X] %08X\n", CS, cpu_state.pc, fetchdat);
#endif
opcode = fetchdat & 0xFF;
fetchdat >>= 8;
trap = cpu_state.flags & T_FLAG;
opcode = fetchdat & 0xFF;
fetchdat >>= 8;
trap = cpu_state.flags & T_FLAG;
cpu_state.pc++;
x86_opcodes[(opcode | cpu_state.op32) & 0x3ff](fetchdat);
if (x86_was_reset)
break;
}
cpu_state.pc++;
x86_opcodes[(opcode | cpu_state.op32) & 0x3ff](fetchdat);
if (x86_was_reset)
break;
}
#ifdef ENABLE_386_LOG
else if (in_smm)
x386_log("[%04X:%08X] ABRT\n", CS, cpu_state.pc);
else if (in_smm)
x386_log("[%04X:%08X] ABRT\n", CS, cpu_state.pc);
#endif
#ifndef USE_NEW_DYNAREC
if (!use32) cpu_state.pc &= 0xffff;
if (!use32)
cpu_state.pc &= 0xffff;
#endif
if (cpu_end_block_after_ins)
cpu_end_block_after_ins--;
if (cpu_end_block_after_ins)
cpu_end_block_after_ins--;
if (cpu_state.abrt) {
flags_rebuild();
tempi = cpu_state.abrt & ABRT_MASK;
cpu_state.abrt = 0;
x86_doabrt(tempi);
if (cpu_state.abrt) {
cpu_state.abrt = 0;
if (cpu_state.abrt) {
flags_rebuild();
tempi = cpu_state.abrt & ABRT_MASK;
cpu_state.abrt = 0;
x86_doabrt(tempi);
if (cpu_state.abrt) {
cpu_state.abrt = 0;
#ifndef USE_NEW_DYNAREC
CS = oldcs;
CS = oldcs;
#endif
cpu_state.pc = cpu_state.oldpc;
x386_log("Double fault\n");
pmodeint(8, 0);
if (cpu_state.abrt) {
cpu_state.abrt = 0;
softresetx86();
cpu_set_edx();
cpu_state.pc = cpu_state.oldpc;
x386_log("Double fault\n");
pmodeint(8, 0);
if (cpu_state.abrt) {
cpu_state.abrt = 0;
softresetx86();
cpu_set_edx();
#ifdef ENABLE_386_LOG
x386_log("Triple fault - reset\n");
x386_log("Triple fault - reset\n");
#endif
}
}
}
}
}
}
if (smi_line)
enter_smm_check(0);
else if (trap) {
flags_rebuild();
dr[6] |= 0x4000;
if (msw&1)
pmodeint(1,0);
else {
writememw(ss, (SP - 2) & 0xFFFF, cpu_state.flags);
writememw(ss, (SP - 4) & 0xFFFF, CS);
writememw(ss, (SP - 6) & 0xFFFF, cpu_state.pc);
SP -= 6;
addr = (1 << 2) + idt.base;
cpu_state.flags &= ~I_FLAG;
cpu_state.flags &= ~T_FLAG;
cpu_state.pc = readmemw(0, addr);
loadcs(readmemw(0, addr + 2));
}
} else if (nmi && nmi_enable && nmi_mask) {
cpu_state.oldpc = cpu_state.pc;
x86_int(2);
nmi_enable = 0;
if (smi_line)
enter_smm_check(0);
else if (trap) {
flags_rebuild();
dr[6] |= 0x4000;
if (msw & 1)
pmodeint(1, 0);
else {
writememw(ss, (SP - 2) & 0xFFFF, cpu_state.flags);
writememw(ss, (SP - 4) & 0xFFFF, CS);
writememw(ss, (SP - 6) & 0xFFFF, cpu_state.pc);
SP -= 6;
addr = (1 << 2) + idt.base;
cpu_state.flags &= ~I_FLAG;
cpu_state.flags &= ~T_FLAG;
cpu_state.pc = readmemw(0, addr);
loadcs(readmemw(0, addr + 2));
}
} else if (nmi && nmi_enable && nmi_mask) {
cpu_state.oldpc = cpu_state.pc;
x86_int(2);
nmi_enable = 0;
#ifdef OLD_NMI_BEHAVIOR
if (nmi_auto_clear) {
nmi_auto_clear = 0;
nmi = 0;
}
if (nmi_auto_clear) {
nmi_auto_clear = 0;
nmi = 0;
}
#else
nmi = 0;
nmi = 0;
#endif
} else if ((cpu_state.flags & I_FLAG) && pic.int_pending && !cpu_end_block_after_ins) {
vector = picinterrupt();
if (vector != -1) {
flags_rebuild();
if (msw & 1)
pmodeint(vector, 0);
else {
writememw(ss, (SP - 2) & 0xFFFF, cpu_state.flags);
writememw(ss, (SP - 4) & 0xFFFF, CS);
writememw(ss, (SP - 6) & 0xFFFF, cpu_state.pc);
SP -= 6;
addr = (vector << 2) + idt.base;
cpu_state.flags &= ~I_FLAG;
cpu_state.flags &= ~T_FLAG;
cpu_state.pc = readmemw(0, addr);
loadcs(readmemw(0, addr + 2));
}
}
}
} else if ((cpu_state.flags & I_FLAG) && pic.int_pending && !cpu_end_block_after_ins) {
vector = picinterrupt();
if (vector != -1) {
flags_rebuild();
if (msw & 1)
pmodeint(vector, 0);
else {
writememw(ss, (SP - 2) & 0xFFFF, cpu_state.flags);
writememw(ss, (SP - 4) & 0xFFFF, CS);
writememw(ss, (SP - 6) & 0xFFFF, cpu_state.pc);
SP -= 6;
addr = (vector << 2) + idt.base;
cpu_state.flags &= ~I_FLAG;
cpu_state.flags &= ~T_FLAG;
cpu_state.pc = readmemw(0, addr);
loadcs(readmemw(0, addr + 2));
}
}
}
ins_cycles -= cycles;
tsc += ins_cycles;
ins_cycles -= cycles;
tsc += ins_cycles;
cycdiff = oldcyc - cycles;
cycdiff = oldcyc - cycles;
if (timetolive) {
timetolive--;
if (!timetolive)
fatal("Life expired\n");
}
if (timetolive) {
timetolive--;
if (!timetolive)
fatal("Life expired\n");
}
if (TIMER_VAL_LESS_THAN_VAL(timer_target, (uint32_t) tsc))
timer_process_inline();
if (TIMER_VAL_LESS_THAN_VAL(timer_target, (uint32_t) tsc))
timer_process_inline();
#ifdef USE_GDBSTUB
if (gdbstub_instruction())
return;
if (gdbstub_instruction())
return;
#endif
}
}
}
}

4672
src/cpu/808x.c
View File

File diff suppressed because it is too large Load Diff

7
src/cpu/codegen_timing_p6.c
View File

@@ -158,10 +158,13 @@ static const macro_op_t lods_op =
};
static const macro_op_t loop_op =
{
.nr_uops = 2,
.nr_uops = 5,
.decode_type = DECODE_COMPLEX,
.uop[0] = {.type = UOP_ALU, .latency = 1},
.uop[1] = {.type = UOP_BRANCH, .latency = 2}
.uop[1] = {.type = UOP_ALU, .latency = 1},
.uop[2] = {.type = UOP_ALU, .latency = 1},
.uop[3] = {.type = UOP_ALU, .latency = 1},
.uop[4] = {.type = UOP_BRANCH, .latency = 1}
};
static const macro_op_t mov_reg_seg_op =
{

5028
src/cpu/cpu.c
View File

File diff suppressed because it is too large Load Diff

886
src/cpu/cpu.h
View File

File diff suppressed because it is too large Load Diff

882
src/cpu/x87_timings.c
View File

@@ -10,80 +10,79 @@
x87_timings_t x87_timings;
x87_timings_t x87_concurrency;
const x87_timings_t x87_timings_8087 =
{
.f2xm1 = (310 + 630) / 2,
.fabs = (10 + 17) / 2,
.fadd = (70 + 100) / 2,
.fadd_32 = (90 + 120) / 2,
.fadd_64 = (95 + 125) / 2,
.fbld = (290 + 310) / 2,
.fbstp = (520 + 540) / 2,
.fchs = (10 + 17) / 2,
.fclex = (2 + 8) / 2,
.fcom = (40 + 50) / 2,
.fcom_32 = (60 + 70) / 2,
.fcom_64 = (65 + 75) / 2,
.fcos = 0, /*387+*/
.fincdecstp = (6 + 12) / 2,
.fdisi_eni = (6 + 12) / 2,
.fdiv = (193 + 203) / 2,
.fdiv_32 = (215 + 225) / 2,
.fdiv_64 = (220 + 230) / 2,
.ffree = (9 + 16) / 2,
.fadd_i16 = (102 + 137) / 2,
.fadd_i32 = (108 + 143) / 2,
.fcom_i16 = (72 + 86) / 2,
.fcom_i32 = (78 + 91) / 2,
.fdiv_i16 = (224 + 238) / 2,
.fdiv_i32 = (230 + 243) / 2,
.fild_16 = (46 + 54) / 2,
.fild_32 = (50 + 60) / 2,
.fild_64 = (60 + 68) / 2,
.fmul_i16 = (124 + 138) / 2,
.fmul_i32 = (130 + 144) / 2,
.finit = (2 + 8) / 2,
.fist_16 = (80 + 90) / 2,
.fist_32 = (82 + 92) / 2,
.fist_64 = (94 + 105) / 2,
.fld = (17 + 22) / 2,
.fld_32 = (38 + 56) / 2,
.fld_64 = (40 + 60) / 2,
.fld_80 = (53 + 65) / 2,
.fld_z1 = (11 + 21) / 2,
.fld_const = (15 + 24) / 2,
.fldcw = (7 + 14) / 2,
.fldenv = (35 + 45) / 2,
.fmul = (90 + 145) / 2,
.fmul_32 = (110 + 125) / 2,
.fmul_64 = (154 + 168) / 2,
.fnop = (10 + 16) / 2,
.fpatan = (250 + 800) / 2,
.fprem = (15 + 190) / 2,
.fprem1 = 0, /*387+*/
.fptan = (30 + 540) / 2,
.frndint = (16 + 50) / 2,
.frstor = (197 + 207) / 2,
.fsave = (197 + 207) / 2,
.fscale = (32 + 38) / 2,
.fsetpm = 0, /*287+*/
.fsin_cos = 0, /*387+*/
.fsincos = 0, /*387+*/
.fsqrt = (180 + 186) / 2,
.fst = (15 + 22) / 2,
.fst_32 = (84 + 90) / 2,
.fst_64 = (96 + 104) / 2,
.fst_80 = (52 + 58) / 2,
.fstcw_sw = (12 + 18) / 2,
.fstenv = (40 + 50) / 2,
.ftst = (38 + 48) / 2,
.fucom = 0, /*387+*/
.fwait = 4,
.fxam = (12 + 23) / 2,
.fxch = (10 + 15) / 2,
.fxtract = (27 + 55) / 2,
.fyl2x = (900 + 1100) / 2,
.fyl2xp1 = (700 + 1000) / 2
const x87_timings_t x87_timings_8087 = {
.f2xm1 = (310 + 630) / 2,
.fabs = (10 + 17) / 2,
.fadd = (70 + 100) / 2,
.fadd_32 = (90 + 120) / 2,
.fadd_64 = (95 + 125) / 2,
.fbld = (290 + 310) / 2,
.fbstp = (520 + 540) / 2,
.fchs = (10 + 17) / 2,
.fclex = (2 + 8) / 2,
.fcom = (40 + 50) / 2,
.fcom_32 = (60 + 70) / 2,
.fcom_64 = (65 + 75) / 2,
.fcos = 0, /*387+*/
.fincdecstp = (6 + 12) / 2,
.fdisi_eni = (6 + 12) / 2,
.fdiv = (193 + 203) / 2,
.fdiv_32 = (215 + 225) / 2,
.fdiv_64 = (220 + 230) / 2,
.ffree = (9 + 16) / 2,
.fadd_i16 = (102 + 137) / 2,
.fadd_i32 = (108 + 143) / 2,
.fcom_i16 = (72 + 86) / 2,
.fcom_i32 = (78 + 91) / 2,
.fdiv_i16 = (224 + 238) / 2,
.fdiv_i32 = (230 + 243) / 2,
.fild_16 = (46 + 54) / 2,
.fild_32 = (50 + 60) / 2,
.fild_64 = (60 + 68) / 2,
.fmul_i16 = (124 + 138) / 2,
.fmul_i32 = (130 + 144) / 2,
.finit = (2 + 8) / 2,
.fist_16 = (80 + 90) / 2,
.fist_32 = (82 + 92) / 2,
.fist_64 = (94 + 105) / 2,
.fld = (17 + 22) / 2,
.fld_32 = (38 + 56) / 2,
.fld_64 = (40 + 60) / 2,
.fld_80 = (53 + 65) / 2,
.fld_z1 = (11 + 21) / 2,
.fld_const = (15 + 24) / 2,
.fldcw = (7 + 14) / 2,
.fldenv = (35 + 45) / 2,
.fmul = (90 + 145) / 2,
.fmul_32 = (110 + 125) / 2,
.fmul_64 = (154 + 168) / 2,
.fnop = (10 + 16) / 2,
.fpatan = (250 + 800) / 2,
.fprem = (15 + 190) / 2,
.fprem1 = 0, /*387+*/
.fptan = (30 + 540) / 2,
.frndint = (16 + 50) / 2,
.frstor = (197 + 207) / 2,
.fsave = (197 + 207) / 2,
.fscale = (32 + 38) / 2,
.fsetpm = 0, /*287+*/
.fsin_cos = 0, /*387+*/
.fsincos = 0, /*387+*/
.fsqrt = (180 + 186) / 2,
.fst = (15 + 22) / 2,
.fst_32 = (84 + 90) / 2,
.fst_64 = (96 + 104) / 2,
.fst_80 = (52 + 58) / 2,
.fstcw_sw = (12 + 18) / 2,
.fstenv = (40 + 50) / 2,
.ftst = (38 + 48) / 2,
.fucom = 0, /*387+*/
.fwait = 4,
.fxam = (12 + 23) / 2,
.fxch = (10 + 15) / 2,
.fxtract = (27 + 55) / 2,
.fyl2x = (900 + 1100) / 2,
.fyl2xp1 = (700 + 1000) / 2
};
const x87_timings_t x87_timings_80187 =
@@ -163,384 +162,379 @@ const x87_timings_t x87_timings_80187 =
};
/*Mostly the same as 8087*/
const x87_timings_t x87_timings_287 =
{
.f2xm1 = (310 + 630) / 2,
.fabs = (10 + 17) / 2,
.fadd = (70 + 100) / 2,
.fadd_32 = (90 + 120) / 2,
.fadd_64 = (95 + 125) / 2,
.fbld = (290 + 310) / 2,
.fbstp = (520 + 540) / 2,
.fchs = (10 + 17) / 2,
.fclex = (2 + 8) / 2,
.fcom = (40 + 50) / 2,
.fcom_32 = (60 + 70) / 2,
.fcom_64 = (65 + 75) / 2,
.fcos = 0, /*387+*/
.fincdecstp = (6 + 12) / 2,
.fdisi_eni = 2,
.fdiv = (193 + 203) / 2,
.fdiv_32 = (215 + 225) / 2,
.fdiv_64 = (220 + 230) / 2,
.ffree = (9 + 16) / 2,
.fadd_i16 = (102 + 137) / 2,
.fadd_i32 = (108 + 143) / 2,
.fcom_i16 = (72 + 86) / 2,
.fcom_i32 = (78 + 91) / 2,
.fdiv_i16 = (224 + 238) / 2,
.fdiv_i32 = (230 + 243) / 2,
.fild_16 = (46 + 54) / 2,
.fild_32 = (50 + 60) / 2,
.fild_64 = (60 + 68) / 2,
.fmul_i16 = (124 + 138) / 2,
.fmul_i32 = (130 + 144) / 2,
.finit = (2 + 8) / 2,
.fist_16 = (80 + 90) / 2,
.fist_32 = (82 + 92) / 2,
.fist_64 = (94 + 105) / 2,
.fld = (17 + 22) / 2,
.fld_32 = (38 + 56) / 2,
.fld_64 = (40 + 60) / 2,
.fld_80 = (53 + 65) / 2,
.fld_z1 = (11 + 21) / 2,
.fld_const = (15 + 24) / 2,
.fldcw = (7 + 14) / 2,
.fldenv = (35 + 45) / 2,
.fmul = (90 + 145) / 2,
.fmul_32 = (110 + 125) / 2,
.fmul_64 = (154 + 168) / 2,
.fnop = (10 + 16) / 2,
.fpatan = (250 + 800) / 2,
.fprem = (15 + 190) / 2,
.fprem1 = 0, /*387+*/
.fptan = (30 + 540) / 2,
.frndint = (16 + 50) / 2,
.frstor = (197 + 207) / 2,
.fsave = (197 + 207) / 2,
.fscale = (32 + 38) / 2,
.fsetpm = (2 + 8) / 2, /*287+*/
.fsin_cos = 0, /*387+*/
.fsincos = 0, /*387+*/
.fsqrt = (180 + 186) / 2,
.fst = (15 + 22) / 2,
.fst_32 = (84 + 90) / 2,
.fst_64 = (96 + 104) / 2,
.fst_80 = (52 + 58) / 2,
.fstcw_sw = (12 + 18) / 2,
.fstenv = (40 + 50) / 2,
.ftst = (38 + 48) / 2,
.fucom = 0, /*387+*/
.fwait = 3,
.fxam = (12 + 23) / 2,
.fxch = (10 + 15) / 2,
.fxtract = (27 + 55) / 2,
.fyl2x = (900 + 1100) / 2,
.fyl2xp1 = (700 + 1000) / 2
const x87_timings_t x87_timings_287 = {
.f2xm1 = (310 + 630) / 2,
.fabs = (10 + 17) / 2,
.fadd = (70 + 100) / 2,
.fadd_32 = (90 + 120) / 2,
.fadd_64 = (95 + 125) / 2,
.fbld = (290 + 310) / 2,
.fbstp = (520 + 540) / 2,
.fchs = (10 + 17) / 2,
.fclex = (2 + 8) / 2,
.fcom = (40 + 50) / 2,
.fcom_32 = (60 + 70) / 2,
.fcom_64 = (65 + 75) / 2,
.fcos = 0, /*387+*/
.fincdecstp = (6 + 12) / 2,
.fdisi_eni = 2,
.fdiv = (193 + 203) / 2,
.fdiv_32 = (215 + 225) / 2,
.fdiv_64 = (220 + 230) / 2,
.ffree = (9 + 16) / 2,
.fadd_i16 = (102 + 137) / 2,
.fadd_i32 = (108 + 143) / 2,
.fcom_i16 = (72 + 86) / 2,
.fcom_i32 = (78 + 91) / 2,
.fdiv_i16 = (224 + 238) / 2,
.fdiv_i32 = (230 + 243) / 2,
.fild_16 = (46 + 54) / 2,
.fild_32 = (50 + 60) / 2,
.fild_64 = (60 + 68) / 2,
.fmul_i16 = (124 + 138) / 2,
.fmul_i32 = (130 + 144) / 2,
.finit = (2 + 8) / 2,
.fist_16 = (80 + 90) / 2,
.fist_32 = (82 + 92) / 2,
.fist_64 = (94 + 105) / 2,
.fld = (17 + 22) / 2,
.fld_32 = (38 + 56) / 2,
.fld_64 = (40 + 60) / 2,
.fld_80 = (53 + 65) / 2,
.fld_z1 = (11 + 21) / 2,
.fld_const = (15 + 24) / 2,
.fldcw = (7 + 14) / 2,
.fldenv = (35 + 45) / 2,
.fmul = (90 + 145) / 2,
.fmul_32 = (110 + 125) / 2,
.fmul_64 = (154 + 168) / 2,
.fnop = (10 + 16) / 2,
.fpatan = (250 + 800) / 2,
.fprem = (15 + 190) / 2,
.fprem1 = 0, /*387+*/
.fptan = (30 + 540) / 2,
.frndint = (16 + 50) / 2,
.frstor = (197 + 207) / 2,
.fsave = (197 + 207) / 2,
.fscale = (32 + 38) / 2,
.fsetpm = (2 + 8) / 2, /*287+*/
.fsin_cos = 0, /*387+*/
.fsincos = 0, /*387+*/
.fsqrt = (180 + 186) / 2,
.fst = (15 + 22) / 2,
.fst_32 = (84 + 90) / 2,
.fst_64 = (96 + 104) / 2,
.fst_80 = (52 + 58) / 2,
.fstcw_sw = (12 + 18) / 2,
.fstenv = (40 + 50) / 2,
.ftst = (38 + 48) / 2,
.fucom = 0, /*387+*/
.fwait = 3,
.fxam = (12 + 23) / 2,
.fxch = (10 + 15) / 2,
.fxtract = (27 + 55) / 2,
.fyl2x = (900 + 1100) / 2,
.fyl2xp1 = (700 + 1000) / 2
};
const x87_timings_t x87_timings_387 =
{
.f2xm1 = (211 + 476) / 2,
.fabs = 22,
.fadd = (23 + 34) / 2,
.fadd_32 = (24 + 32) / 2,
.fadd_64 = (29 + 37) / 2,
.fbld = (266 + 275) / 2,
.fbstp = (512 + 534) / 2,
.fchs = (24 + 25) / 2,
.fclex = 11,
.fcom = 24,
.fcom_32 = 26,
.fcom_64 = 31,
.fcos = (122 + 772) / 2,
.fincdecstp = 22,
.fdisi_eni = 2,
.fdiv = (88 + 91) / 2,
.fdiv_32 = 89,
.fdiv_64 = 94,
.ffree = 18,
.fadd_i16 = (71 + 85) / 2,
.fadd_i32 = (57 + 72) / 2,
.fcom_i16 = (71 + 75) / 2,
.fcom_i32 = (56 + 63) / 2,
.fdiv_i16 = (136 + 140) / 2,
.fdiv_i32 = (120 + 127) / 2,
.fild_16 = (61 + 65) / 2,
.fild_32 = (45 + 52) / 2,
.fild_64 = (56 + 67) / 2,
.fmul_i16 = (76 + 87) / 2,
.fmul_i32 = (61 + 82) / 2,
.finit = 33,
.fist_16 = (82 + 95) / 2,
.fist_32 = (79 + 93) / 2,
.fist_64 = (80 + 97) / 2,
.fld = 14,
.fld_32 = 20,
.fld_64 = 25,
.fld_80 = 44,
.fld_z1 = (20 + 24) / 2,
.fld_const = 40,
.fldcw = 19,
.fldenv = 71,
.fmul = (29 + 57) / 2,
.fmul_32 = (27 + 35) / 2,
.fmul_64 = (32 + 57) / 2,
.fnop = 12,
.fpatan = (314 + 487) / 2,
.fprem = (74 + 155) / 2,
.fprem1 = (95 + 185) / 2,
.fptan = (191 + 497) / 2,
.frndint = (66 + 80) / 2,
.frstor = 308,
.fsave = 375,
.fscale = (67 + 86) / 2,
.fsetpm = 12,
.fsin_cos = (122 + 771) / 2,
.fsincos = (194 + 809) / 2,
.fsqrt = (122 + 129) / 2,
.fst = 11,
.fst_32 = 44,
.fst_64 = 45,
.fst_80 = 53,
.fstcw_sw = 15,
.fstenv = 103,
.ftst = 28,
.fucom = 24,
.fwait = 6,
.fxam = (30 + 38) / 2,
.fxch = 18,
.fxtract = (70 + 76) / 2,
.fyl2x = (120 + 538) / 2,
.fyl2xp1 = (257 + 547) / 2
const x87_timings_t x87_timings_387 = {
.f2xm1 = (211 + 476) / 2,
.fabs = 22,
.fadd = (23 + 34) / 2,
.fadd_32 = (24 + 32) / 2,
.fadd_64 = (29 + 37) / 2,
.fbld = (266 + 275) / 2,
.fbstp = (512 + 534) / 2,
.fchs = (24 + 25) / 2,
.fclex = 11,
.fcom = 24,
.fcom_32 = 26,
.fcom_64 = 31,
.fcos = (122 + 772) / 2,
.fincdecstp = 22,
.fdisi_eni = 2,
.fdiv = (88 + 91) / 2,
.fdiv_32 = 89,
.fdiv_64 = 94,
.ffree = 18,
.fadd_i16 = (71 + 85) / 2,
.fadd_i32 = (57 + 72) / 2,
.fcom_i16 = (71 + 75) / 2,
.fcom_i32 = (56 + 63) / 2,
.fdiv_i16 = (136 + 140) / 2,
.fdiv_i32 = (120 + 127) / 2,
.fild_16 = (61 + 65) / 2,
.fild_32 = (45 + 52) / 2,
.fild_64 = (56 + 67) / 2,
.fmul_i16 = (76 + 87) / 2,
.fmul_i32 = (61 + 82) / 2,
.finit = 33,
.fist_16 = (82 + 95) / 2,
.fist_32 = (79 + 93) / 2,
.fist_64 = (80 + 97) / 2,
.fld = 14,
.fld_32 = 20,
.fld_64 = 25,
.fld_80 = 44,
.fld_z1 = (20 + 24) / 2,
.fld_const = 40,
.fldcw = 19,
.fldenv = 71,
.fmul = (29 + 57) / 2,
.fmul_32 = (27 + 35) / 2,
.fmul_64 = (32 + 57) / 2,
.fnop = 12,
.fpatan = (314 + 487) / 2,
.fprem = (74 + 155) / 2,
.fprem1 = (95 + 185) / 2,
.fptan = (191 + 497) / 2,
.frndint = (66 + 80) / 2,
.frstor = 308,
.fsave = 375,
.fscale = (67 + 86) / 2,
.fsetpm = 12,
.fsin_cos = (122 + 771) / 2,
.fsincos = (194 + 809) / 2,
.fsqrt = (122 + 129) / 2,
.fst = 11,
.fst_32 = 44,
.fst_64 = 45,
.fst_80 = 53,
.fstcw_sw = 15,
.fstenv = 103,
.ftst = 28,
.fucom = 24,
.fwait = 6,
.fxam = (30 + 38) / 2,
.fxch = 18,
.fxtract = (70 + 76) / 2,
.fyl2x = (120 + 538) / 2,
.fyl2xp1 = (257 + 547) / 2
};
const x87_timings_t x87_timings_486 =
{
.f2xm1 = (140 + 270) / 2,
.fabs = 3,
.fadd = (8 + 20) / 2,
.fadd_32 = (8 + 20) / 2,
.fadd_64 = (8 + 20) / 2,
.fbld = (70 + 103) / 2,
.fbstp = (172 + 176) / 2,
.fchs = 6,
.fclex = 7,
.fcom = 4,
.fcom_32 = 4,
.fcom_64 = 4,
.fcos = (257 + 354) / 2,
.fincdecstp = 3,
.fdisi_eni = 3,
.fdiv = 73,
.fdiv_32 = 73,
.fdiv_64 = 73,
.ffree = 3,
.fadd_i16 = (20 + 35) / 2,
.fadd_i32 = (19 + 32) / 2,
.fcom_i16 = (16 + 20) / 2,
.fcom_i32 = (15 + 17) / 2,
.fdiv_i16 = (85 + 89) / 2,
.fdiv_i32 = (84 + 86) / 2,
.fild_16 = (13 + 16) / 2,
.fild_32 = (9 + 12) / 2,
.fild_64 = (10 + 18) / 2,
.fmul_i16 = (23 + 27) / 2,
.fmul_i32 = (22 + 24) / 2,
.finit = 17,
.fist_16 = (29 + 34) / 2,
.fist_32 = (28 + 34) / 2,
.fist_64 = (29 + 34) / 2,
.fld = 4,
.fld_32 = 3,
.fld_64 = 3,
.fld_80 = 6,
.fld_z1 = 4,
.fld_const = 8,
.fldcw = 4,
.fldenv = 34,
.fmul = 16,
.fmul_32 = 11,
.fmul_64 = 14,
.fnop = 3,
.fpatan = (218 + 303) / 2,
.fprem = (70 + 138) / 2,
.fprem1 = (72 + 167) / 2,
.fptan = (200 + 273) / 2,
.frndint = (21 + 30) / 2,
.frstor = 120,
.fsave = 143,
.fscale = (30 + 32) / 2,
.fsetpm = 3,
.fsin_cos = (257 + 354) / 2,
.fsincos = (292 + 365) / 2,
.fsqrt = (83 + 87) / 2,
.fst = 3,
.fst_32 = 7,
.fst_64 = 8,
.fst_80 = 6,
.fstcw_sw = 3,
.fstenv = 56,
.ftst = 4,
.fucom = 4,
.fwait = (1 + 3) / 2,
.fxam = 8,
.fxch = 4,
.fxtract = (16 + 20) / 2,
.fyl2x = (196 + 329) / 2,
.fyl2xp1 = (171 + 326) / 2
const x87_timings_t x87_timings_486 = {
.f2xm1 = (140 + 270) / 2,
.fabs = 3,
.fadd = (8 + 20) / 2,
.fadd_32 = (8 + 20) / 2,
.fadd_64 = (8 + 20) / 2,
.fbld = (70 + 103) / 2,
.fbstp = (172 + 176) / 2,
.fchs = 6,
.fclex = 7,
.fcom = 4,
.fcom_32 = 4,
.fcom_64 = 4,
.fcos = (257 + 354) / 2,
.fincdecstp = 3,
.fdisi_eni = 3,
.fdiv = 73,
.fdiv_32 = 73,
.fdiv_64 = 73,
.ffree = 3,
.fadd_i16 = (20 + 35) / 2,
.fadd_i32 = (19 + 32) / 2,
.fcom_i16 = (16 + 20) / 2,
.fcom_i32 = (15 + 17) / 2,
.fdiv_i16 = (85 + 89) / 2,
.fdiv_i32 = (84 + 86) / 2,
.fild_16 = (13 + 16) / 2,
.fild_32 = (9 + 12) / 2,
.fild_64 = (10 + 18) / 2,
.fmul_i16 = (23 + 27) / 2,
.fmul_i32 = (22 + 24) / 2,
.finit = 17,
.fist_16 = (29 + 34) / 2,
.fist_32 = (28 + 34) / 2,
.fist_64 = (29 + 34) / 2,
.fld = 4,
.fld_32 = 3,
.fld_64 = 3,
.fld_80 = 6,
.fld_z1 = 4,
.fld_const = 8,
.fldcw = 4,
.fldenv = 34,
.fmul = 16,
.fmul_32 = 11,
.fmul_64 = 14,
.fnop = 3,
.fpatan = (218 + 303) / 2,
.fprem = (70 + 138) / 2,
.fprem1 = (72 + 167) / 2,
.fptan = (200 + 273) / 2,
.frndint = (21 + 30) / 2,
.frstor = 120,
.fsave = 143,
.fscale = (30 + 32) / 2,
.fsetpm = 3,
.fsin_cos = (257 + 354) / 2,
.fsincos = (292 + 365) / 2,
.fsqrt = (83 + 87) / 2,
.fst = 3,
.fst_32 = 7,
.fst_64 = 8,
.fst_80 = 6,
.fstcw_sw = 3,
.fstenv = 56,
.ftst = 4,
.fucom = 4,
.fwait = (1 + 3) / 2,
.fxam = 8,
.fxch = 4,
.fxtract = (16 + 20) / 2,
.fyl2x = (196 + 329) / 2,
.fyl2xp1 = (171 + 326) / 2
};
/* this should be used for FPUs with no concurrency.
some pre-486DX Cyrix FPUs reportedly are like this. */
const x87_timings_t x87_concurrency_none =
{
.f2xm1 = 0,
.fabs = 0,
.fadd = 0,
.fadd_32 = 0,
.fadd_64 = 0,
.fbld = 0,
.fbstp = 0,
.fchs = 0,
.fclex = 0,
.fcom = 0,
.fcom_32 = 0,
.fcom_64 = 0,
.fcos = 0,
.fincdecstp = 0,
.fdisi_eni = 0,
.fdiv = 0,
.fdiv_32 = 0,
.fdiv_64 = 0,
.ffree = 0,
.fadd_i16 = 0,
.fadd_i32 = 0,
.fcom_i16 = 0,
.fcom_i32 = 0,
.fdiv_i16 = 0,
.fdiv_i32 = 0,
.fild_16 = 0,
.fild_32 = 0,
.fild_64 = 0,
.fmul_i16 = 0,
.fmul_i32 = 0,
.finit = 0,
.fist_16 = 0,
.fist_32 = 0,
.fist_64 = 0,
.fld = 0,
.fld_32 = 0,
.fld_64 = 0,
.fld_80 = 0,
.fld_z1 = 0,
.fld_const = 0,
.fldcw = 0,
.fldenv = 0,
.fmul = 0,
.fmul_32 = 0,
.fmul_64 = 0,
.fnop = 0,
.fpatan = 0,
.fprem = 0,
.fprem1 = 0,
.fptan = 0,
.frndint = 0,
.frstor = 0,
.fsave = 0,
.fscale = 0,
.fsetpm = 0,
.fsin_cos = 0,
.fsincos = 0,
.fsqrt = 0,
.fst = 0,
.fst_32 = 0,
.fst_64 = 0,
.fst_80 = 0,
.fstcw_sw = 0,
.fstenv = 0,
.ftst = 0,
.fucom = 0,
.fwait = 0,
.fxam = 0,
.fxch = 0,
.fxtract = 0,
.fyl2x = 0,
.fyl2xp1 = 0,
const x87_timings_t x87_concurrency_none = {
.f2xm1 = 0,
.fabs = 0,
.fadd = 0,
.fadd_32 = 0,
.fadd_64 = 0,
.fbld = 0,
.fbstp = 0,
.fchs = 0,
.fclex = 0,
.fcom = 0,
.fcom_32 = 0,
.fcom_64 = 0,
.fcos = 0,
.fincdecstp = 0,
.fdisi_eni = 0,
.fdiv = 0,
.fdiv_32 = 0,
.fdiv_64 = 0,
.ffree = 0,
.fadd_i16 = 0,
.fadd_i32 = 0,
.fcom_i16 = 0,
.fcom_i32 = 0,
.fdiv_i16 = 0,
.fdiv_i32 = 0,
.fild_16 = 0,
.fild_32 = 0,
.fild_64 = 0,
.fmul_i16 = 0,
.fmul_i32 = 0,
.finit = 0,
.fist_16 = 0,
.fist_32 = 0,
.fist_64 = 0,
.fld = 0,
.fld_32 = 0,
.fld_64 = 0,
.fld_80 = 0,
.fld_z1 = 0,
.fld_const = 0,
.fldcw = 0,
.fldenv = 0,
.fmul = 0,
.fmul_32 = 0,
.fmul_64 = 0,
.fnop = 0,
.fpatan = 0,
.fprem = 0,
.fprem1 = 0,
.fptan = 0,
.frndint = 0,
.frstor = 0,
.fsave = 0,
.fscale = 0,
.fsetpm = 0,
.fsin_cos = 0,
.fsincos = 0,
.fsqrt = 0,
.fst = 0,
.fst_32 = 0,
.fst_64 = 0,
.fst_80 = 0,
.fstcw_sw = 0,
.fstenv = 0,
.ftst = 0,
.fucom = 0,
.fwait = 0,
.fxam = 0,
.fxch = 0,
.fxtract = 0,
.fyl2x = 0,
.fyl2xp1 = 0,
};
const x87_timings_t x87_concurrency_486 =
{
.f2xm1 = 2,
.fabs = 0,
.fadd = 7,
.fadd_32 = 7,
.fadd_64 = 7,
.fbld = 8,
.fbstp = 0,
.fchs = 0,
.fclex = 0,
.fcom = 1,
.fcom_32 = 1,
.fcom_64 = 1,
.fcos = 2,
.fincdecstp = 0,
.fdisi_eni = 0,
.fdiv = 70,
.fdiv_32 = 70,
.fdiv_64 = 70,
.ffree = 0,
.fadd_i16 = 7,
.fadd_i32 = 7,
.fcom_i16 = 1,
.fcom_i32 = 1,
.fdiv_i16 = 70,
.fdiv_i32 = 70,
.fild_16 = 4,
.fild_32 = 4,
.fild_64 = 8,
.fmul_i16 = 8,
.fmul_i32 = 8,
.finit = 0,
.fist_16 = 0,
.fist_32 = 0,
.fist_64 = 0,
.fld = 0,
.fld_32 = 0,
.fld_64 = 0,
.fld_80 = 0,
.fld_z1 = 0,
.fld_const = 2,
.fldcw = 0,
.fldenv = 0,
.fmul = 13,
.fmul_32 = 8,
.fmul_64 = 11,
.fnop = 0,
.fpatan = 5,
.fprem = 2,
.fprem1 = 6,
.fptan = 70,
.frndint = 0,
.frstor = 0,
.fsave = 0,
.fscale = 2,
.fsetpm = 0,
.fsin_cos = 2,
.fsincos = 2,
.fsqrt = 70,
.fst = 0,
.fst_32 = 0,
.fst_64 = 0,
.fst_80 = 0,
.fstcw_sw = 0,
.fstenv = 0,
.ftst = 1,
.fucom = 1,
.fwait = 0,
.fxam = 0,
.fxch = 0,
.fxtract = 4,
.fyl2x = 13,
.fyl2xp1 = 13,
const x87_timings_t x87_concurrency_486 = {
.f2xm1 = 2,
.fabs = 0,
.fadd = 7,
.fadd_32 = 7,
.fadd_64 = 7,
.fbld = 8,
.fbstp = 0,
.fchs = 0,
.fclex = 0,
.fcom = 1,
.fcom_32 = 1,
.fcom_64 = 1,
.fcos = 2,
.fincdecstp = 0,
.fdisi_eni = 0,
.fdiv = 70,
.fdiv_32 = 70,
.fdiv_64 = 70,
.ffree = 0,
.fadd_i16 = 7,
.fadd_i32 = 7,
.fcom_i16 = 1,
.fcom_i32 = 1,
.fdiv_i16 = 70,
.fdiv_i32 = 70,
.fild_16 = 4,
.fild_32 = 4,
.fild_64 = 8,
.fmul_i16 = 8,
.fmul_i32 = 8,
.finit = 0,
.fist_16 = 0,
.fist_32 = 0,
.fist_64 = 0,
.fld = 0,
.fld_32 = 0,
.fld_64 = 0,
.fld_80 = 0,
.fld_z1 = 0,
.fld_const = 2,
.fldcw = 0,
.fldenv = 0,
.fmul = 13,
.fmul_32 = 8,
.fmul_64 = 11,
.fnop = 0,
.fpatan = 5,
.fprem = 2,
.fprem1 = 6,
.fptan = 70,
.frndint = 0,
.frstor = 0,
.fsave = 0,
.fscale = 2,
.fsetpm = 0,
.fsin_cos = 2,
.fsincos = 2,
.fsqrt = 70,
.fst = 0,
.fst_32 = 0,
.fst_64 = 0,
.fst_80 = 0,
.fstcw_sw = 0,
.fstenv = 0,
.ftst = 1,
.fucom = 1,
.fwait = 0,
.fxam = 0,
.fxch = 0,
.fxtract = 4,
.fyl2x = 13,
.fyl2xp1 = 13,
};

92
src/cpu/x87_timings.h
View File

@@ -1,51 +1,51 @@
typedef struct
{
int f2xm1;
int fabs;
int fadd, fadd_32, fadd_64;
int fbld;
int fbstp;
int fchs;
int fclex;
int fcom, fcom_32, fcom_64;
int fcos;
int fincdecstp;
int fdisi_eni;
int fdiv, fdiv_32, fdiv_64;
int ffree;
int fadd_i16, fadd_i32;
int fcom_i16, fcom_i32;
int fdiv_i16, fdiv_i32;
int fild_16, fild_32, fild_64;
int fmul_i16, fmul_i32;
int finit;
int fist_16, fist_32, fist_64;
int fld, fld_32, fld_64, fld_80;
int fld_z1, fld_const;
int fldcw;
int fldenv;
int fmul, fmul_32, fmul_64;
int fnop;
int fpatan;
int fprem, fprem1;
int fptan;
int frndint;
int frstor;
int fsave;
int fscale;
int fsetpm;
int fsin_cos, fsincos;
int fsqrt;
int fst, fst_32, fst_64, fst_80;
int fstcw_sw;
int fstenv;
int ftst;
int fucom;
int fwait;
int fxam;
int fxch;
int fxtract;
int fyl2x, fyl2xp1;
int f2xm1;
int fabs;
int fadd, fadd_32, fadd_64;
int fbld;
int fbstp;
int fchs;
int fclex;
int fcom, fcom_32, fcom_64;
int fcos;
int fincdecstp;
int fdisi_eni;
int fdiv, fdiv_32, fdiv_64;
int ffree;
int fadd_i16, fadd_i32;
int fcom_i16, fcom_i32;
int fdiv_i16, fdiv_i32;
int fild_16, fild_32, fild_64;
int fmul_i16, fmul_i32;
int finit;
int fist_16, fist_32, fist_64;
int fld, fld_32, fld_64, fld_80;
int fld_z1, fld_const;
int fldcw;
int fldenv;
int fmul, fmul_32, fmul_64;
int fnop;
int fpatan;
int fprem, fprem1;
int fptan;
int frndint;
int frstor;
int fsave;
int fscale;
int fsetpm;
int fsin_cos, fsincos;
int fsqrt;
int fst, fst_32, fst_64, fst_80;
int fstcw_sw;
int fstenv;
int ftst;
int fucom;
int fwait;
int fxam;
int fxch;
int fxtract;
int fyl2x, fyl2xp1;
} x87_timings_t;
extern const x87_timings_t x87_timings_8087;