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

Rewrote the recompiler interrupt checking in assembly (and removed it for the new dynamic compiler because the requires uops are not present), brings performance back up, and also did a number of CPU-related clean-ups (mostly removal of dead variables and associated code).

Browse Source
This commit is contained in:
OBattler
2020-07-13 19:46:19 +02:00
parent 0cd0d83cee
commit a862bda04c
32 changed files with 196 additions and 291 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/codegen/codegen.h
View File

@@ -308,13 +308,6 @@ void codegen_check_flush(page_t *page, uint64_t mask, uint32_t phys_addr);
extern int cpu_block_end;
extern uint32_t codegen_endpc;
extern int cpu_recomp_blocks, cpu_recomp_full_ins, cpu_new_blocks;
extern int cpu_recomp_blocks_latched, cpu_recomp_ins_latched, cpu_recomp_full_ins_latched, cpu_new_blocks_latched;
extern int cpu_recomp_flushes, cpu_recomp_flushes_latched;
extern int cpu_recomp_evicted, cpu_recomp_evicted_latched;
extern int cpu_recomp_reuse, cpu_recomp_reuse_latched;
extern int cpu_recomp_removed, cpu_recomp_removed_latched;
extern int codegen_block_cycles;
extern void (*codegen_timing_start)();

3
src/codegen/codegen_accumulate.h
View File

@@ -1,7 +1,6 @@
enum
{
ACCREG_ins = 0,
ACCREG_cycles = 1,
ACCREG_cycles = 0,
ACCREG_COUNT
};

52
src/codegen/codegen_accumulate_x86-64.c
View File

@@ -13,7 +13,6 @@ static struct
uintptr_t dest_reg;
} acc_regs[] =
{
[ACCREG_ins] = {0, (uintptr_t) &(ins)},
[ACCREG_cycles] = {0, (uintptr_t) &(cycles)},
};
@@ -22,37 +21,42 @@ void codegen_accumulate(int acc_reg, int delta)
acc_regs[acc_reg].count += delta;
if ((acc_reg == ACCREG_cycles) && (delta != 0)) {
addbyte(0x81); /*ADD $acc_regs[c].count,acc_regs[c].dest*/
addbyte(0x04);
addbyte(0x25);
addlong((uint32_t) (uintptr_t) &(acycs));
addlong(-delta);
if (delta == -1) {
/* -delta = 1 */
addbyte(0xff); /*inc dword ptr[&acycs]*/
addbyte(0x04);
addbyte(0x25);
addlong((uint32_t) (uintptr_t) &(acycs));
} else if (delta == 1) {
/* -delta = -1 */
addbyte(0xff); /*dec dword ptr[&acycs]*/
addbyte(0x0c);
addbyte(0x25);
addlong((uint32_t) (uintptr_t) &(acycs));
} else {
addbyte(0x81); /*ADD $acc_regs[c].count,acc_regs[c].dest*/
addbyte(0x04);
addbyte(0x25);
addlong((uint32_t) (uintptr_t) &(acycs));
addlong(-delta);
}
}
}
void codegen_accumulate_flush(void)
{
int c;
for (c = 0; c < ACCREG_COUNT; c++)
{
if (acc_regs[c].count)
{
addbyte(0x81); /*ADD $acc_regs[c].count,acc_regs[c].dest*/
addbyte(0x04);
addbyte(0x25);
addlong((uint32_t) acc_regs[c].dest_reg);
addlong(acc_regs[c].count);
}
if (acc_regs[0].count) {
addbyte(0x81); /*ADD $acc_regs[0].count,acc_regs[0].dest*/
addbyte(0x04);
addbyte(0x25);
addlong((uint32_t) acc_regs[0].dest_reg);
addlong(acc_regs[0].count);
}
acc_regs[c].count = 0;
}
acc_regs[0].count = 0;
}
void codegen_accumulate_reset()
{
int c;
for (c = 0; c < ACCREG_COUNT; c++)
acc_regs[c].count = 0;
acc_regs[0].count = 0;
}

46
src/codegen/codegen_accumulate_x86.c
View File

@@ -13,7 +13,6 @@ static struct
uintptr_t dest_reg;
} acc_regs[] =
{
[ACCREG_ins] = {0, (uintptr_t) &(ins)},
[ACCREG_cycles] = {0, (uintptr_t) &(cycles)}
};
@@ -22,35 +21,38 @@ void codegen_accumulate(int acc_reg, int delta)
acc_regs[acc_reg].count += delta;
if ((acc_reg == ACCREG_cycles) && (delta != 0)) {
addbyte(0x81); /*ADD $acc_regs[c].count,acc_regs[c].dest*/
addbyte(0x05);
addlong((uint32_t) (uintptr_t) &(acycs));
addlong((uintptr_t) -delta);
if (delta == -1) {
/* -delta = 1 */
addbyte(0xff); /*inc dword ptr[&acycs]*/
addbyte(0x05);
addlong((uint32_t) (uintptr_t) &(acycs));
} else if (delta == 1) {
/* -delta = -1 */
addbyte(0xff); /*dec dword ptr[&acycs]*/
addbyte(0x0d);
addlong((uint32_t) (uintptr_t) &(acycs));
} else {
addbyte(0x81); /*ADD $acc_regs[c].count,acc_regs[c].dest*/
addbyte(0x05);
addlong((uint32_t) (uintptr_t) &(acycs));
addlong((uintptr_t) -delta);
}
}
}
void codegen_accumulate_flush(void)
{
int c;
for (c = 0; c < ACCREG_COUNT; c++)
{
if (acc_regs[c].count)
{
addbyte(0x81); /*ADD $acc_regs[c].count,acc_regs[c].dest*/
addbyte(0x05);
addlong((uint32_t) acc_regs[c].dest_reg);
addlong(acc_regs[c].count);
}
if (acc_regs[0].count) {
addbyte(0x81); /*ADD $acc_regs[0].count,acc_regs[0].dest*/
addbyte(0x05);
addlong((uint32_t) acc_regs[0].dest_reg);
addlong(acc_regs[0].count);
}
acc_regs[c].count = 0;
}
acc_regs[0].count = 0;
}
void codegen_accumulate_reset()
{
int c;
for (c = 0; c < ACCREG_COUNT; c++)
acc_regs[c].count = 0;
acc_regs[0].count = 0;
}

1
src/codegen/codegen_ops_misc.h
View File

@@ -19,6 +19,7 @@ static uint32_t ropCLI(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32
if (!IOPLp && (cr4 & (CR4_VME | CR4_PVI)))
return 0;
CLEAR_BITS((uintptr_t)&cpu_state.flags, I_FLAG);
CLEAR_BITS((uintptr_t)&pic_pending, 0xffffffff);
return op_pc;
}
static uint32_t ropSTI(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)

29
src/codegen/codegen_x86-64.c
View File

@@ -49,11 +49,6 @@ int block_current = 0;
static int block_num;
int block_pos;
int cpu_recomp_flushes, cpu_recomp_flushes_latched;
int cpu_recomp_evicted, cpu_recomp_evicted_latched;
int cpu_recomp_reuse, cpu_recomp_reuse_latched;
int cpu_recomp_removed, cpu_recomp_removed_latched;
uint32_t codegen_endpc;
int codegen_block_cycles;
@@ -224,7 +219,6 @@ void codegen_check_flush(page_t *page, uint64_t mask, uint32_t phys_addr)
if (mask & block->page_mask)
{
delete_block(block);
cpu_recomp_evicted++;
}
if (block == block->next)
fatal("Broken 1\n");
@@ -238,7 +232,6 @@ void codegen_check_flush(page_t *page, uint64_t mask, uint32_t phys_addr)
if (mask & block->page_mask2)
{
delete_block(block);
cpu_recomp_evicted++;
}
if (block == block->next_2)
fatal("Broken 2\n");
@@ -260,7 +253,6 @@ void codegen_block_init(uint32_t phys_addr)
if (block->valid != 0)
{
delete_block(block);
cpu_recomp_reuse++;
}
block_num = HASH(phys_addr);
codeblock_hash[block_num] = &codeblock[block_current];
@@ -393,7 +385,6 @@ void codegen_block_remove()
codeblock_t *block = &codeblock[block_current];
delete_block(block);
cpu_recomp_removed++;
recomp_page = -1;
}
@@ -1049,7 +1040,6 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
generate_call:
codegen_timing_opcode(opcode, fetchdat, op_32, op_pc);
codegen_accumulate(ACCREG_ins, 1);
codegen_accumulate(ACCREG_cycles, -codegen_block_cycles);
codegen_block_cycles = 0;
@@ -1096,10 +1086,11 @@ generate_call:
codegen_endpc = (cs + cpu_state.pc) + 8;
/* Check for interrupts. */
call(block, (uintptr_t)int_check);
addbyte(0x85); /*OR %eax, %eax*/
addbyte(0xc0);
addbyte(0xf6); /* test byte ptr[&pic_pending],1 */
addbyte(0x04);
addbyte(0x25);
addlong((uint32_t) (uintptr_t) &pic_pending);
addbyte(0x01);
addbyte(0x0F); addbyte(0x85); /*JNZ 0*/
addlong((uint32_t)(uintptr_t)&block->data[BLOCK_EXIT_OFFSET] - (uint32_t)(uintptr_t)(&block->data[block_pos + 4]));
@@ -1175,13 +1166,11 @@ generate_call:
block->ins++;
addbyte(0x85); /*OR %eax, %eax*/
addbyte(0xc0);
addbyte(0x0F); addbyte(0x85); /*JNZ 0*/
addlong((uint32_t)(uintptr_t)&block->data[BLOCK_EXIT_OFFSET] - (uint32_t)(uintptr_t)(&block->data[block_pos + 4]));
/* Check for interrupts. */
call(block, (uintptr_t)int_check);
addbyte(0x0a); /* or al,byte ptr[&pic_pending] */
addbyte(0x04);
addbyte(0x25);
addlong((uint32_t) (uintptr_t) &pic_pending);
addbyte(0x85); /*OR %eax, %eax*/
addbyte(0xc0);

32
src/codegen/codegen_x86.c
View File

@@ -50,6 +50,9 @@
#include "x86_flags.h"
#include "x86_ops.h"
#include "x87.h"
/*ex*/
#include <86box/nmi.h>
#include <86box/pic.h>
#include "386_common.h"
@@ -88,11 +91,6 @@ int block_current = 0;
static int block_num;
int block_pos;
int cpu_recomp_flushes, cpu_recomp_flushes_latched;
int cpu_recomp_evicted, cpu_recomp_evicted_latched;
int cpu_recomp_reuse, cpu_recomp_reuse_latched;
int cpu_recomp_removed, cpu_recomp_removed_latched;
uint32_t codegen_endpc;
int codegen_block_cycles;
@@ -1367,7 +1365,6 @@ void codegen_check_flush(page_t *page, uint64_t mask, uint32_t phys_addr)
if (mask & block->page_mask)
{
delete_block(block);
cpu_recomp_evicted++;
}
if (block == block->next)
fatal("Broken 1\n");
@@ -1381,7 +1378,6 @@ void codegen_check_flush(page_t *page, uint64_t mask, uint32_t phys_addr)
if (mask & block->page_mask2)
{
delete_block(block);
cpu_recomp_evicted++;
}
if (block == block->next_2)
fatal("Broken 2\n");
@@ -1403,7 +1399,6 @@ void codegen_block_init(uint32_t phys_addr)
if (block->valid != 0)
{
delete_block(block);
cpu_recomp_reuse++;
}
block_num = HASH(phys_addr);
codeblock_hash[block_num] = &codeblock[block_current];
@@ -1512,7 +1507,6 @@ void codegen_block_remove()
codeblock_t *block = &codeblock[block_current];
delete_block(block);
cpu_recomp_removed++;
recomp_page = -1;
}
@@ -2016,7 +2010,6 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
generate_call:
codegen_timing_opcode(opcode, fetchdat, op_32, op_pc);
codegen_accumulate(ACCREG_ins, 1);
codegen_accumulate(ACCREG_cycles, -codegen_block_cycles);
codegen_block_cycles = 0;
@@ -2063,11 +2056,10 @@ generate_call:
codegen_endpc = (cs + cpu_state.pc) + 8;
/* Check for interrupts. */
addbyte(0xE8); /*CALL*/
addlong(((uint8_t *)int_check - (uint8_t *)(&block->data[block_pos + 4])));
addbyte(0x09); /*OR %eax, %eax*/
addbyte(0xc0);
addbyte(0xf6); /* test byte ptr[&pic_pending],1 */
addbyte(0x05);
addlong((uint32_t) (uintptr_t) &pic_pending);
addbyte(0x01);
addbyte(0x0F); addbyte(0x85); /*JNZ 0*/
addlong((uint32_t)&block->data[BLOCK_EXIT_OFFSET] - (uint32_t)(&block->data[block_pos + 4]));
@@ -2154,14 +2146,10 @@ generate_call:
block->ins++;
addbyte(0x09); /*OR %eax, %eax*/
addbyte(0xc0);
addbyte(0x0F); addbyte(0x85); /*JNZ 0*/
addlong((uint32_t)&block->data[BLOCK_EXIT_OFFSET] - (uint32_t)(&block->data[block_pos + 4]));
/* Check for interrupts. */
addbyte(0xE8); /*CALL*/
addlong(((uint8_t *)int_check - (uint8_t *)(&block->data[block_pos + 4])));
addbyte(0x0a); /* or al,byte ptr[&pic_pending] */
addbyte(0x05);
addlong((uint32_t) (uintptr_t) &pic_pending);
addbyte(0x09); /*OR %eax, %eax*/
addbyte(0xc0);

14
src/codegen/x86seg.c
View File

@@ -918,11 +918,9 @@ void loadcscall(uint16_t seg)
int type;
uint16_t tempw;
int csout = output;
if (msw&1 && !(cpu_state.eflags&VM_FLAG))
{
if (csout) x86seg_log("Protected mode CS load! %04X\n",seg);
x86seg_log("Protected mode CS load! %04X\n", seg);
if (!(seg&~3))
{
x86gpf("loadcscall(): Protected mode selector is zero",0);
@@ -956,7 +954,7 @@ void loadcscall(uint16_t seg)
newpc=segdat[0];
if (type&0x800) newpc|=segdat[3]<<16;
if (csout) x86seg_log("Code seg call - %04X - %04X %04X %04X\n",seg,segdat[0],segdat[1],segdat[2]);
x86seg_log("Code seg call - %04X - %04X %04X %04X\n",seg,segdat[0],segdat[1],segdat[2]);
if (segdat[2]&0x1000)
{
if (!(segdat[2]&0x400)) /*Not conforming*/
@@ -1001,14 +999,16 @@ void loadcscall(uint16_t seg)
CS=seg;
do_seg_load(&cpu_state.seg_cs, segdat);
if (CPL==3 && oldcpl!=3) flushmmucache_cr3();
if (csout) x86seg_log("Complete\n");
#ifdef ENABLE_X86SEG_LOG
x86seg_log("Complete\n");
#endif
cycles -= timing_call_pm;
}
else
{
type=segdat[2]&0xF00;
if (csout) x86seg_log("Type %03X\n",type);
x86seg_log("Type %03X\n",type);
switch (type)
{
case 0x400: /*Call gate*/