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

Reorganized some CPU structures a bit and fixed inappropriate behavior of some mem.c functions on 64-bit binaries (and made mem.c aware of ARM64 as well), fixes both 64-bit recompilers, closes #1215.

Browse Source
This commit is contained in:
OBattler
2021-08-27 13:31:47 +02:00
parent eb827a6629
commit ddc7c5b78a
20 changed files with 181 additions and 176 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/codegen_new/codegen_allocator.c
View File

@@ -1,6 +1,7 @@
#if defined(__linux__) || defined(__APPLE__)
#include <sys/mman.h>
#include <unistd.h>
#include <stdlib.h>
#endif
#if defined WIN32 || defined _WIN32 || defined _WIN32
#include <windows.h>
@@ -35,6 +36,10 @@ void codegen_allocator_init()
#if defined WIN32 || defined _WIN32 || defined _WIN32
mem_block_alloc = VirtualAlloc(NULL, MEM_BLOCK_NR * MEM_BLOCK_SIZE, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
/* TODO: check deployment target: older Intel-based versions of macOS don't play
nice with MAP_JIT. */
#elif defined(__APPLE__) && defined(MAP_JIT)
mem_block_alloc = mmap(0, MEM_BLOCK_NR * MEM_BLOCK_SIZE, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_ANON|MAP_PRIVATE|MAP_JIT, 0, 0);
#else
mem_block_alloc = mmap(0, MEM_BLOCK_NR * MEM_BLOCK_SIZE, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_ANON|MAP_PRIVATE, 0, 0);
#endif
@@ -115,11 +120,7 @@ void codegen_allocator_clean_blocks(struct mem_block_t *block)
#if defined __ARM_EABI__ || defined _ARM_ || defined __aarch64__ || defined _M_ARM || defined _M_ARM64
while (1)
{
#ifndef _MSC_VER
__clear_cache(&mem_block_alloc[block->offset], &mem_block_alloc[block->offset + MEM_BLOCK_SIZE]);
#else
FlushInstructionCache(GetCurrentProcess(), &mem_block_alloc[block->offset], MEM_BLOCK_SIZE);
#endif
if (block->next)
block = &mem_blocks[block->next - 1];
else

4
src/codegen_new/codegen_backend_arm.c
View File

@@ -327,13 +327,9 @@ printf("block_pos=%i\n", block_pos);
block_write_data = NULL;
//fatal("block_pos=%i\n", block_pos);
#if !defined _MSC_VER || defined __clang__
asm("vmrs %0, fpscr\n"
: "=r" (cpu_state.old_fp_control)
);
#else
cpu_state.old_fp_control = _controlfp();
#endif
if ((cpu_state.old_fp_control >> 22) & 3)
fatal("VFP not in nearest rounding mode\n");
}

4
src/codegen_new/codegen_backend_arm64.c
View File

@@ -332,13 +332,9 @@ void codegen_backend_init()
codegen_allocator_clean_blocks(block->head_mem_block);
#if !defined _MSC_VER || defined __clang__
asm("mrs %0, fpcr\n"
: "=r" (cpu_state.old_fp_control)
);
#else
cpu_state.old_fp_control = _controlfp();
#endif
}
void codegen_set_rounding_mode(int mode)

9
src/codegen_new/codegen_backend_x86-64.c
View File

@@ -22,8 +22,6 @@
#include <windows.h>
#endif
#include <xmmintrin.h>
void *codegen_mem_load_byte;
void *codegen_mem_load_word;
void *codegen_mem_load_long;
@@ -326,7 +324,6 @@ void codegen_backend_init()
host_x86_XOR32_REG_REG(block, REG_EDI, REG_EDI);
host_x86_XOR32_REG_REG(block, REG_ESI, REG_ESI);
#endif
/* host_x86_CALL(block, (uintptr_t)x86gpf); */
host_x86_CALL(block, (void *)x86gpf);
codegen_exit_rout = &codeblock[block_current].data[block_pos];
host_x86_ADD64_REG_IMM(block, REG_RSP, 0x38);
@@ -342,7 +339,11 @@ void codegen_backend_init()
block_write_data = NULL;
cpu_state.trunc_fp_control = _mm_getcsr() | 0x6000;
asm(
"stmxcsr %0\n"
: "=m" (cpu_state.old_fp_control)
);
cpu_state.trunc_fp_control = cpu_state.old_fp_control | 0x6000;
}
void codegen_set_rounding_mode(int mode)

2
src/codegen_new/codegen_backend_x86-64_defs.h
View File

@@ -44,7 +44,7 @@
#define REG_XMM6 6
#define REG_XMM7 7
#define REG_XMM_TEMP REG_XMM7
#define REG_XMM_TEMP REG_XMM0
#define CODEGEN_HOST_REGS 3
#define CODEGEN_HOST_FP_REGS 7

4
src/codegen_new/codegen_backend_x86-64_ops.c
View File

@@ -32,7 +32,7 @@ static inline void call(codeblock_t *block, uintptr_t func)
codegen_alloc_bytes(block, 5);
diff = func - (uintptr_t)&block_write_data[block_pos + 5];
if (diff >= -0x80000000ULL && diff < 0x7fffffffULL)
if (diff >= -0x80000000 && diff < 0x7fffffff)
{
codegen_addbyte(block, 0xE8); /*CALL*/
codegen_addlong(block, (uint32_t)diff);
@@ -53,7 +53,7 @@ static inline void jmp(codeblock_t *block, uintptr_t func)
codegen_alloc_bytes(block, 5);
diff = func - (uintptr_t)&block_write_data[block_pos + 5];
if (diff >= -0x80000000ULL && diff < 0x7fffffffULL)
if (diff >= -0x80000000 && diff < 0x7fffffff)
{
codegen_addbyte(block, 0xe9); /*JMP*/
codegen_addlong(block, (uint32_t)diff);

18
src/codegen_new/codegen_backend_x86-64_uops.c
View File

@@ -199,9 +199,9 @@ static int codegen_CALL_FUNC(codeblock_t *block, uop_t *uop)
static int codegen_CALL_FUNC_RESULT(codeblock_t *block, uop_t *uop)
{
int dest_reg = HOST_REG_GET(uop->dest_reg_a_real);
/* int dest_size = IREG_GET_SIZE(uop->dest_reg_a_real); */
#ifdef RECOMPILER_DEBUG
int dest_size = IREG_GET_SIZE(uop->dest_reg_a_real);
if (!REG_IS_L(dest_size))
fatal("CALL_FUNC_RESULT %02x\n", uop->dest_reg_a_real);
#endif
@@ -922,9 +922,9 @@ static int codegen_LOAD_FUNC_ARG3_IMM(codeblock_t *block, uop_t *uop)
static int codegen_LOAD_SEG(codeblock_t *block, uop_t *uop)
{
int src_reg = HOST_REG_GET(uop->src_reg_a_real);
/* int src_size = IREG_GET_SIZE(uop->src_reg_a_real); */
#ifdef RECOMPILER_DEBUG
int src_size = IREG_GET_SIZE(uop->src_reg_a_real);
if (!REG_IS_W(src_size))
fatal("LOAD_SEG %02x %p\n", uop->src_reg_a_real, uop->p);
#endif
@@ -1033,9 +1033,9 @@ static int codegen_MEM_LOAD_REG(codeblock_t *block, uop_t *uop)
static int codegen_MEM_LOAD_SINGLE(codeblock_t *block, uop_t *uop)
{
int dest_reg = HOST_REG_GET(uop->dest_reg_a_real), seg_reg = HOST_REG_GET(uop->src_reg_a_real), addr_reg = HOST_REG_GET(uop->src_reg_b_real);
/* int dest_size = IREG_GET_SIZE(uop->dest_reg_a_real); */
#ifdef RECOMPILER_DEBUG
int dest_size = IREG_GET_SIZE(uop->dest_reg_a_real);
if (!REG_IS_D(dest_size))
fatal("MEM_LOAD_SINGLE - %02x\n", uop->dest_reg_a_real);
#endif
@@ -1052,9 +1052,9 @@ static int codegen_MEM_LOAD_SINGLE(codeblock_t *block, uop_t *uop)
static int codegen_MEM_LOAD_DOUBLE(codeblock_t *block, uop_t *uop)
{
int dest_reg = HOST_REG_GET(uop->dest_reg_a_real), seg_reg = HOST_REG_GET(uop->src_reg_a_real), addr_reg = HOST_REG_GET(uop->src_reg_b_real);
/* int dest_size = IREG_GET_SIZE(uop->dest_reg_a_real); */
#ifdef RECOMPILER_DEBUG
int dest_size = IREG_GET_SIZE(uop->dest_reg_a_real);
if (!REG_IS_D(dest_size))
fatal("MEM_LOAD_DOUBLE - %02x\n", uop->dest_reg_a_real);
#endif
@@ -1178,9 +1178,9 @@ static int codegen_MEM_STORE_REG(codeblock_t *block, uop_t *uop)
static int codegen_MEM_STORE_SINGLE(codeblock_t *block, uop_t *uop)
{
int seg_reg = HOST_REG_GET(uop->src_reg_a_real), addr_reg = HOST_REG_GET(uop->src_reg_b_real), src_reg = HOST_REG_GET(uop->src_reg_c_real);
/* int src_size = IREG_GET_SIZE(uop->src_reg_c_real); */
#ifdef RECOMPILER_DEBUG
int src_size = IREG_GET_SIZE(uop->src_reg_c_real);
if (!REG_IS_D(src_size))
fatal("MEM_STORE_SINGLE - %02x\n", uop->src_reg_b_real);
#endif
@@ -1197,9 +1197,9 @@ static int codegen_MEM_STORE_SINGLE(codeblock_t *block, uop_t *uop)
static int codegen_MEM_STORE_DOUBLE(codeblock_t *block, uop_t *uop)
{
int seg_reg = HOST_REG_GET(uop->src_reg_a_real), addr_reg = HOST_REG_GET(uop->src_reg_b_real), src_reg = HOST_REG_GET(uop->src_reg_c_real);
/* int src_size = IREG_GET_SIZE(uop->src_reg_c_real); */
#ifdef RECOMPILER_DEBUG
int src_size = IREG_GET_SIZE(uop->src_reg_c_real);
if (!REG_IS_D(src_size))
fatal("MEM_STORE_DOUBLE - %02x\n", uop->src_reg_b_real);
#endif
@@ -1499,7 +1499,7 @@ static int codegen_OR(codeblock_t *block, uop_t *uop)
}
static int codegen_OR_IMM(codeblock_t *block, uop_t *uop)
{
int dest_reg = HOST_REG_GET(uop->dest_reg_a_real)/*, src_reg = HOST_REG_GET(uop->src_reg_a_real)*/;
int dest_reg = HOST_REG_GET(uop->dest_reg_a_real);
int dest_size = IREG_GET_SIZE(uop->dest_reg_a_real), src_size = IREG_GET_SIZE(uop->src_reg_a_real);
if (REG_IS_L(dest_size) && REG_IS_L(src_size))
@@ -2770,7 +2770,7 @@ static int codegen_TEST_JS_DEST(codeblock_t *block, uop_t *uop)
static int codegen_XOR(codeblock_t *block, uop_t *uop)
{
int dest_reg = HOST_REG_GET(uop->dest_reg_a_real)/*, src_reg_a = HOST_REG_GET(uop->src_reg_a_real)*/, src_reg_b = HOST_REG_GET(uop->src_reg_b_real);
int dest_reg = HOST_REG_GET(uop->dest_reg_a_real), src_reg_b = HOST_REG_GET(uop->src_reg_b_real);
int dest_size = IREG_GET_SIZE(uop->dest_reg_a_real), src_size_a = IREG_GET_SIZE(uop->src_reg_a_real), src_size_b = IREG_GET_SIZE(uop->src_reg_b_real);
if (REG_IS_Q(dest_size) && REG_IS_Q(src_size_a) && REG_IS_Q(src_size_b) && uop->dest_reg_a_real == uop->src_reg_a_real)
@@ -2797,7 +2797,7 @@ static int codegen_XOR(codeblock_t *block, uop_t *uop)
}
static int codegen_XOR_IMM(codeblock_t *block, uop_t *uop)
{
int dest_reg = HOST_REG_GET(uop->dest_reg_a_real)/*, src_reg = HOST_REG_GET(uop->src_reg_a_real)*/;
int dest_reg = HOST_REG_GET(uop->dest_reg_a_real);
int dest_size = IREG_GET_SIZE(uop->dest_reg_a_real), src_size = IREG_GET_SIZE(uop->src_reg_a_real);
if (REG_IS_L(dest_size) && REG_IS_L(src_size))

8
src/codegen_new/codegen_backend_x86.c
View File

@@ -304,20 +304,12 @@ void codegen_backend_init()
block_write_data = NULL;
cpu_state.old_fp_control = 0;
#ifndef _MSC_VER
asm(
"fstcw %0\n"
"stmxcsr %1\n"
: "=m" (cpu_state.old_fp_control2),
"=m" (cpu_state.old_fp_control)
);
#else
__asm
{
fstcw cpu_state.old_fp_control2
stmxcsr cpu_state.old_fp_control
}
#endif
cpu_state.trunc_fp_control = cpu_state.old_fp_control | 0x6000;
}

2
src/codegen_new/codegen_ops_helpers.c
View File

@@ -62,7 +62,7 @@ int codegen_can_unroll_full(codeblock_t *block, ir_data_t *ir, uint32_t next_pc,
return 0;
max_unroll = UNROLL_MAX_UOPS / ((ir->wr_pos-start)+6);
if ((max_version_refcount != 0) && (max_unroll > (UNROLL_MAX_REG_REFERENCES / max_version_refcount)))
if (max_unroll > (UNROLL_MAX_REG_REFERENCES / max_version_refcount))
max_unroll = (UNROLL_MAX_REG_REFERENCES / max_version_refcount);
if (max_unroll > UNROLL_MAX_COUNT)
max_unroll = UNROLL_MAX_COUNT;

6
src/codegen_new/codegen_ops_jump.c
View File

@@ -13,7 +13,7 @@
uint32_t ropJMP_r8(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
{
int32_t offset = (int32_t)(int8_t)fastreadb(cs + op_pc);
uint32_t offset = (int32_t)(int8_t)fastreadb(cs + op_pc);
uint32_t dest_addr = op_pc+1+offset;
if (!(op_32 & 0x100))
@@ -26,7 +26,7 @@ uint32_t ropJMP_r8(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t f
}
uint32_t ropJMP_r16(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
{
int32_t offset = (int32_t)(int16_t)fastreadw(cs + op_pc);
uint32_t offset = (int32_t)(int16_t)fastreadw(cs + op_pc);
uint32_t dest_addr = op_pc+2+offset;
dest_addr &= 0xffff;
@@ -38,7 +38,7 @@ uint32_t ropJMP_r16(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t
}
uint32_t ropJMP_r32(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
{
int32_t offset = fastreadl(cs + op_pc);
uint32_t offset = fastreadl(cs + op_pc);
uint32_t dest_addr = op_pc+4+offset;
if (offset < 0)

24
src/codegen_new/codegen_ops_logic.c
View File

@@ -65,7 +65,7 @@ uint32_t ropAND_b_rm(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t
{
int src_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_AND(ir, IREG_8(dest_reg), IREG_8(dest_reg), IREG_8(src_reg));
uop_AND(ir, IREG_8(dest_reg), IREG_8(dest_reg), IREG_8(src_reg));
}
else
{
@@ -93,7 +93,7 @@ uint32_t ropAND_b_rmw(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_
{
int dest_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_AND(ir, IREG_8(dest_reg), IREG_8(dest_reg), IREG_8(src_reg));
uop_AND(ir, IREG_8(dest_reg), IREG_8(dest_reg), IREG_8(src_reg));
uop_MOVZX(ir, IREG_flags_res, IREG_8(dest_reg));
}
else
@@ -124,7 +124,7 @@ uint32_t ropAND_w_rm(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t
{
int src_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_AND(ir, IREG_16(dest_reg), IREG_16(dest_reg), IREG_16(src_reg));
uop_AND(ir, IREG_16(dest_reg), IREG_16(dest_reg), IREG_16(src_reg));
}
else
{
@@ -152,7 +152,7 @@ uint32_t ropAND_w_rmw(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_
{
int dest_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_AND(ir, IREG_16(dest_reg), IREG_16(dest_reg), IREG_16(src_reg));
uop_AND(ir, IREG_16(dest_reg), IREG_16(dest_reg), IREG_16(src_reg));
uop_MOVZX(ir, IREG_flags_res, IREG_16(dest_reg));
}
else
@@ -183,7 +183,7 @@ uint32_t ropAND_l_rm(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t
{
int src_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_AND(ir, IREG_32(dest_reg), IREG_32(dest_reg), IREG_32(src_reg));
uop_AND(ir, IREG_32(dest_reg), IREG_32(dest_reg), IREG_32(src_reg));
}
else
{
@@ -211,7 +211,7 @@ uint32_t ropAND_l_rmw(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_
{
int dest_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_AND(ir, IREG_32(dest_reg), IREG_32(dest_reg), IREG_32(src_reg));
uop_AND(ir, IREG_32(dest_reg), IREG_32(dest_reg), IREG_32(src_reg));
uop_MOV(ir, IREG_flags_res, IREG_32(dest_reg));
}
else
@@ -288,7 +288,7 @@ uint32_t ropOR_b_rm(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t
{
int src_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_OR(ir, IREG_8(dest_reg), IREG_8(dest_reg), IREG_8(src_reg));
uop_OR(ir, IREG_8(dest_reg), IREG_8(dest_reg), IREG_8(src_reg));
}
else
{
@@ -316,7 +316,7 @@ uint32_t ropOR_b_rmw(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t
{
int dest_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_OR(ir, IREG_8(dest_reg), IREG_8(dest_reg), IREG_8(src_reg));
uop_OR(ir, IREG_8(dest_reg), IREG_8(dest_reg), IREG_8(src_reg));
uop_MOVZX(ir, IREG_flags_res, IREG_8(dest_reg));
}
else
@@ -347,7 +347,7 @@ uint32_t ropOR_w_rm(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t
{
int src_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_OR(ir, IREG_16(dest_reg), IREG_16(dest_reg), IREG_16(src_reg));
uop_OR(ir, IREG_16(dest_reg), IREG_16(dest_reg), IREG_16(src_reg));
}
else
{
@@ -375,7 +375,7 @@ uint32_t ropOR_w_rmw(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t
{
int dest_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_OR(ir, IREG_16(dest_reg), IREG_16(dest_reg), IREG_16(src_reg));
uop_OR(ir, IREG_16(dest_reg), IREG_16(dest_reg), IREG_16(src_reg));
uop_MOVZX(ir, IREG_flags_res, IREG_16(dest_reg));
}
else
@@ -406,7 +406,7 @@ uint32_t ropOR_l_rm(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t
{
int src_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_OR(ir, IREG_32(dest_reg), IREG_32(dest_reg), IREG_32(src_reg));
uop_OR(ir, IREG_32(dest_reg), IREG_32(dest_reg), IREG_32(src_reg));
}
else
{
@@ -434,7 +434,7 @@ uint32_t ropOR_l_rmw(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t
{
int dest_reg = fetchdat & 7;
if(src_reg != dest_reg) uop_OR(ir, IREG_32(dest_reg), IREG_32(dest_reg), IREG_32(src_reg));
uop_OR(ir, IREG_32(dest_reg), IREG_32(dest_reg), IREG_32(src_reg));
uop_MOV(ir, IREG_flags_res, IREG_32(dest_reg));
}
else

21
src/codegen_new/codegen_reg.c
View File

@@ -282,7 +282,7 @@ static void codegen_reg_load(host_reg_set_t *reg_set, codeblock_t *block, int c,
fatal("codegen_reg_load - REG_WORD !REG_INTEGER\n");
#endif
if ((uintptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p < 256)
codegen_direct_read_16_stack(block, reg_set->reg_list[c].reg, (int)(uintptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p);
codegen_direct_read_16_stack(block, reg_set->reg_list[c].reg, (intptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p);
else
codegen_direct_read_16(block, reg_set->reg_list[c].reg, ireg_data[IREG_GET_REG(ir_reg.reg)].p);
break;
@@ -293,7 +293,7 @@ static void codegen_reg_load(host_reg_set_t *reg_set, codeblock_t *block, int c,
fatal("codegen_reg_load - REG_DWORD !REG_INTEGER\n");
#endif
if ((uintptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p < 256)
codegen_direct_read_32_stack(block, reg_set->reg_list[c].reg, (int)(uintptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p);
codegen_direct_read_32_stack(block, reg_set->reg_list[c].reg, (intptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p);
else
codegen_direct_read_32(block, reg_set->reg_list[c].reg, ireg_data[IREG_GET_REG(ir_reg.reg)].p);
break;
@@ -304,7 +304,7 @@ static void codegen_reg_load(host_reg_set_t *reg_set, codeblock_t *block, int c,
fatal("codegen_reg_load - REG_QWORD !REG_FP\n");
#endif
if ((uintptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p < 256)
codegen_direct_read_64_stack(block, reg_set->reg_list[c].reg, (int)(uintptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p);
codegen_direct_read_64_stack(block, reg_set->reg_list[c].reg, (intptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p);
else
codegen_direct_read_64(block, reg_set->reg_list[c].reg, ireg_data[IREG_GET_REG(ir_reg.reg)].p);
break;
@@ -315,7 +315,7 @@ static void codegen_reg_load(host_reg_set_t *reg_set, codeblock_t *block, int c,
fatal("codegen_reg_load - REG_POINTER !REG_INTEGER\n");
#endif
if ((uintptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p < 256)
codegen_direct_read_pointer_stack(block, reg_set->reg_list[c].reg, (int)(uintptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p);
codegen_direct_read_pointer_stack(block, reg_set->reg_list[c].reg, (intptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p);
else
codegen_direct_read_pointer(block, reg_set->reg_list[c].reg, ireg_data[IREG_GET_REG(ir_reg.reg)].p);
break;
@@ -326,7 +326,7 @@ static void codegen_reg_load(host_reg_set_t *reg_set, codeblock_t *block, int c,
fatal("codegen_reg_load - REG_DOUBLE !REG_FP\n");
#endif
if ((uintptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p < 256)
codegen_direct_read_double_stack(block, reg_set->reg_list[c].reg, (int)(uintptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p);
codegen_direct_read_double_stack(block, reg_set->reg_list[c].reg, (intptr_t)ireg_data[IREG_GET_REG(ir_reg.reg)].p);
else
codegen_direct_read_double(block, reg_set->reg_list[c].reg, ireg_data[IREG_GET_REG(ir_reg.reg)].p);
break;
@@ -364,7 +364,8 @@ static void codegen_reg_load(host_reg_set_t *reg_set, codeblock_t *block, int c,
codegen_direct_read_st_double(block, reg_set->reg_list[c].reg, &cpu_state.ST[0], ir_reg.reg & 7);
break;
default: fatal("codegen_reg_load - native_size=%i reg=%i\n", ireg_data[IREG_GET_REG(ir_reg.reg)].native_size, IREG_GET_REG(ir_reg.reg));
default:
fatal("codegen_reg_load - native_size=%i reg=%i\n", ireg_data[IREG_GET_REG(ir_reg.reg)].native_size, IREG_GET_REG(ir_reg.reg));
}
reg_set->regs[c] = ir_reg;
@@ -407,7 +408,7 @@ static void codegen_reg_writeback(host_reg_set_t *reg_set, codeblock_t *block, i
fatal("codegen_reg_writeback - REG_DWORD !REG_INTEGER\n");
#endif
if ((uintptr_t)p < 256)
codegen_direct_write_32_stack(block, (int)(uintptr_t)p, reg_set->reg_list[c].reg);
codegen_direct_write_32_stack(block, (intptr_t)p, reg_set->reg_list[c].reg);
else
codegen_direct_write_32(block, p, reg_set->reg_list[c].reg);
break;
@@ -418,7 +419,7 @@ static void codegen_reg_writeback(host_reg_set_t *reg_set, codeblock_t *block, i
fatal("codegen_reg_writeback - REG_QWORD !REG_FP\n");
#endif
if ((uintptr_t)p < 256)
codegen_direct_write_64_stack(block, (int)(uintptr_t)p, reg_set->reg_list[c].reg);
codegen_direct_write_64_stack(block, (intptr_t)p, reg_set->reg_list[c].reg);
else
codegen_direct_write_64(block, p, reg_set->reg_list[c].reg);
break;
@@ -439,7 +440,7 @@ static void codegen_reg_writeback(host_reg_set_t *reg_set, codeblock_t *block, i
fatal("codegen_reg_writeback - REG_DOUBLE !REG_FP\n");
#endif
if ((uintptr_t)p < 256)
codegen_direct_write_double_stack(block, (int)(uintptr_t)p, reg_set->reg_list[c].reg);
codegen_direct_write_double_stack(block, (intptr_t)p, reg_set->reg_list[c].reg);
else
codegen_direct_write_double(block, p, reg_set->reg_list[c].reg);
break;
@@ -512,7 +513,7 @@ void codegen_reg_write_imm(codeblock_t *block, ir_reg_t ir_reg, uint32_t imm_dat
case REG_DWORD:
if ((uintptr_t)p < 256)
codegen_direct_write_32_imm_stack(block, (int)(intptr_t)p, imm_data);
codegen_direct_write_32_imm_stack(block, (int)p, imm_data);
else
codegen_direct_write_32_imm(block, p, imm_data);
break;

2
src/codegen_new/codegen_reg.h
View File

@@ -343,6 +343,7 @@ static inline ir_reg_t codegen_reg_read(int reg)
CPU_BLOCK_END();
if (version->refcount > max_version_refcount)
max_version_refcount = version->refcount;
// pclog("codegen_reg_read: %i %i %i\n", reg & IREG_REG_MASK, ireg.version, reg_version_refcount[IREG_GET_REG(ireg.reg)][ireg.version]);
return ireg;
}
@@ -383,6 +384,7 @@ static inline ir_reg_t codegen_reg_write(int reg, int uop_nr)
version->refcount = 0;
version->flags = 0;
version->parent_uop = uop_nr;
// pclog("codegen_reg_write: %i\n", reg & IREG_REG_MASK);
return ireg;
}

8
src/cpu/386_common.c
View File

@@ -65,9 +65,8 @@ uint32_t old_rammask = 0xffffffff;
int soft_reset_mask = 0;
int in_smm = 0, smi_line = 0, smi_latched = 0, smm_in_hlt = 0;
int smi_block = 0;
uint32_t smbase = 0x30000;
int smi_latched = 0;
int smm_in_hlt = 0, smi_block = 0;
uint32_t addr64, addr64_2;
uint32_t addr64a[8], addr64a_2[8];
@@ -1845,7 +1844,6 @@ sysret(uint32_t fetchdat)
/* This is for compatibility with new x87 code. */
void codegen_set_rounding_mode(int mode)
{
/* cpu_state.new_npxc = (cpu_state.old_npxc & ~0xc00) | (cpu_state.npxc & 0xc00); */
cpu_state.new_npxc = (cpu_state.old_npxc & ~0xc00) | (mode << 10);
/* cpu_state.new_npxc = (cpu_state.old_npxc & ~0xc00) | (mode << 10); */
}
#endif

36
src/cpu/386_dynarec.c
View File

@@ -391,10 +391,10 @@ exec386_dynarec_dyn(void)
codeblock_t *block = codeblock_hash[hash];
#endif
int valid_block = 0;
#ifdef USE_NEW_DYNAREC
if (!cpu_state.abrt)
#else
if (block && !cpu_state.abrt)
#endif
{
@@ -535,6 +535,9 @@ exec386_dynarec_dyn(void)
cpu_block_end = 0;
x86_was_reset = 0;
#if defined(__APPLE__) && defined(__aarch64__)
pthread_jit_write_protect_np(0);
#endif
codegen_block_start_recompile(block);
codegen_in_recompile = 1;
@@ -585,21 +588,21 @@ exec386_dynarec_dyn(void)
#endif
CPU_BLOCK_END();
if (cpu_state.flags & T_FLAG)
CPU_BLOCK_END();
if (smi_line)
CPU_BLOCK_END();
if (nmi && nmi_enable && nmi_mask)
CPU_BLOCK_END();
if ((cpu_state.flags & I_FLAG) && pic.int_pending && !cpu_end_block_after_ins)
CPU_BLOCK_END();
if (cpu_end_block_after_ins) {
cpu_end_block_after_ins--;
if (!cpu_end_block_after_ins)
CPU_BLOCK_END();
}
if (smi_line)
CPU_BLOCK_END();
else if (cpu_state.flags & T_FLAG)
CPU_BLOCK_END();
else if (nmi && nmi_enable && nmi_mask)
CPU_BLOCK_END();
else if ((cpu_state.flags & I_FLAG) && pic.int_pending && !cpu_end_block_after_ins)
CPU_BLOCK_END();
if (cpu_state.abrt) {
if (!(cpu_state.abrt & ABRT_EXPECTED))
codegen_block_remove();
@@ -616,6 +619,9 @@ exec386_dynarec_dyn(void)
codegen_reset();
codegen_in_recompile = 0;
#if defined(__APPLE__) && defined(__aarch64__)
pthread_jit_write_protect_np(1);
#endif
} else if (!cpu_state.abrt) {
/* Mark block but do not recompile */
#ifdef USE_NEW_DYNAREC
@@ -642,8 +648,8 @@ exec386_dynarec_dyn(void)
cpu_state.ssegs = 0;
codegen_endpc = (cs + cpu_state.pc) + 8;
fetchdat = fastreadl(cs + cpu_state.pc);
#ifdef ENABLE_386_DYNAREC_LOG
if (in_smm)
x386_dynarec_log("[%04X:%08X] fetchdat = %08X\n", CS, cpu_state.pc, fetchdat);
@@ -677,13 +683,13 @@ exec386_dynarec_dyn(void)
#endif
CPU_BLOCK_END();
if (cpu_state.flags & T_FLAG)
CPU_BLOCK_END();
if (smi_line)
CPU_BLOCK_END();
else if (cpu_state.flags & T_FLAG)
if (nmi && nmi_enable && nmi_mask)
CPU_BLOCK_END();
else if (nmi && nmi_enable && nmi_mask)
CPU_BLOCK_END();
else if ((cpu_state.flags & I_FLAG) && pic.int_pending && !cpu_end_block_after_ins)
if ((cpu_state.flags & I_FLAG) && pic.int_pending && !cpu_end_block_after_ins)
CPU_BLOCK_END();
if (cpu_end_block_after_ins) {

27
src/cpu/cpu.h
View File

@@ -210,11 +210,12 @@ typedef union {
} x86reg;
typedef struct {
uint32_t base;
uint32_t limit;
uint8_t access, ar_high;
int8_t checked; /*Non-zero if selector is known to be valid*/
uint16_t seg;
uint32_t base, limit,
limit_low, limit_high;
uint32_t limit_low, limit_high;
int checked; /*Non-zero if selector is known to be valid*/
} x86seg;
typedef union {
@@ -352,9 +353,9 @@ typedef struct {
} rm_data;
uint8_t ssegs, ismmx,
abrt, pad;
abrt, _smi_line;
int _cycles;
int _cycles, _in_smm;
uint16_t npxs, npxc;
@@ -364,8 +365,6 @@ typedef struct {
MMX_REG MM[8];
uint16_t old_npxc, new_npxc;
#ifdef USE_NEW_DYNAREC
uint32_t old_fp_control, new_fp_control;
#if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined _M_IX86
@@ -374,6 +373,8 @@ typedef struct {
#if defined i386 || defined __i386 || defined __i386__ || defined _X86_ || defined _M_IX86 || defined __amd64__ || defined _M_X64
uint32_t trunc_fp_control;
#endif
#else
uint16_t old_npxc, new_npxc;
#endif
x86seg seg_cs, seg_ds, seg_es, seg_ss,
@@ -385,8 +386,17 @@ typedef struct {
} CR0;
uint16_t flags, eflags;
uint32_t _smbase;
} cpu_state_t;
#define in_smm cpu_state._in_smm
#define smi_line cpu_state._smi_line
#define smbase cpu_state._smbase
/*The cpu_state.flags below must match in both cpu_cur_status and block->status for a block
to be valid*/
#define CPU_STATUS_USE32 (1 << 0)
@@ -495,9 +505,8 @@ extern int hasfpu;
extern uint32_t cpu_features;
extern int in_smm, smi_line, smi_latched, smm_in_hlt;
extern int smi_latched, smm_in_hlt;
extern int smi_block;
extern uint32_t smbase;
#ifdef USE_NEW_DYNAREC
extern uint16_t cpu_cur_status;

30
src/mem/mem.c
View File

@@ -244,10 +244,15 @@ mem_flush_write_page(uint32_t addr, uint32_t virt)
{
page_t *page_target = &pages[addr >> 12];
int c;
#if (!(defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64))
uint32_t a;
#endif
for (c = 0; c < 256; c++) {
if (writelookup[c] != (int) 0xffffffff) {
#if (defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64)
uintptr_t target = (uintptr_t)&ram[(uintptr_t)(addr & ~0xfff) - (virt & ~0xfff)];
#else
a = (uintptr_t)(addr & ~0xfff) - (virt & ~0xfff);
uintptr_t target;
@@ -255,6 +260,7 @@ mem_flush_write_page(uint32_t addr, uint32_t virt)
target = (uintptr_t)&ram2[a - (1 << 30)];
else
target = (uintptr_t)&ram[a];
#endif
if (writelookup2[writelookup[c]] == target || page_lookup[writelookup[c]] == page_target) {
writelookup2[writelookup[c]] = LOOKUP_INV;
@@ -556,9 +562,7 @@ mem_addr_translate(uint32_t addr, uint32_t chunk_start, uint32_t len)
void
addreadlookup(uint32_t virt, uint32_t phys)
{
#if (defined __amd64__ || defined _M_X64)
uint64_t a;
#else
#if (!(defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64))
uint32_t a;
#endif
@@ -572,16 +576,16 @@ addreadlookup(uint32_t virt, uint32_t phys)
readlookup2[readlookup[readlnext]] = LOOKUP_INV;
}
#if (defined __amd64__ || defined _M_X64)
a = ((uint64_t)(phys & ~0xfff) - (uint64_t)(virt & ~0xfff));
#if (defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64)
readlookup2[virt>>12] = (uintptr_t)&ram[(uintptr_t)(phys & ~0xFFF) - (uintptr_t)(virt & ~0xfff)];
#else
a = ((uint32_t)(phys & ~0xfff) - (uint32_t)(virt & ~0xfff));
#endif
if ((phys & ~0xfff) >= (1 << 30))
readlookup2[virt>>12] = (uintptr_t)&ram2[a - (1 << 30)];
else
readlookup2[virt>>12] = (uintptr_t)&ram[a];
#endif
readlookupp[virt>>12] = mmu_perm;
readlookup[readlnext++] = virt >> 12;
@@ -594,9 +598,7 @@ addreadlookup(uint32_t virt, uint32_t phys)
void
addwritelookup(uint32_t virt, uint32_t phys)
{
#if (defined __amd64__ || defined _M_X64)
uint64_t a;
#else
#if (!(defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64))
uint32_t a;
#endif
@@ -625,16 +627,16 @@ addwritelookup(uint32_t virt, uint32_t phys)
page_lookup[virt >> 12] = &pages[phys >> 12];
page_lookupp[virt >> 12] = mmu_perm;
} else {
#if (defined __amd64__ || defined _M_X64)
a = ((uint64_t)(phys & ~0xfff) - (uint64_t)(virt & ~0xfff));
#if (defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64)
writelookup2[virt>>12] = (uintptr_t)&ram[(uintptr_t)(phys & ~0xFFF) - (uintptr_t)(virt & ~0xfff)];
#else
a = ((uint32_t)(phys & ~0xfff) - (uint32_t)(virt & ~0xfff));
#endif
if ((phys & ~0xfff) >= (1 << 30))
writelookup2[virt>>12] = (uintptr_t)&ram2[a - (1 << 30)];
else
writelookup2[virt>>12] = (uintptr_t)&ram[a];
#endif
}
writelookupp[virt>>12] = mmu_perm;
@@ -2607,7 +2609,7 @@ mem_reset(void)
free(ram);
ram = NULL;
}
#if (!(defined __amd64__ || defined _M_X64))
#if (!(defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64))
if (ram2 != NULL) {
free(ram2);
ram2 = NULL;
@@ -2619,7 +2621,7 @@ mem_reset(void)
m = 1024UL * mem_size;
#if (!(defined __amd64__ || defined _M_X64))
#if (!(defined __amd64__ || defined _M_X64 || defined __aarch64__ || defined _M_ARM64))
if (mem_size > 1048576) {
ram = (uint8_t *)malloc(1 << 30); /* allocate and clear the RAM block of the first 1 GB */
if (ram == NULL) {

9
src/win/Makefile.mingw
View File

@@ -605,10 +605,11 @@ MAINOBJ := 86box.o config.o random.o timer.o io.o acpi.o apm.o dma.o ddma.o \
MEMOBJ := catalyst_flash.o i2c_eeprom.o intel_flash.o mem.o rom.o smram.o spd.o sst_flash.o
CPUOBJ := cpu.o cpu_table.o fpu.o x86.o \
808x.o 386.o 386_common.o 386_dynarec.o 386_dynarec_ops.o $(CGTOBJ) \
x86seg.o x87.o x87_timings.o \
$(DYNARECOBJ)
CPUOBJ := $(DYNARECOBJ) \
$(CGTOBJ) \
cpu.o cpu_table.o fpu.o x86.o \
808x.o 386.o 386_common.o 386_dynarec.o 386_dynarec_ops.o \
x86seg.o x87.o x87_timings.o
CHIPSETOBJ := acc2168.o cs8230.o ali1217.o ali1429.o ali1489.o et6000.o headland.o intel_82335.o cs4031.o \
intel_420ex.o intel_4x0.o intel_sio.o intel_piix.o ioapic.o \