Retro/pcem
Watch
1
0
Fork
You've already forked pcem
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

x86-32 recompiler compiles FPU instructions using direct access to register stack when top-of-stack is static. Minor speedup.

Browse source
This commit is contained in:
SarahW 2016-09-11 15:38:05 +01:00
commit 75f9316c11
5 changed files with 1395 additions and 785 deletions
Download patch file Download diff file Expand all files Collapse all files

7
src/386_dynarec.c
View file

@ -1348,6 +1348,13 @@ void exec386_dynarec(int cycs)
valid_block = 0; valid_block = 0;
} }
} }
if (valid_block && block->was_recompiled && (block->flags & CODEBLOCK_STATIC_TOP) && block->TOP != cpu_state.TOP)
{
/*FPU top-of-stack does not match the value this block was compiled
with, re-compile using dynamic top-of-stack*/
block->flags &= ~CODEBLOCK_STATIC_TOP;
block->was_recompiled = 0;
}
} }
if (valid_block && block->was_recompiled) if (valid_block && block->was_recompiled)

7
src/codegen.h
View file

@ -56,10 +56,17 @@ typedef struct codeblock_t
uint64_t page_mask, page_mask2; uint64_t page_mask, page_mask2;
int was_recompiled; int was_recompiled;
uint32_t flags;
int TOP;
uint8_t data[2048]; uint8_t data[2048];
} codeblock_t; } codeblock_t;
/*Code block uses FPU*/
#define CODEBLOCK_HAS_FPU 1
/*Code block is always entered with the same FPU top-of-stack*/
#define CODEBLOCK_STATIC_TOP 2
static inline codeblock_t *codeblock_tree_find(uint32_t phys) static inline codeblock_t *codeblock_tree_find(uint32_t phys)
{ {
codeblock_t *block = pages[phys >> 12].head; codeblock_t *block = pages[phys >> 12].head;

2190
src/codegen_ops_x86.h
View file

File diff suppressed because it is too large Load diff

9
src/codegen_x86-64.c
View file

@ -281,6 +281,7 @@ void codegen_block_init(uint32_t phys_addr)
block->next = block->prev = NULL; block->next = block->prev = NULL;
block->next_2 = block->prev_2 = NULL; block->next_2 = block->prev_2 = NULL;
block->page_mask = 0; block->page_mask = 0;
block->flags = 0;
block->was_recompiled = 0; block->was_recompiled = 0;
@ -1005,6 +1006,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xd9: case 0xd9:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_d9_a32 : x86_dynarec_opcodes_d9_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_d9_a32 : x86_dynarec_opcodes_d9_a16;
@ -1013,6 +1015,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xda: case 0xda:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_da_a32 : x86_dynarec_opcodes_da_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_da_a32 : x86_dynarec_opcodes_da_a16;
@ -1021,6 +1024,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xdb: case 0xdb:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_db_a32 : x86_dynarec_opcodes_db_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_db_a32 : x86_dynarec_opcodes_db_a16;
@ -1029,6 +1033,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xdc: case 0xdc:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_dc_a32 : x86_dynarec_opcodes_dc_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_dc_a32 : x86_dynarec_opcodes_dc_a16;
@ -1038,6 +1043,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xdd: case 0xdd:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_dd_a32 : x86_dynarec_opcodes_dd_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_dd_a32 : x86_dynarec_opcodes_dd_a16;
@ -1046,6 +1052,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xde: case 0xde:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_de_a32 : x86_dynarec_opcodes_de_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_de_a32 : x86_dynarec_opcodes_de_a16;
@ -1054,6 +1061,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xdf: case 0xdf:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_df_a32 : x86_dynarec_opcodes_df_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_df_a32 : x86_dynarec_opcodes_df_a16;
@ -1062,6 +1070,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xf0: /*LOCK*/ case 0xf0: /*LOCK*/

13
src/codegen_x86.c
View file

@ -281,6 +281,7 @@ void codegen_block_init(uint32_t phys_addr)
block->next = block->prev = NULL; block->next = block->prev = NULL;
block->next_2 = block->prev_2 = NULL; block->next_2 = block->prev_2 = NULL;
block->page_mask = 0; block->page_mask = 0;
block->flags = CODEBLOCK_STATIC_TOP;
block->was_recompiled = 0; block->was_recompiled = 0;
@ -359,6 +360,7 @@ void codegen_block_start_recompile(codeblock_t *block)
_ds.checked = _es.checked = _fs.checked = _gs.checked = (cr0 & 1) ? 0 : 1; _ds.checked = _es.checked = _fs.checked = _gs.checked = (cr0 & 1) ? 0 : 1;
block->TOP = cpu_state.TOP;
block->was_recompiled = 1; block->was_recompiled = 1;
} }
@ -484,6 +486,9 @@ void codegen_block_end_recompile(codeblock_t *block)
codegen_block_generate_end_mask(); codegen_block_generate_end_mask();
add_to_block_list(block); add_to_block_list(block);
// pclog("End block %i\n", block_num); // pclog("End block %i\n", block_num);
if (!(block->flags & CODEBLOCK_HAS_FPU))
block->flags &= ~CODEBLOCK_STATIC_TOP;
} }
void codegen_flush() void codegen_flush()
@ -821,6 +826,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xd9: case 0xd9:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_d9_a32 : x86_dynarec_opcodes_d9_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_d9_a32 : x86_dynarec_opcodes_d9_a16;
@ -829,6 +835,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xda: case 0xda:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_da_a32 : x86_dynarec_opcodes_da_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_da_a32 : x86_dynarec_opcodes_da_a16;
@ -837,6 +844,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xdb: case 0xdb:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_db_a32 : x86_dynarec_opcodes_db_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_db_a32 : x86_dynarec_opcodes_db_a16;
@ -845,6 +853,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xdc: case 0xdc:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_dc_a32 : x86_dynarec_opcodes_dc_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_dc_a32 : x86_dynarec_opcodes_dc_a16;
@ -854,6 +863,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xdd: case 0xdd:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_dd_a32 : x86_dynarec_opcodes_dd_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_dd_a32 : x86_dynarec_opcodes_dd_a16;
@ -862,6 +872,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xde: case 0xde:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_de_a32 : x86_dynarec_opcodes_de_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_de_a32 : x86_dynarec_opcodes_de_a16;
@ -870,6 +881,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xdf: case 0xdf:
op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_df_a32 : x86_dynarec_opcodes_df_a16; op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_df_a32 : x86_dynarec_opcodes_df_a16;
@ -878,6 +890,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
over = 1; over = 1;
pc_off = -1; pc_off = -1;
test_modrm = 0; test_modrm = 0;
block->flags |= CODEBLOCK_HAS_FPU;
break; break;
case 0xf0: /*LOCK*/ case 0xf0: /*LOCK*/