Skip to content

Commit 8105736

Browse files
committed
py: Clean up x86-64 native assembler; allow use of extended regs.
Native x86-64 now has 3 locals in registers.
1 parent 25d9041 commit 8105736

6 files changed

Lines changed: 130 additions & 105 deletions

File tree

py/asmarm.h

Lines changed: 0 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -48,12 +48,6 @@
4848
#define REG_LR (REG_R14)
4949
#define REG_PC (REG_R15)
5050

51-
#define REG_RET REG_R0
52-
#define REG_ARG_1 REG_R0
53-
#define REG_ARG_2 REG_R1
54-
#define REG_ARG_3 REG_R2
55-
#define REG_ARG_4 REG_R3
56-
5751
#define ARM_CC_EQ (0x0 << 28)
5852
#define ARM_CC_NE (0x1 << 28)
5953
#define ARM_CC_CS (0x2 << 28)

py/asmthumb.h

Lines changed: 0 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -45,12 +45,6 @@
4545
#define REG_R15 (15)
4646
#define REG_LR (REG_R14)
4747

48-
#define REG_RET REG_R0
49-
#define REG_ARG_1 REG_R0
50-
#define REG_ARG_2 REG_R1
51-
#define REG_ARG_3 REG_R2
52-
#define REG_ARG_4 REG_R3
53-
5448
#define THUMB_CC_EQ (0x0)
5549
#define THUMB_CC_NE (0x1)
5650
#define THUMB_CC_CS (0x2)

py/asmx64.c

Lines changed: 64 additions & 49 deletions
Original file line numberDiff line numberDiff line change
@@ -41,15 +41,15 @@
4141
#define WORD_SIZE (8)
4242

4343
#define OPCODE_NOP (0x90)
44-
#define OPCODE_PUSH_R64 (0x50)
44+
#define OPCODE_PUSH_R64 (0x50) /* +rq */
4545
#define OPCODE_PUSH_I64 (0x68)
4646
#define OPCODE_PUSH_M64 (0xff) /* /6 */
47-
#define OPCODE_POP_R64 (0x58)
47+
#define OPCODE_POP_R64 (0x58) /* +rq */
4848
#define OPCODE_RET (0xc3)
4949
#define OPCODE_MOV_I8_TO_R8 (0xb0) /* +rb */
50-
#define OPCODE_MOV_I64_TO_R64 (0xb8)
50+
#define OPCODE_MOV_I64_TO_R64 (0xb8) /* +rq */
5151
#define OPCODE_MOV_I32_TO_RM32 (0xc7)
52-
#define OPCODE_MOV_R64_TO_RM64 (0x89)
52+
#define OPCODE_MOV_R64_TO_RM64 (0x89) /* /r */
5353
#define OPCODE_MOV_RM64_TO_R64 (0x8b)
5454
#define OPCODE_LEA_MEM_TO_R64 (0x8d) /* /r */
5555
#define OPCODE_XOR_R64_TO_RM64 (0x31) /* /r */
@@ -78,12 +78,12 @@
7878
#define OPCODE_CALL_RM32 (0xff) /* /2 */
7979
#define OPCODE_LEAVE (0xc9)
8080

81-
#define MODRM_R64(x) ((x) << 3)
81+
#define MODRM_R64(x) (((x) & 0x7) << 3)
8282
#define MODRM_RM_DISP0 (0x00)
8383
#define MODRM_RM_DISP8 (0x40)
8484
#define MODRM_RM_DISP32 (0x80)
8585
#define MODRM_RM_REG (0xc0)
86-
#define MODRM_RM_R64(x) (x)
86+
#define MODRM_RM_R64(x) ((x) & 0x7)
8787

8888
#define REX_PREFIX (0x40)
8989
#define REX_W (0x08) // width
@@ -248,6 +248,7 @@ STATIC void asm_x64_write_word32_to(asm_x64_t *as, int offset, int w32) {
248248
*/
249249

250250
STATIC void asm_x64_write_r64_disp(asm_x64_t *as, int r64, int disp_r64, int disp_offset) {
251+
assert(disp_r64 < 8);
251252
assert(disp_r64 != REG_RSP);
252253

253254
if (disp_offset == 0 && disp_r64 != REG_RBP) {
@@ -265,21 +266,32 @@ void asm_x64_nop(asm_x64_t *as) {
265266
}
266267

267268
void asm_x64_push_r64(asm_x64_t *as, int src_r64) {
268-
asm_x64_write_byte_1(as, OPCODE_PUSH_R64 | src_r64);
269+
if (src_r64 < 8) {
270+
asm_x64_write_byte_1(as, OPCODE_PUSH_R64 | src_r64);
271+
} else {
272+
asm_x64_write_byte_2(as, REX_PREFIX | REX_B, OPCODE_PUSH_R64 | (src_r64 & 7));
273+
}
269274
}
270275

276+
/*
271277
void asm_x64_push_i32(asm_x64_t *as, int src_i32) {
272278
asm_x64_write_byte_1(as, OPCODE_PUSH_I64);
273279
asm_x64_write_word32(as, src_i32); // will be sign extended to 64 bits
274280
}
281+
*/
275282

276283
void asm_x64_push_disp(asm_x64_t *as, int src_r64, int src_offset) {
284+
assert(src_r64 < 8);
277285
asm_x64_write_byte_1(as, OPCODE_PUSH_M64);
278286
asm_x64_write_r64_disp(as, 6, src_r64, src_offset);
279287
}
280288

281289
void asm_x64_pop_r64(asm_x64_t *as, int dest_r64) {
282-
asm_x64_write_byte_1(as, OPCODE_POP_R64 | dest_r64);
290+
if (dest_r64 < 8) {
291+
asm_x64_write_byte_1(as, OPCODE_POP_R64 | dest_r64);
292+
} else {
293+
asm_x64_write_byte_2(as, REX_PREFIX | REX_B, OPCODE_POP_R64 | (dest_r64 & 7));
294+
}
283295
}
284296

285297
STATIC void asm_x64_ret(asm_x64_t *as) {
@@ -288,45 +300,57 @@ STATIC void asm_x64_ret(asm_x64_t *as) {
288300

289301
void asm_x64_mov_r32_to_r32(asm_x64_t *as, int src_r32, int dest_r32) {
290302
// defaults to 32 bit operation
303+
assert(src_r32 < 8);
304+
assert(dest_r32 < 8);
291305
asm_x64_write_byte_2(as, OPCODE_MOV_R64_TO_RM64, MODRM_R64(src_r32) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
292306
}
293307

294308
void asm_x64_mov_r64_to_r64(asm_x64_t *as, int src_r64, int dest_r64) {
295309
// use REX prefix for 64 bit operation
296-
asm_x64_write_byte_3(as, REX_PREFIX | REX_W, OPCODE_MOV_R64_TO_RM64, MODRM_R64(src_r64) | MODRM_RM_REG | MODRM_RM_R64(dest_r64));
310+
asm_x64_write_byte_3(as, REX_PREFIX | REX_W | (src_r64 < 8 ? 0 : REX_R) | (dest_r64 < 8 ? 0 : REX_B), OPCODE_MOV_R64_TO_RM64, MODRM_R64(src_r64) | MODRM_RM_REG | MODRM_RM_R64(dest_r64));
297311
}
298312

299313
void asm_x64_mov_r64_to_disp(asm_x64_t *as, int src_r64, int dest_r64, int dest_disp) {
300314
// use REX prefix for 64 bit operation
301-
asm_x64_write_byte_2(as, REX_PREFIX | REX_W, OPCODE_MOV_R64_TO_RM64);
315+
assert(dest_r64 < 8);
316+
asm_x64_write_byte_2(as, REX_PREFIX | REX_W | (src_r64 < 8 ? 0 : REX_R), OPCODE_MOV_R64_TO_RM64);
302317
asm_x64_write_r64_disp(as, src_r64, dest_r64, dest_disp);
303318
}
304319

305320
void asm_x64_mov_disp_to_r64(asm_x64_t *as, int src_r64, int src_disp, int dest_r64) {
306321
// use REX prefix for 64 bit operation
307-
asm_x64_write_byte_2(as, REX_PREFIX | REX_W, OPCODE_MOV_RM64_TO_R64);
322+
assert(src_r64 < 8);
323+
asm_x64_write_byte_2(as, REX_PREFIX | REX_W | (dest_r64 < 8 ? 0 : REX_R), OPCODE_MOV_RM64_TO_R64);
308324
asm_x64_write_r64_disp(as, dest_r64, src_r64, src_disp);
309325
}
310326

311327
void asm_x64_lea_disp_to_r64(asm_x64_t *as, int src_r64, int src_disp, int dest_r64) {
312328
// use REX prefix for 64 bit operation
329+
assert(src_r64 < 8);
330+
assert(dest_r64 < 8);
313331
asm_x64_write_byte_2(as, REX_PREFIX | REX_W, OPCODE_LEA_MEM_TO_R64);
314332
asm_x64_write_r64_disp(as, dest_r64, src_r64, src_disp);
315333
}
316334

317335
void asm_x64_mov_i8_to_r8(asm_x64_t *as, int src_i8, int dest_r64) {
336+
assert(dest_r64 < 8);
318337
asm_x64_write_byte_2(as, OPCODE_MOV_I8_TO_R8 | dest_r64, src_i8);
319338
}
320339

321-
void asm_x64_mov_i32_to_r64(asm_x64_t *as, int src_i32, int dest_r64) {
340+
STATIC void asm_x64_mov_i32_to_r64(asm_x64_t *as, int src_i32, int dest_r64) {
322341
// cpu defaults to i32 to r64, with zero extension
323-
asm_x64_write_byte_1(as, OPCODE_MOV_I64_TO_R64 | dest_r64);
342+
if (dest_r64 < 8) {
343+
asm_x64_write_byte_1(as, OPCODE_MOV_I64_TO_R64 | dest_r64);
344+
} else {
345+
asm_x64_write_byte_2(as, REX_PREFIX | REX_B, OPCODE_MOV_I64_TO_R64 | (dest_r64 & 7));
346+
}
324347
asm_x64_write_word32(as, src_i32);
325348
}
326349

327350
void asm_x64_mov_i64_to_r64(asm_x64_t *as, int64_t src_i64, int dest_r64) {
328351
// cpu defaults to i32 to r64
329352
// to mov i64 to r64 need to use REX prefix
353+
assert(dest_r64 < 8);
330354
asm_x64_write_byte_2(as, REX_PREFIX | REX_W, OPCODE_MOV_I64_TO_R64 | dest_r64);
331355
asm_x64_write_word64(as, src_i64);
332356
}
@@ -350,22 +374,19 @@ void asm_x64_mov_i64_to_r64_aligned(asm_x64_t *as, int64_t src_i64, int dest_r64
350374
asm_x64_mov_i64_to_r64(as, src_i64, dest_r64);
351375
}
352376

353-
void asm_x64_mov_i32_to_disp(asm_x64_t *as, int src_i32, int dest_r32, int dest_disp)
354-
{
355-
assert(0);
356-
asm_x64_write_byte_1(as, OPCODE_MOV_I32_TO_RM32);
357-
//asm_x64_write_r32_disp(as, 0, dest_r32, dest_disp);
358-
asm_x64_write_word32(as, src_i32);
359-
}
360-
361377
void asm_x64_xor_r64_to_r64(asm_x64_t *as, int src_r64, int dest_r64) {
378+
assert(src_r64 < 8);
379+
assert(dest_r64 < 8);
362380
asm_x64_write_byte_3(as, REX_PREFIX | REX_W, OPCODE_XOR_R64_TO_RM64, MODRM_R64(src_r64) | MODRM_RM_REG | MODRM_RM_R64(dest_r64));
363381
}
364382

365383
void asm_x64_add_r64_to_r64(asm_x64_t *as, int src_r64, int dest_r64) {
384+
assert(src_r64 < 8);
385+
assert(dest_r64 < 8);
366386
asm_x64_write_byte_3(as, REX_PREFIX | REX_W, OPCODE_ADD_R64_TO_RM64, MODRM_R64(src_r64) | MODRM_RM_REG | MODRM_RM_R64(dest_r64));
367387
}
368388

389+
/*
369390
void asm_x64_add_i32_to_r32(asm_x64_t *as, int src_i32, int dest_r32)
370391
{
371392
assert(dest_r32 != REG_RSP); // in this case i think src_i32 must be 64 bits
@@ -380,17 +401,23 @@ void asm_x64_add_i32_to_r32(asm_x64_t *as, int src_i32, int dest_r32)
380401
asm_x64_write_word32(as, src_i32);
381402
}
382403
}
404+
*/
383405

406+
/*
384407
void asm_x64_sub_r32_from_r32(asm_x64_t *as, int src_r32, int dest_r32) {
385408
// defaults to 32 bit operation
386409
asm_x64_write_byte_2(as, OPCODE_SUB_R64_FROM_RM64, MODRM_R64(src_r32) | MODRM_RM_REG | MODRM_RM_R64(dest_r32));
387410
}
411+
*/
388412

389413
void asm_x64_sub_r64_from_r64(asm_x64_t *as, int src_r64, int dest_r64) {
390414
// use REX prefix for 64 bit operation
415+
assert(src_r64 < 8);
416+
assert(dest_r64 < 8);
391417
asm_x64_write_byte_3(as, REX_PREFIX | REX_W, OPCODE_SUB_R64_FROM_RM64, MODRM_R64(src_r64) | MODRM_RM_REG | MODRM_RM_R64(dest_r64));
392418
}
393419

420+
/*
394421
void asm_x64_sub_i32_from_r32(asm_x64_t *as, int src_i32, int dest_r32) {
395422
if (SIGNED_FIT8(src_i32)) {
396423
// defaults to 32 bit operation
@@ -402,8 +429,10 @@ void asm_x64_sub_i32_from_r32(asm_x64_t *as, int src_i32, int dest_r32) {
402429
asm_x64_write_word32(as, src_i32);
403430
}
404431
}
432+
*/
405433

406434
void asm_x64_sub_i32_from_r64(asm_x64_t *as, int src_i32, int dest_r64) {
435+
assert(dest_r64 < 8);
407436
if (SIGNED_FIT8(src_i32)) {
408437
// use REX prefix for 64 bit operation
409438
asm_x64_write_byte_3(as, REX_PREFIX | REX_W, OPCODE_SUB_I8_FROM_RM64, MODRM_R64(5) | MODRM_RM_REG | MODRM_RM_R64(dest_r64));
@@ -415,7 +444,7 @@ void asm_x64_sub_i32_from_r64(asm_x64_t *as, int src_i32, int dest_r64) {
415444
}
416445
}
417446

418-
/* shifts not tested */
447+
/*
419448
void asm_x64_shl_r32_by_imm(asm_x64_t *as, int r32, int imm) {
420449
asm_x64_write_byte_2(as, OPCODE_SHL_RM32_BY_I8, MODRM_R64(4) | MODRM_RM_REG | MODRM_RM_R64(r32));
421450
asm_x64_write_byte_1(as, imm);
@@ -430,23 +459,15 @@ void asm_x64_sar_r32_by_imm(asm_x64_t *as, int r32, int imm) {
430459
asm_x64_write_byte_2(as, OPCODE_SAR_RM32_BY_I8, MODRM_R64(7) | MODRM_RM_REG | MODRM_RM_R64(r32));
431460
asm_x64_write_byte_1(as, imm);
432461
}
462+
*/
433463

434464
void asm_x64_cmp_r64_with_r64(asm_x64_t *as, int src_r64_a, int src_r64_b) {
465+
assert(src_r64_a < 8);
466+
assert(src_r64_b < 8);
435467
asm_x64_write_byte_3(as, REX_PREFIX | REX_W, OPCODE_CMP_R64_WITH_RM64, MODRM_R64(src_r64_a) | MODRM_RM_REG | MODRM_RM_R64(src_r64_b));
436468
}
437469

438-
void asm_x64_cmp_r32_with_disp(asm_x64_t *as, int src_r32_a, int src_r32_b, int src_disp_b) {
439-
assert(0);
440-
asm_x64_write_byte_1(as, OPCODE_CMP_R64_WITH_RM64);
441-
//asm_x64_write_r32_disp(as, src_r32_a, src_r32_b, src_disp_b);
442-
}
443-
444-
void asm_x64_cmp_disp_with_r32(asm_x64_t *as, int src_r32_a, int src_disp_a, int src_r32_b) {
445-
assert(0);
446-
asm_x64_write_byte_1(as, OPCODE_CMP_RM32_WITH_R32);
447-
//asm_x64_write_r32_disp(as, src_r32_b, src_r32_a, src_disp_a);
448-
}
449-
470+
/*
450471
void asm_x64_cmp_i32_with_r32(asm_x64_t *as, int src_i32, int src_r32) {
451472
if (SIGNED_FIT8(src_i32)) {
452473
asm_x64_write_byte_2(as, OPCODE_CMP_I8_WITH_RM32, MODRM_R64(7) | MODRM_RM_REG | MODRM_RM_R64(src_r32));
@@ -456,6 +477,7 @@ void asm_x64_cmp_i32_with_r32(asm_x64_t *as, int src_i32, int src_r32) {
456477
asm_x64_write_word32(as, src_i32);
457478
}
458479
}
480+
*/
459481

460482
void asm_x64_test_r8_with_r8(asm_x64_t *as, int src_r64_a, int src_r64_b) {
461483
// TODO implement for other registers
@@ -465,6 +487,7 @@ void asm_x64_test_r8_with_r8(asm_x64_t *as, int src_r64_a, int src_r64_b) {
465487
}
466488

467489
void asm_x64_setcc_r8(asm_x64_t *as, int jcc_type, int dest_r8) {
490+
assert(dest_r8 < 8);
468491
asm_x64_write_byte_3(as, OPCODE_SETCC_RM8_A, OPCODE_SETCC_RM8_B | jcc_type, MODRM_R64(0) | MODRM_RM_REG | MODRM_RM_R64(dest_r8));
469492
}
470493

@@ -539,30 +562,19 @@ void asm_x64_entry(asm_x64_t *as, int num_locals) {
539562
num_locals |= 1; // make it odd so stack is aligned on 16 byte boundary
540563
asm_x64_sub_i32_from_r64(as, num_locals * WORD_SIZE, REG_RSP);
541564
asm_x64_push_r64(as, REG_RBX);
565+
asm_x64_push_r64(as, REG_R12);
566+
asm_x64_push_r64(as, REG_R13);
542567
as->num_locals = num_locals;
543568
}
544569

545570
void asm_x64_exit(asm_x64_t *as) {
571+
asm_x64_pop_r64(as, REG_R13);
572+
asm_x64_pop_r64(as, REG_R12);
546573
asm_x64_pop_r64(as, REG_RBX);
547574
asm_x64_write_byte_1(as, OPCODE_LEAVE);
548575
asm_x64_ret(as);
549576
}
550577

551-
void asm_x64_push_arg(asm_x64_t *as, int src_arg_num) {
552-
assert(0);
553-
asm_x64_push_disp(as, REG_RBP, 8 + src_arg_num * WORD_SIZE);
554-
}
555-
556-
void asm_x64_mov_arg_to_r32(asm_x64_t *as, int src_arg_num, int dest_r32) {
557-
assert(0);
558-
//asm_x64_mov_disp_to_r32(as, REG_RBP, 8 + src_arg_num * WORD_SIZE, dest_r32);
559-
}
560-
561-
void asm_x64_mov_r32_to_arg(asm_x64_t *as, int src_r32, int dest_arg_num) {
562-
assert(0);
563-
//asm_x64_mov_r32_to_disp(as, src_r32, REG_RBP, 8 + dest_arg_num * WORD_SIZE);
564-
}
565-
566578
// locals:
567579
// - stored on the stack in ascending order
568580
// - numbered 0 through as->num_locals-1
@@ -595,6 +607,7 @@ void asm_x64_mov_local_addr_to_r64(asm_x64_t *as, int local_num, int dest_r64) {
595607
}
596608
}
597609

610+
/*
598611
void asm_x64_push_local(asm_x64_t *as, int local_num) {
599612
asm_x64_push_disp(as, REG_RBP, asm_x64_local_offset_from_ebp(as, local_num));
600613
}
@@ -605,6 +618,7 @@ void asm_x64_push_local_addr(asm_x64_t *as, int local_num, int temp_r64)
605618
asm_x64_add_i32_to_r32(as, asm_x64_local_offset_from_ebp(as, local_num), temp_r64);
606619
asm_x64_push_r64(as, temp_r64);
607620
}
621+
*/
608622

609623
/*
610624
can't use these because code might be relocated when resized
@@ -630,6 +644,7 @@ void asm_x64_call_i1(asm_x64_t *as, void* func, int i1)
630644
*/
631645

632646
void asm_x64_call_ind(asm_x64_t *as, void *ptr, int temp_r64) {
647+
assert(temp_r64 < 8);
633648
#ifdef __LP64__
634649
asm_x64_mov_i64_to_r64_optimised(as, (int64_t)ptr, temp_r64);
635650
#else

0 commit comments

Comments
 (0)