diff options
Diffstat (limited to 'src/3rdparty/pcre/sljit/sljitNativeARM_32.c')
| -rw-r--r-- | src/3rdparty/pcre/sljit/sljitNativeARM_32.c | 499 |
1 files changed, 263 insertions, 236 deletions
diff --git a/src/3rdparty/pcre/sljit/sljitNativeARM_32.c b/src/3rdparty/pcre/sljit/sljitNativeARM_32.c index 6747c4f617..aca1d31fdf 100644 --- a/src/3rdparty/pcre/sljit/sljitNativeARM_32.c +++ b/src/3rdparty/pcre/sljit/sljitNativeARM_32.c @@ -36,13 +36,13 @@ SLJIT_API_FUNC_ATTRIBUTE SLJIT_CONST char* sljit_get_platform_name(void) } /* Last register + 1. */ -#define TMP_REG1 (SLJIT_NO_REGISTERS + 1) -#define TMP_REG2 (SLJIT_NO_REGISTERS + 2) -#define TMP_REG3 (SLJIT_NO_REGISTERS + 3) -#define TMP_PC (SLJIT_NO_REGISTERS + 4) +#define TMP_REG1 (SLJIT_NUMBER_OF_REGISTERS + 2) +#define TMP_REG2 (SLJIT_NUMBER_OF_REGISTERS + 3) +#define TMP_REG3 (SLJIT_NUMBER_OF_REGISTERS + 4) +#define TMP_PC (SLJIT_NUMBER_OF_REGISTERS + 5) #define TMP_FREG1 (0) -#define TMP_FREG2 (SLJIT_FLOAT_REG6 + 1) +#define TMP_FREG2 (SLJIT_NUMBER_OF_FLOAT_REGISTERS + 1) /* In ARM instruction words. Cache lines are usually 32 byte aligned. */ @@ -55,8 +55,8 @@ SLJIT_API_FUNC_ATTRIBUTE SLJIT_CONST char* sljit_get_platform_name(void) (((max_diff) / (sljit_si)sizeof(sljit_uw)) - (CONST_POOL_ALIGNMENT - 1)) /* See sljit_emit_enter and sljit_emit_op0 if you want to change them. */ -static SLJIT_CONST sljit_ub reg_map[SLJIT_NO_REGISTERS + 5] = { - 0, 0, 1, 2, 10, 11, 4, 5, 6, 7, 8, 13, 3, 12, 14, 15 +static SLJIT_CONST sljit_ub reg_map[SLJIT_NUMBER_OF_REGISTERS + 6] = { + 0, 0, 1, 2, 11, 10, 9, 8, 7, 6, 5, 4, 13, 3, 12, 14, 15 }; #define RM(rm) (reg_map[rm]) @@ -102,8 +102,12 @@ static SLJIT_CONST sljit_ub reg_map[SLJIT_NO_REGISTERS + 5] = { #define VABS_F32 0xeeb00ac0 #define VADD_F32 0xee300a00 #define VCMP_F32 0xeeb40a40 +#define VCVT_F32_S32 0xeeb80ac0 +#define VCVT_F64_F32 0xeeb70ac0 +#define VCVT_S32_F32 0xeebd0ac0 #define VDIV_F32 0xee800a00 #define VMOV_F32 0xeeb00a40 +#define VMOV 0xee000a10 #define VMRS 0xeef1fa10 #define VMUL_F32 0xee200a00 #define VNEG_F32 0xeeb10a40 @@ -311,11 +315,13 @@ struct future_patch { sljit_si value; }; -static SLJIT_INLINE sljit_si resolve_const_pool_index(struct future_patch **first_patch, sljit_uw cpool_current_index, sljit_uw *cpool_start_address, sljit_uw *buf_ptr) +static sljit_si resolve_const_pool_index(struct sljit_compiler *compiler, struct future_patch **first_patch, sljit_uw cpool_current_index, sljit_uw *cpool_start_address, sljit_uw *buf_ptr) { sljit_si value; struct future_patch *curr_patch, *prev_patch; + SLJIT_UNUSED_ARG(compiler); + /* Using the values generated by patch_pc_relative_loads. */ if (!*first_patch) value = (sljit_si)cpool_start_address[cpool_current_index]; @@ -333,7 +339,7 @@ static SLJIT_INLINE sljit_si resolve_const_pool_index(struct future_patch **firs prev_patch->next = curr_patch->next; else *first_patch = curr_patch->next; - SLJIT_FREE(curr_patch); + SLJIT_FREE(curr_patch, compiler->allocator_data); break; } prev_patch = curr_patch; @@ -343,12 +349,12 @@ static SLJIT_INLINE sljit_si resolve_const_pool_index(struct future_patch **firs if (value >= 0) { if ((sljit_uw)value > cpool_current_index) { - curr_patch = (struct future_patch*)SLJIT_MALLOC(sizeof(struct future_patch)); + curr_patch = (struct future_patch*)SLJIT_MALLOC(sizeof(struct future_patch), compiler->allocator_data); if (!curr_patch) { while (*first_patch) { curr_patch = *first_patch; *first_patch = (*first_patch)->next; - SLJIT_FREE(curr_patch); + SLJIT_FREE(curr_patch, compiler->allocator_data); } return SLJIT_ERR_ALLOC_FAILED; } @@ -570,7 +576,7 @@ SLJIT_API_FUNC_ATTRIBUTE void* sljit_generate_code(struct sljit_compiler *compil struct sljit_const *const_; CHECK_ERROR_PTR(); - check_sljit_generate_code(compiler); + CHECK_PTR(check_sljit_generate_code(compiler)); reverse_buf(compiler); /* Second code generation pass. */ @@ -619,7 +625,7 @@ SLJIT_API_FUNC_ATTRIBUTE void* sljit_generate_code(struct sljit_compiler *compil cpool_skip_alignment--; } else { - if (SLJIT_UNLIKELY(resolve_const_pool_index(&first_patch, cpool_current_index, cpool_start_address, buf_ptr))) { + if (SLJIT_UNLIKELY(resolve_const_pool_index(compiler, &first_patch, cpool_current_index, cpool_start_address, buf_ptr))) { SLJIT_FREE_EXEC(code); compiler->error = SLJIT_ERR_ALLOC_FAILED; return NULL; @@ -709,7 +715,7 @@ SLJIT_API_FUNC_ATTRIBUTE void* sljit_generate_code(struct sljit_compiler *compil buf_end = buf_ptr + compiler->cpool_fill; cpool_current_index = 0; while (buf_ptr < buf_end) { - if (SLJIT_UNLIKELY(resolve_const_pool_index(&first_patch, cpool_current_index, cpool_start_address, buf_ptr))) { + if (SLJIT_UNLIKELY(resolve_const_pool_index(compiler, &first_patch, cpool_current_index, cpool_start_address, buf_ptr))) { SLJIT_FREE_EXEC(code); compiler->error = SLJIT_ERR_ALLOC_FAILED; return NULL; @@ -810,9 +816,6 @@ SLJIT_API_FUNC_ATTRIBUTE void* sljit_generate_code(struct sljit_compiler *compil #define SIGNED_DATA 0x40 #define LOAD_DATA 0x80 -#define EMIT_INSTRUCTION(inst) \ - FAIL_IF(push_inst(compiler, (inst))) - /* Condition: AL. */ #define EMIT_DATA_PROCESS_INS(opcode, set_flags, dst, src1, src2) \ (0xe0000000 | ((opcode) << 21) | (set_flags) | RD(dst) | RN(src1) | (src2)) @@ -822,111 +825,85 @@ static sljit_si emit_op(struct sljit_compiler *compiler, sljit_si op, sljit_si i sljit_si src1, sljit_sw src1w, sljit_si src2, sljit_sw src2w); -SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_enter(struct sljit_compiler *compiler, sljit_si args, sljit_si scratches, sljit_si saveds, sljit_si local_size) +SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_enter(struct sljit_compiler *compiler, + sljit_si options, sljit_si args, sljit_si scratches, sljit_si saveds, + sljit_si fscratches, sljit_si fsaveds, sljit_si local_size) { - sljit_si size; + sljit_si size, i, tmp; sljit_uw push; CHECK_ERROR(); - check_sljit_emit_enter(compiler, args, scratches, saveds, local_size); - - compiler->scratches = scratches; - compiler->saveds = saveds; -#if (defined SLJIT_DEBUG && SLJIT_DEBUG) - compiler->logical_local_size = local_size; -#endif + CHECK(check_sljit_emit_enter(compiler, options, args, scratches, saveds, fscratches, fsaveds, local_size)); + set_emit_enter(compiler, options, args, scratches, saveds, fscratches, fsaveds, local_size); /* Push saved registers, temporary registers stmdb sp!, {..., lr} */ push = PUSH | (1 << 14); - if (scratches >= 5) - push |= 1 << 11; - if (scratches >= 4) - push |= 1 << 10; - if (saveds >= 5) - push |= 1 << 8; - if (saveds >= 4) - push |= 1 << 7; - if (saveds >= 3) - push |= 1 << 6; - if (saveds >= 2) - push |= 1 << 5; - if (saveds >= 1) - push |= 1 << 4; - EMIT_INSTRUCTION(push); + + tmp = saveds < SLJIT_NUMBER_OF_SAVED_REGISTERS ? (SLJIT_S0 + 1 - saveds) : SLJIT_FIRST_SAVED_REG; + for (i = SLJIT_S0; i >= tmp; i--) + push |= 1 << reg_map[i]; + + for (i = scratches; i >= SLJIT_FIRST_SAVED_REG; i--) + push |= 1 << reg_map[i]; + + FAIL_IF(push_inst(compiler, push)); /* Stack must be aligned to 8 bytes: */ - size = (1 + saveds) * sizeof(sljit_uw); - if (scratches >= 4) - size += (scratches - 3) * sizeof(sljit_uw); - local_size += size; - local_size = (local_size + 7) & ~7; - local_size -= size; + size = GET_SAVED_REGISTERS_SIZE(scratches, saveds, 1); + local_size = ((size + local_size + 7) & ~7) - size; compiler->local_size = local_size; if (local_size > 0) - FAIL_IF(emit_op(compiler, SLJIT_SUB, ALLOW_IMM, SLJIT_LOCALS_REG, 0, SLJIT_LOCALS_REG, 0, SLJIT_IMM, local_size)); + FAIL_IF(emit_op(compiler, SLJIT_SUB, ALLOW_IMM, SLJIT_SP, 0, SLJIT_SP, 0, SLJIT_IMM, local_size)); if (args >= 1) - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, SLJIT_SAVED_REG1, SLJIT_UNUSED, RM(SLJIT_SCRATCH_REG1))); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, SLJIT_S0, SLJIT_UNUSED, RM(SLJIT_R0)))); if (args >= 2) - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, SLJIT_SAVED_REG2, SLJIT_UNUSED, RM(SLJIT_SCRATCH_REG2))); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, SLJIT_S1, SLJIT_UNUSED, RM(SLJIT_R1)))); if (args >= 3) - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, SLJIT_SAVED_REG3, SLJIT_UNUSED, RM(SLJIT_SCRATCH_REG3))); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, SLJIT_S2, SLJIT_UNUSED, RM(SLJIT_R2)))); return SLJIT_SUCCESS; } -SLJIT_API_FUNC_ATTRIBUTE void sljit_set_context(struct sljit_compiler *compiler, sljit_si args, sljit_si scratches, sljit_si saveds, sljit_si local_size) +SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_set_context(struct sljit_compiler *compiler, + sljit_si options, sljit_si args, sljit_si scratches, sljit_si saveds, + sljit_si fscratches, sljit_si fsaveds, sljit_si local_size) { sljit_si size; - CHECK_ERROR_VOID(); - check_sljit_set_context(compiler, args, scratches, saveds, local_size); - - compiler->scratches = scratches; - compiler->saveds = saveds; -#if (defined SLJIT_DEBUG && SLJIT_DEBUG) - compiler->logical_local_size = local_size; -#endif + CHECK_ERROR(); + CHECK(check_sljit_set_context(compiler, options, args, scratches, saveds, fscratches, fsaveds, local_size)); + set_set_context(compiler, options, args, scratches, saveds, fscratches, fsaveds, local_size); - size = (1 + saveds) * sizeof(sljit_uw); - if (scratches >= 4) - size += (scratches - 3) * sizeof(sljit_uw); - local_size += size; - local_size = (local_size + 7) & ~7; - local_size -= size; - compiler->local_size = local_size; + size = GET_SAVED_REGISTERS_SIZE(scratches, saveds, 1); + compiler->local_size = ((size + local_size + 7) & ~7) - size; + return SLJIT_SUCCESS; } SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_return(struct sljit_compiler *compiler, sljit_si op, sljit_si src, sljit_sw srcw) { + sljit_si i, tmp; sljit_uw pop; CHECK_ERROR(); - check_sljit_emit_return(compiler, op, src, srcw); + CHECK(check_sljit_emit_return(compiler, op, src, srcw)); FAIL_IF(emit_mov_before_return(compiler, op, src, srcw)); if (compiler->local_size > 0) - FAIL_IF(emit_op(compiler, SLJIT_ADD, ALLOW_IMM, SLJIT_LOCALS_REG, 0, SLJIT_LOCALS_REG, 0, SLJIT_IMM, compiler->local_size)); + FAIL_IF(emit_op(compiler, SLJIT_ADD, ALLOW_IMM, SLJIT_SP, 0, SLJIT_SP, 0, SLJIT_IMM, compiler->local_size)); - pop = POP | (1 << 15); /* Push saved registers, temporary registers ldmia sp!, {..., pc} */ - if (compiler->scratches >= 5) - pop |= 1 << 11; - if (compiler->scratches >= 4) - pop |= 1 << 10; - if (compiler->saveds >= 5) - pop |= 1 << 8; - if (compiler->saveds >= 4) - pop |= 1 << 7; - if (compiler->saveds >= 3) - pop |= 1 << 6; - if (compiler->saveds >= 2) - pop |= 1 << 5; - if (compiler->saveds >= 1) - pop |= 1 << 4; + pop = POP | (1 << 15); + + tmp = compiler->saveds < SLJIT_NUMBER_OF_SAVED_REGISTERS ? (SLJIT_S0 + 1 - compiler->saveds) : SLJIT_FIRST_SAVED_REG; + for (i = SLJIT_S0; i >= tmp; i--) + pop |= 1 << reg_map[i]; + + for (i = compiler->scratches; i >= SLJIT_FIRST_SAVED_REG; i--) + pop |= 1 << reg_map[i]; return push_inst(compiler, pop); } @@ -1031,7 +1008,7 @@ static SLJIT_INLINE sljit_si emit_single_op(struct sljit_compiler *compiler, slj #if (defined SLJIT_CONFIG_ARM_V5 && SLJIT_CONFIG_ARM_V5) if (op == SLJIT_MOV_UB) return push_inst(compiler, EMIT_DATA_PROCESS_INS(AND_DP, 0, dst, src2, SRC2_IMM | 0xff)); - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, dst, SLJIT_UNUSED, (24 << 7) | reg_map[src2])); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, dst, SLJIT_UNUSED, (24 << 7) | reg_map[src2]))); return push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, dst, SLJIT_UNUSED, (24 << 7) | (op == SLJIT_MOV_UB ? 0x20 : 0x40) | reg_map[dst])); #else return push_inst(compiler, (op == SLJIT_MOV_UB ? UXTB : SXTB) | RD(dst) | RM(src2)); @@ -1050,7 +1027,7 @@ static SLJIT_INLINE sljit_si emit_single_op(struct sljit_compiler *compiler, slj SLJIT_ASSERT(src1 == TMP_REG1 && !(flags & ARGS_SWAPPED)); if ((flags & (REG_DEST | REG_SOURCE)) == (REG_DEST | REG_SOURCE)) { #if (defined SLJIT_CONFIG_ARM_V5 && SLJIT_CONFIG_ARM_V5) - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, dst, SLJIT_UNUSED, (16 << 7) | reg_map[src2])); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, dst, SLJIT_UNUSED, (16 << 7) | reg_map[src2]))); return push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, dst, SLJIT_UNUSED, (16 << 7) | (op == SLJIT_MOV_UH ? 0x20 : 0x40) | reg_map[dst])); #else return push_inst(compiler, (op == SLJIT_MOV_UH ? UXTH : SXTH) | RD(dst) | RM(src2)); @@ -1303,8 +1280,8 @@ static sljit_si generate_int(struct sljit_compiler *compiler, sljit_si reg, slji return 0; } - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(positive ? MOV_DP : MVN_DP, 0, reg, SLJIT_UNUSED, imm1)); - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(positive ? ORR_DP : BIC_DP, 0, reg, reg, imm2)); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(positive ? MOV_DP : MVN_DP, 0, reg, SLJIT_UNUSED, imm1))); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(positive ? ORR_DP : BIC_DP, 0, reg, reg, imm2))); return 1; } #endif @@ -1320,16 +1297,12 @@ static sljit_si load_immediate(struct sljit_compiler *compiler, sljit_si reg, sl /* Create imm by 1 inst. */ tmp = get_imm(imm); - if (tmp) { - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, reg, SLJIT_UNUSED, tmp)); - return SLJIT_SUCCESS; - } + if (tmp) + return push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, reg, SLJIT_UNUSED, tmp)); tmp = get_imm(~imm); - if (tmp) { - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MVN_DP, 0, reg, SLJIT_UNUSED, tmp)); - return SLJIT_SUCCESS; - } + if (tmp) + return push_inst(compiler, EMIT_DATA_PROCESS_INS(MVN_DP, 0, reg, SLJIT_UNUSED, tmp)); #if (defined SLJIT_CONFIG_ARM_V5 && SLJIT_CONFIG_ARM_V5) /* Create imm by 2 inst. */ @@ -1369,14 +1342,14 @@ static sljit_si getput_arg_fast(struct sljit_compiler *compiler, sljit_si inp_fl if (imm) { if (inp_flags & ARG_TEST) return 1; - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, reg, SLJIT_UNUSED, imm)); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, reg, SLJIT_UNUSED, imm))); return -1; } imm = get_imm(~argw); if (imm) { if (inp_flags & ARG_TEST) return 1; - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MVN_DP, 0, reg, SLJIT_UNUSED, imm)); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MVN_DP, 0, reg, SLJIT_UNUSED, imm))); return -1; } return 0; @@ -1394,8 +1367,8 @@ static sljit_si getput_arg_fast(struct sljit_compiler *compiler, sljit_si inp_fl if (inp_flags & ARG_TEST) return 1; - EMIT_INSTRUCTION(EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, - RM(OFFS_REG(arg)) | (IS_TYPE1_TRANSFER(inp_flags) ? SRC2_IMM : 0) | ((argw & 0x3) << 7))); + FAIL_IF(push_inst(compiler, EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, + RM(OFFS_REG(arg)) | (IS_TYPE1_TRANSFER(inp_flags) ? SRC2_IMM : 0) | ((argw & 0x3) << 7)))); return -1; } @@ -1403,13 +1376,13 @@ static sljit_si getput_arg_fast(struct sljit_compiler *compiler, sljit_si inp_fl if (argw >= 0 && argw <= 0xfff) { if (inp_flags & ARG_TEST) return 1; - EMIT_INSTRUCTION(EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, argw)); + FAIL_IF(push_inst(compiler, EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, argw))); return -1; } if (argw < 0 && argw >= -0xfff) { if (inp_flags & ARG_TEST) return 1; - EMIT_INSTRUCTION(EMIT_DATA_TRANSFER(inp_flags, 0, inp_flags & WRITE_BACK, reg, arg & REG_MASK, -argw)); + FAIL_IF(push_inst(compiler, EMIT_DATA_TRANSFER(inp_flags, 0, inp_flags & WRITE_BACK, reg, arg & REG_MASK, -argw))); return -1; } } @@ -1417,14 +1390,14 @@ static sljit_si getput_arg_fast(struct sljit_compiler *compiler, sljit_si inp_fl if (argw >= 0 && argw <= 0xff) { if (inp_flags & ARG_TEST) return 1; - EMIT_INSTRUCTION(EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, TYPE2_TRANSFER_IMM(argw))); + FAIL_IF(push_inst(compiler, EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, TYPE2_TRANSFER_IMM(argw)))); return -1; } if (argw < 0 && argw >= -0xff) { if (inp_flags & ARG_TEST) return 1; argw = -argw; - EMIT_INSTRUCTION(EMIT_DATA_TRANSFER(inp_flags, 0, inp_flags & WRITE_BACK, reg, arg & REG_MASK, TYPE2_TRANSFER_IMM(argw))); + FAIL_IF(push_inst(compiler, EMIT_DATA_TRANSFER(inp_flags, 0, inp_flags & WRITE_BACK, reg, arg & REG_MASK, TYPE2_TRANSFER_IMM(argw)))); return -1; } } @@ -1477,7 +1450,7 @@ static sljit_si can_cache(sljit_si arg, sljit_sw argw, sljit_si next_arg, sljit_ /* This can only happen for stores */ \ /* since ldr reg, [reg, ...]! has no meaning */ \ SLJIT_ASSERT(!(inp_flags & LOAD_DATA)); \ - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, TMP_REG3, SLJIT_UNUSED, RM(reg))); \ + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, TMP_REG3, SLJIT_UNUSED, RM(reg)))); \ reg = TMP_REG3; \ } \ } @@ -1537,9 +1510,8 @@ static sljit_si getput_arg(struct sljit_compiler *compiler, sljit_si inp_flags, SLJIT_ASSERT((argw & 0x3) && !(max_delta & 0xf00)); if (inp_flags & WRITE_BACK) tmp_r = arg & REG_MASK; - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(ADD_DP, 0, tmp_r, arg & REG_MASK, RM(OFFS_REG(arg)) | ((argw & 0x3) << 7))); - EMIT_INSTRUCTION(EMIT_DATA_TRANSFER(inp_flags, 1, 0, reg, tmp_r, TYPE2_TRANSFER_IMM(0))); - return SLJIT_SUCCESS; + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(ADD_DP, 0, tmp_r, arg & REG_MASK, RM(OFFS_REG(arg)) | ((argw & 0x3) << 7)))); + return push_inst(compiler, EMIT_DATA_TRANSFER(inp_flags, 1, 0, reg, tmp_r, TYPE2_TRANSFER_IMM(0))); } imm = (sljit_uw)(argw - compiler->cache_argw); @@ -1558,7 +1530,7 @@ static sljit_si getput_arg(struct sljit_compiler *compiler, sljit_si inp_flags, imm = get_imm(argw & ~max_delta); if (imm) { TEST_WRITE_BACK(); - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(ADD_DP, 0, tmp_r, arg & REG_MASK, imm)); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(ADD_DP, 0, tmp_r, arg & REG_MASK, imm))); GETPUT_ARG_DATA_TRANSFER(1, inp_flags & WRITE_BACK, reg, tmp_r, argw & max_delta); return SLJIT_SUCCESS; } @@ -1567,15 +1539,14 @@ static sljit_si getput_arg(struct sljit_compiler *compiler, sljit_si inp_flags, if (imm) { argw = -argw; TEST_WRITE_BACK(); - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(SUB_DP, 0, tmp_r, arg & REG_MASK, imm)); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(SUB_DP, 0, tmp_r, arg & REG_MASK, imm))); GETPUT_ARG_DATA_TRANSFER(0, inp_flags & WRITE_BACK, reg, tmp_r, argw & max_delta); return SLJIT_SUCCESS; } if ((compiler->cache_arg & SLJIT_IMM) && compiler->cache_argw == argw) { TEST_WRITE_BACK(); - EMIT_INSTRUCTION(EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, RM(TMP_REG3) | (max_delta & 0xf00 ? SRC2_IMM : 0))); - return SLJIT_SUCCESS; + return push_inst(compiler, EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, RM(TMP_REG3) | (max_delta & 0xf00 ? SRC2_IMM : 0))); } if (argw == next_argw && (next_arg & SLJIT_MEM)) { @@ -1586,15 +1557,14 @@ static sljit_si getput_arg(struct sljit_compiler *compiler, sljit_si inp_flags, compiler->cache_argw = argw; TEST_WRITE_BACK(); - EMIT_INSTRUCTION(EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, RM(TMP_REG3) | (max_delta & 0xf00 ? SRC2_IMM : 0))); - return SLJIT_SUCCESS; + return push_inst(compiler, EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, RM(TMP_REG3) | (max_delta & 0xf00 ? SRC2_IMM : 0))); } imm = (sljit_uw)(argw - next_argw); if (arg == next_arg && !(inp_flags & WRITE_BACK) && (imm <= (sljit_uw)max_delta || imm >= (sljit_uw)-max_delta)) { SLJIT_ASSERT(inp_flags & LOAD_DATA); FAIL_IF(load_immediate(compiler, TMP_REG3, argw)); - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(ADD_DP, 0, TMP_REG3, TMP_REG3, reg_map[arg & REG_MASK])); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(ADD_DP, 0, TMP_REG3, TMP_REG3, reg_map[arg & REG_MASK]))); compiler->cache_arg = arg; compiler->cache_argw = argw; @@ -1610,8 +1580,7 @@ static sljit_si getput_arg(struct sljit_compiler *compiler, sljit_si inp_flags, } FAIL_IF(load_immediate(compiler, tmp_r, argw)); - EMIT_INSTRUCTION(EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, reg_map[tmp_r] | (max_delta & 0xf00 ? SRC2_IMM : 0))); - return SLJIT_SUCCESS; + return push_inst(compiler, EMIT_DATA_TRANSFER(inp_flags, 1, inp_flags & WRITE_BACK, reg, arg & REG_MASK, reg_map[tmp_r] | (max_delta & 0xf00 ? SRC2_IMM : 0))); } static SLJIT_INLINE sljit_si emit_op_mem(struct sljit_compiler *compiler, sljit_si flags, sljit_si reg, sljit_si arg, sljit_sw argw) @@ -1838,39 +1807,39 @@ extern int __aeabi_idivmod(int numerator, int denominator); SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_op0(struct sljit_compiler *compiler, sljit_si op) { CHECK_ERROR(); - check_sljit_emit_op0(compiler, op); + CHECK(check_sljit_emit_op0(compiler, op)); op = GET_OPCODE(op); switch (op) { case SLJIT_BREAKPOINT: - EMIT_INSTRUCTION(BKPT); + FAIL_IF(push_inst(compiler, BKPT)); break; case SLJIT_NOP: - EMIT_INSTRUCTION(NOP); + FAIL_IF(push_inst(compiler, NOP)); break; - case SLJIT_UMUL: - case SLJIT_SMUL: + case SLJIT_LUMUL: + case SLJIT_LSMUL: #if (defined SLJIT_CONFIG_ARM_V7 && SLJIT_CONFIG_ARM_V7) - return push_inst(compiler, (op == SLJIT_UMUL ? UMULL : SMULL) - | (reg_map[SLJIT_SCRATCH_REG2] << 16) - | (reg_map[SLJIT_SCRATCH_REG1] << 12) - | (reg_map[SLJIT_SCRATCH_REG1] << 8) - | reg_map[SLJIT_SCRATCH_REG2]); + return push_inst(compiler, (op == SLJIT_LUMUL ? UMULL : SMULL) + | (reg_map[SLJIT_R1] << 16) + | (reg_map[SLJIT_R0] << 12) + | (reg_map[SLJIT_R0] << 8) + | reg_map[SLJIT_R1]); #else - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, TMP_REG1, SLJIT_UNUSED, RM(SLJIT_SCRATCH_REG2))); - return push_inst(compiler, (op == SLJIT_UMUL ? UMULL : SMULL) - | (reg_map[SLJIT_SCRATCH_REG2] << 16) - | (reg_map[SLJIT_SCRATCH_REG1] << 12) - | (reg_map[SLJIT_SCRATCH_REG1] << 8) + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, TMP_REG1, SLJIT_UNUSED, RM(SLJIT_R1)))); + return push_inst(compiler, (op == SLJIT_LUMUL ? UMULL : SMULL) + | (reg_map[SLJIT_R1] << 16) + | (reg_map[SLJIT_R0] << 12) + | (reg_map[SLJIT_R0] << 8) | reg_map[TMP_REG1]); #endif - case SLJIT_UDIV: - case SLJIT_SDIV: + case SLJIT_LUDIV: + case SLJIT_LSDIV: if (compiler->scratches >= 3) - EMIT_INSTRUCTION(0xe52d2008 /* str r2, [sp, #-8]! */); + FAIL_IF(push_inst(compiler, 0xe52d2008 /* str r2, [sp, #-8]! */)); #if defined(__GNUC__) FAIL_IF(sljit_emit_ijump(compiler, SLJIT_FAST_CALL, SLJIT_IMM, - (op == SLJIT_UDIV ? SLJIT_FUNC_OFFSET(__aeabi_uidivmod) : SLJIT_FUNC_OFFSET(__aeabi_idivmod)))); + (op == SLJIT_LUDIV ? SLJIT_FUNC_OFFSET(__aeabi_uidivmod) : SLJIT_FUNC_OFFSET(__aeabi_idivmod)))); #else #error "Software divmod functions are needed" #endif @@ -1887,7 +1856,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_op1(struct sljit_compiler *compiler sljit_si src, sljit_sw srcw) { CHECK_ERROR(); - check_sljit_emit_op1(compiler, op, dst, dstw, src, srcw); + CHECK(check_sljit_emit_op1(compiler, op, dst, dstw, src, srcw)); ADJUST_LOCAL_OFFSET(dst, dstw); ADJUST_LOCAL_OFFSET(src, srcw); @@ -1932,7 +1901,8 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_op1(struct sljit_compiler *compiler return emit_op(compiler, op, ALLOW_ANY_IMM, dst, dstw, TMP_REG1, 0, src, srcw); case SLJIT_NEG: -#if (defined SLJIT_VERBOSE && SLJIT_VERBOSE) || (defined SLJIT_DEBUG && SLJIT_DEBUG) +#if (defined SLJIT_VERBOSE && SLJIT_VERBOSE) \ + || (defined SLJIT_ARGUMENT_CHECKS && SLJIT_ARGUMENT_CHECKS) compiler->skip_checks = 1; #endif return sljit_emit_op2(compiler, SLJIT_SUB | GET_ALL_FLAGS(op), dst, dstw, SLJIT_IMM, 0, src, srcw); @@ -1950,7 +1920,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_op2(struct sljit_compiler *compiler sljit_si src2, sljit_sw src2w) { CHECK_ERROR(); - check_sljit_emit_op2(compiler, op, dst, dstw, src1, src1w, src2, src2w); + CHECK(check_sljit_emit_op2(compiler, op, dst, dstw, src1, src1w, src2, src2w)); ADJUST_LOCAL_OFFSET(dst, dstw); ADJUST_LOCAL_OFFSET(src1, src1w); ADJUST_LOCAL_OFFSET(src2, src2w); @@ -1988,22 +1958,21 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_op2(struct sljit_compiler *compiler SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_get_register_index(sljit_si reg) { - check_sljit_get_register_index(reg); + CHECK_REG_INDEX(check_sljit_get_register_index(reg)); return reg_map[reg]; } SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_get_float_register_index(sljit_si reg) { - check_sljit_get_float_register_index(reg); - return reg; + CHECK_REG_INDEX(check_sljit_get_float_register_index(reg)); + return reg << 1; } SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_op_custom(struct sljit_compiler *compiler, void *instruction, sljit_si size) { CHECK_ERROR(); - check_sljit_emit_op_custom(compiler, instruction, size); - SLJIT_ASSERT(size == 4); + CHECK(check_sljit_emit_op_custom(compiler, instruction, size)); return push_inst(compiler, *(sljit_uw*)instruction); } @@ -2064,7 +2033,7 @@ static sljit_si emit_fop_mem(struct sljit_compiler *compiler, sljit_si flags, sl SLJIT_ASSERT(arg & SLJIT_MEM); if (SLJIT_UNLIKELY(arg & OFFS_REG_MASK)) { - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(ADD_DP, 0, TMP_REG1, arg & REG_MASK, RM(OFFS_REG(arg)) | ((argw & 0x3) << 7))); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(ADD_DP, 0, TMP_REG1, arg & REG_MASK, RM(OFFS_REG(arg)) | ((argw & 0x3) << 7)))); arg = SLJIT_MEM | TMP_REG1; argw = 0; } @@ -2097,13 +2066,13 @@ static sljit_si emit_fop_mem(struct sljit_compiler *compiler, sljit_si flags, sl } imm = get_imm(argw & ~0x3fc); if (imm) { - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(ADD_DP, 0, TMP_REG1, arg & REG_MASK, imm)); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(ADD_DP, 0, TMP_REG1, arg & REG_MASK, imm))); return push_inst(compiler, EMIT_FPU_DATA_TRANSFER(inst, 1, TMP_REG1, reg, (argw & 0x3fc) >> 2)); } imm = get_imm(-argw & ~0x3fc); if (imm) { argw = -argw; - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(SUB_DP, 0, TMP_REG1, arg & REG_MASK, imm)); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(SUB_DP, 0, TMP_REG1, arg & REG_MASK, imm))); return push_inst(compiler, EMIT_FPU_DATA_TRANSFER(inst, 0, TMP_REG1, reg, (argw & 0x3fc) >> 2)); } } @@ -2112,7 +2081,7 @@ static sljit_si emit_fop_mem(struct sljit_compiler *compiler, sljit_si flags, sl compiler->cache_argw = argw; if (arg & REG_MASK) { FAIL_IF(load_immediate(compiler, TMP_REG1, argw)); - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(ADD_DP, 0, TMP_REG3, arg & REG_MASK, reg_map[TMP_REG1])); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(ADD_DP, 0, TMP_REG3, arg & REG_MASK, reg_map[TMP_REG1]))); } else FAIL_IF(load_immediate(compiler, TMP_REG3, argw)); @@ -2120,60 +2089,114 @@ static sljit_si emit_fop_mem(struct sljit_compiler *compiler, sljit_si flags, sl return push_inst(compiler, EMIT_FPU_DATA_TRANSFER(inst, 1, TMP_REG3, reg, 0)); } +static SLJIT_INLINE sljit_si sljit_emit_fop1_convw_fromd(struct sljit_compiler *compiler, sljit_si op, + sljit_si dst, sljit_sw dstw, + sljit_si src, sljit_sw srcw) +{ + if (src & SLJIT_MEM) { + FAIL_IF(emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP) | FPU_LOAD, TMP_FREG1, src, srcw)); + src = TMP_FREG1; + } + + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VCVT_S32_F32, op & SLJIT_SINGLE_OP, TMP_FREG1, src, 0))); + + if (dst == SLJIT_UNUSED) + return SLJIT_SUCCESS; + + if (FAST_IS_REG(dst)) + return push_inst(compiler, VMOV | (1 << 20) | RD(dst) | (TMP_FREG1 << 16)); + + /* Store the integer value from a VFP register. */ + return emit_fop_mem(compiler, 0, TMP_FREG1, dst, dstw); +} + +static SLJIT_INLINE sljit_si sljit_emit_fop1_convd_fromw(struct sljit_compiler *compiler, sljit_si op, + sljit_si dst, sljit_sw dstw, + sljit_si src, sljit_sw srcw) +{ + sljit_si dst_r = FAST_IS_REG(dst) ? dst : TMP_FREG1; + + if (FAST_IS_REG(src)) + FAIL_IF(push_inst(compiler, VMOV | RD(src) | (TMP_FREG1 << 16))); + else if (src & SLJIT_MEM) { + /* Load the integer value into a VFP register. */ + FAIL_IF(emit_fop_mem(compiler, FPU_LOAD, TMP_FREG1, src, srcw)); + } + else { + FAIL_IF(load_immediate(compiler, TMP_REG1, srcw)); + FAIL_IF(push_inst(compiler, VMOV | RD(TMP_REG1) | (TMP_FREG1 << 16))); + } + + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VCVT_F32_S32, op & SLJIT_SINGLE_OP, dst_r, TMP_FREG1, 0))); + + if (dst & SLJIT_MEM) + return emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP), TMP_FREG1, dst, dstw); + return SLJIT_SUCCESS; +} + +static SLJIT_INLINE sljit_si sljit_emit_fop1_cmp(struct sljit_compiler *compiler, sljit_si op, + sljit_si src1, sljit_sw src1w, + sljit_si src2, sljit_sw src2w) +{ + if (src1 & SLJIT_MEM) { + FAIL_IF(emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP) | FPU_LOAD, TMP_FREG1, src1, src1w)); + src1 = TMP_FREG1; + } + + if (src2 & SLJIT_MEM) { + FAIL_IF(emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP) | FPU_LOAD, TMP_FREG2, src2, src2w)); + src2 = TMP_FREG2; + } + + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VCMP_F32, op & SLJIT_SINGLE_OP, src1, src2, 0))); + return push_inst(compiler, VMRS); +} + SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fop1(struct sljit_compiler *compiler, sljit_si op, sljit_si dst, sljit_sw dstw, sljit_si src, sljit_sw srcw) { - sljit_si dst_fr; + sljit_si dst_r; CHECK_ERROR(); - check_sljit_emit_fop1(compiler, op, dst, dstw, src, srcw); - SLJIT_COMPILE_ASSERT((SLJIT_SINGLE_OP == 0x100), float_transfer_bit_error); - compiler->cache_arg = 0; compiler->cache_argw = 0; - op ^= SLJIT_SINGLE_OP; + if (GET_OPCODE(op) != SLJIT_CONVD_FROMS) + op ^= SLJIT_SINGLE_OP; - if (GET_OPCODE(op) == SLJIT_CMPD) { - if (dst & SLJIT_MEM) { - FAIL_IF(emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP) | FPU_LOAD, TMP_FREG1, dst, dstw)); - dst = TMP_FREG1; - } - if (src & SLJIT_MEM) { - FAIL_IF(emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP) | FPU_LOAD, TMP_FREG2, src, srcw)); - src = TMP_FREG2; - } - EMIT_INSTRUCTION(EMIT_FPU_OPERATION(VCMP_F32, op & SLJIT_SINGLE_OP, dst, src, 0)); - EMIT_INSTRUCTION(VMRS); - return SLJIT_SUCCESS; - } + SLJIT_COMPILE_ASSERT((SLJIT_SINGLE_OP == 0x100), float_transfer_bit_error); + SELECT_FOP1_OPERATION_WITH_CHECKS(compiler, op, dst, dstw, src, srcw); - dst_fr = FAST_IS_REG(dst) ? dst : TMP_FREG1; + dst_r = FAST_IS_REG(dst) ? dst : TMP_FREG1; if (src & SLJIT_MEM) { - FAIL_IF(emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP) | FPU_LOAD, dst_fr, src, srcw)); - src = dst_fr; + FAIL_IF(emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP) | FPU_LOAD, dst_r, src, srcw)); + src = dst_r; } switch (GET_OPCODE(op)) { - case SLJIT_MOVD: - if (src != dst_fr && dst_fr != TMP_FREG1) - EMIT_INSTRUCTION(EMIT_FPU_OPERATION(VMOV_F32, op & SLJIT_SINGLE_OP, dst_fr, src, 0)); - break; - case SLJIT_NEGD: - EMIT_INSTRUCTION(EMIT_FPU_OPERATION(VNEG_F32, op & SLJIT_SINGLE_OP, dst_fr, src, 0)); - break; - case SLJIT_ABSD: - EMIT_INSTRUCTION(EMIT_FPU_OPERATION(VABS_F32, op & SLJIT_SINGLE_OP, dst_fr, src, 0)); - break; - } - - if (dst_fr == TMP_FREG1) { - if (GET_OPCODE(op) == SLJIT_MOVD) - dst_fr = src; - FAIL_IF(emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP), dst_fr, dst, dstw)); + case SLJIT_DMOV: + if (src != dst_r) { + if (dst_r != TMP_FREG1) + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VMOV_F32, op & SLJIT_SINGLE_OP, dst_r, src, 0))); + else + dst_r = src; + } + break; + case SLJIT_DNEG: + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VNEG_F32, op & SLJIT_SINGLE_OP, dst_r, src, 0))); + break; + case SLJIT_DABS: + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VABS_F32, op & SLJIT_SINGLE_OP, dst_r, src, 0))); + break; + case SLJIT_CONVD_FROMS: + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VCVT_F64_F32, op & SLJIT_SINGLE_OP, dst_r, src, 0))); + op ^= SLJIT_SINGLE_OP; + break; } + if (dst & SLJIT_MEM) + return emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP), dst_r, dst, dstw); return SLJIT_SUCCESS; } @@ -2182,16 +2205,19 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fop2(struct sljit_compiler *compile sljit_si src1, sljit_sw src1w, sljit_si src2, sljit_sw src2w) { - sljit_si dst_fr; + sljit_si dst_r; CHECK_ERROR(); - check_sljit_emit_fop2(compiler, op, dst, dstw, src1, src1w, src2, src2w); + CHECK(check_sljit_emit_fop2(compiler, op, dst, dstw, src1, src1w, src2, src2w)); + ADJUST_LOCAL_OFFSET(dst, dstw); + ADJUST_LOCAL_OFFSET(src1, src1w); + ADJUST_LOCAL_OFFSET(src2, src2w); compiler->cache_arg = 0; compiler->cache_argw = 0; op ^= SLJIT_SINGLE_OP; - dst_fr = FAST_IS_REG(dst) ? dst : TMP_FREG1; + dst_r = FAST_IS_REG(dst) ? dst : TMP_FREG1; if (src2 & SLJIT_MEM) { FAIL_IF(emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP) | FPU_LOAD, TMP_FREG2, src2, src2w)); @@ -2204,24 +2230,24 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fop2(struct sljit_compiler *compile } switch (GET_OPCODE(op)) { - case SLJIT_ADDD: - EMIT_INSTRUCTION(EMIT_FPU_OPERATION(VADD_F32, op & SLJIT_SINGLE_OP, dst_fr, src2, src1)); + case SLJIT_DADD: + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VADD_F32, op & SLJIT_SINGLE_OP, dst_r, src2, src1))); break; - case SLJIT_SUBD: - EMIT_INSTRUCTION(EMIT_FPU_OPERATION(VSUB_F32, op & SLJIT_SINGLE_OP, dst_fr, src2, src1)); + case SLJIT_DSUB: + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VSUB_F32, op & SLJIT_SINGLE_OP, dst_r, src2, src1))); break; - case SLJIT_MULD: - EMIT_INSTRUCTION(EMIT_FPU_OPERATION(VMUL_F32, op & SLJIT_SINGLE_OP, dst_fr, src2, src1)); + case SLJIT_DMUL: + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VMUL_F32, op & SLJIT_SINGLE_OP, dst_r, src2, src1))); break; - case SLJIT_DIVD: - EMIT_INSTRUCTION(EMIT_FPU_OPERATION(VDIV_F32, op & SLJIT_SINGLE_OP, dst_fr, src2, src1)); + case SLJIT_DDIV: + FAIL_IF(push_inst(compiler, EMIT_FPU_OPERATION(VDIV_F32, op & SLJIT_SINGLE_OP, dst_r, src2, src1))); break; } - if (dst_fr == TMP_FREG1) + if (dst_r == TMP_FREG1) FAIL_IF(emit_fop_mem(compiler, (op & SLJIT_SINGLE_OP), TMP_FREG1, dst, dstw)); return SLJIT_SUCCESS; @@ -2238,7 +2264,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fop2(struct sljit_compiler *compile SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fast_enter(struct sljit_compiler *compiler, sljit_si dst, sljit_sw dstw) { CHECK_ERROR(); - check_sljit_emit_fast_enter(compiler, dst, dstw); + CHECK(check_sljit_emit_fast_enter(compiler, dst, dstw)); ADJUST_LOCAL_OFFSET(dst, dstw); /* For UNUSED dst. Uncommon, but possible. */ @@ -2252,7 +2278,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fast_enter(struct sljit_compiler *c if (getput_arg_fast(compiler, WORD_DATA, TMP_REG3, dst, dstw)) return compiler->error; /* TMP_REG3 is used for caching. */ - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, TMP_REG2, SLJIT_UNUSED, RM(TMP_REG3))); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, TMP_REG2, SLJIT_UNUSED, RM(TMP_REG3)))); compiler->cache_arg = 0; compiler->cache_argw = 0; return getput_arg(compiler, WORD_DATA, TMP_REG2, dst, dstw, 0, 0); @@ -2261,11 +2287,11 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fast_enter(struct sljit_compiler *c SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fast_return(struct sljit_compiler *compiler, sljit_si src, sljit_sw srcw) { CHECK_ERROR(); - check_sljit_emit_fast_return(compiler, src, srcw); + CHECK(check_sljit_emit_fast_return(compiler, src, srcw)); ADJUST_LOCAL_OFFSET(src, srcw); if (FAST_IS_REG(src)) - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, TMP_REG3, SLJIT_UNUSED, RM(src))); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, TMP_REG3, SLJIT_UNUSED, RM(src)))); else if (src & SLJIT_MEM) { if (getput_arg_fast(compiler, WORD_DATA | LOAD_DATA, TMP_REG3, src, srcw)) FAIL_IF(compiler->error); @@ -2273,7 +2299,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fast_return(struct sljit_compiler * compiler->cache_arg = 0; compiler->cache_argw = 0; FAIL_IF(getput_arg(compiler, WORD_DATA | LOAD_DATA, TMP_REG2, src, srcw, 0, 0)); - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, TMP_REG3, SLJIT_UNUSED, RM(TMP_REG2))); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, TMP_REG3, SLJIT_UNUSED, RM(TMP_REG2)))); } } else if (src & SLJIT_IMM) @@ -2288,53 +2314,54 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_fast_return(struct sljit_compiler * static sljit_uw get_cc(sljit_si type) { switch (type) { - case SLJIT_C_EQUAL: - case SLJIT_C_MUL_NOT_OVERFLOW: - case SLJIT_C_FLOAT_EQUAL: + case SLJIT_EQUAL: + case SLJIT_MUL_NOT_OVERFLOW: + case SLJIT_D_EQUAL: return 0x00000000; - case SLJIT_C_NOT_EQUAL: - case SLJIT_C_MUL_OVERFLOW: - case SLJIT_C_FLOAT_NOT_EQUAL: + case SLJIT_NOT_EQUAL: + case SLJIT_MUL_OVERFLOW: + case SLJIT_D_NOT_EQUAL: return 0x10000000; - case SLJIT_C_LESS: - case SLJIT_C_FLOAT_LESS: + case SLJIT_LESS: + case SLJIT_D_LESS: return 0x30000000; - case SLJIT_C_GREATER_EQUAL: - case SLJIT_C_FLOAT_GREATER_EQUAL: + case SLJIT_GREATER_EQUAL: + case SLJIT_D_GREATER_EQUAL: return 0x20000000; - case SLJIT_C_GREATER: - case SLJIT_C_FLOAT_GREATER: + case SLJIT_GREATER: + case SLJIT_D_GREATER: return 0x80000000; - case SLJIT_C_LESS_EQUAL: - case SLJIT_C_FLOAT_LESS_EQUAL: + case SLJIT_LESS_EQUAL: + case SLJIT_D_LESS_EQUAL: return 0x90000000; - case SLJIT_C_SIG_LESS: + case SLJIT_SIG_LESS: return 0xb0000000; - case SLJIT_C_SIG_GREATER_EQUAL: + case SLJIT_SIG_GREATER_EQUAL: return 0xa0000000; - case SLJIT_C_SIG_GREATER: + case SLJIT_SIG_GREATER: return 0xc0000000; - case SLJIT_C_SIG_LESS_EQUAL: + case SLJIT_SIG_LESS_EQUAL: return 0xd0000000; - case SLJIT_C_OVERFLOW: - case SLJIT_C_FLOAT_UNORDERED: + case SLJIT_OVERFLOW: + case SLJIT_D_UNORDERED: return 0x60000000; - case SLJIT_C_NOT_OVERFLOW: - case SLJIT_C_FLOAT_ORDERED: + case SLJIT_NOT_OVERFLOW: + case SLJIT_D_ORDERED: return 0x70000000; - default: /* SLJIT_JUMP */ + default: + SLJIT_ASSERT(type >= SLJIT_JUMP && type <= SLJIT_CALL3); return 0xe0000000; } } @@ -2344,7 +2371,7 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_label* sljit_emit_label(struct sljit_compi struct sljit_label *label; CHECK_ERROR_PTR(); - check_sljit_emit_label(compiler); + CHECK_PTR(check_sljit_emit_label(compiler)); if (compiler->last_label && compiler->last_label->size == compiler->size) return compiler->last_label; @@ -2360,7 +2387,7 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_jump* sljit_emit_jump(struct sljit_compile struct sljit_jump *jump; CHECK_ERROR_PTR(); - check_sljit_emit_jump(compiler, type); + CHECK_PTR(check_sljit_emit_jump(compiler, type)); jump = (struct sljit_jump*)ensure_abuf(compiler, sizeof(struct sljit_jump)); PTR_FAIL_IF(!jump); @@ -2401,7 +2428,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_ijump(struct sljit_compiler *compil struct sljit_jump *jump; CHECK_ERROR(); - check_sljit_emit_ijump(compiler, type, src, srcw); + CHECK(check_sljit_emit_ijump(compiler, type, src, srcw)); ADJUST_LOCAL_OFFSET(src, srcw); /* In ARM, we don't need to touch the arguments. */ @@ -2442,7 +2469,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_op_flags(struct sljit_compiler *com sljit_uw cc, ins; CHECK_ERROR(); - check_sljit_emit_op_flags(compiler, op, dst, dstw, src, srcw, type); + CHECK(check_sljit_emit_op_flags(compiler, op, dst, dstw, src, srcw, type)); ADJUST_LOCAL_OFFSET(dst, dstw); ADJUST_LOCAL_OFFSET(src, srcw); @@ -2450,18 +2477,18 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_op_flags(struct sljit_compiler *com return SLJIT_SUCCESS; op = GET_OPCODE(op); - cc = get_cc(type); + cc = get_cc(type & 0xff); dst_r = FAST_IS_REG(dst) ? dst : TMP_REG2; if (op < SLJIT_ADD) { - EMIT_INSTRUCTION(EMIT_DATA_PROCESS_INS(MOV_DP, 0, dst_r, SLJIT_UNUSED, SRC2_IMM | 0)); - EMIT_INSTRUCTION((EMIT_DATA_PROCESS_INS(MOV_DP, 0, dst_r, SLJIT_UNUSED, SRC2_IMM | 1) & ~COND_MASK) | cc); + FAIL_IF(push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, 0, dst_r, SLJIT_UNUSED, SRC2_IMM | 0))); + FAIL_IF(push_inst(compiler, (EMIT_DATA_PROCESS_INS(MOV_DP, 0, dst_r, SLJIT_UNUSED, SRC2_IMM | 1) & ~COND_MASK) | cc)); return (dst_r == TMP_REG2) ? emit_op_mem(compiler, WORD_DATA, TMP_REG2, dst, dstw) : SLJIT_SUCCESS; } ins = (op == SLJIT_AND ? AND_DP : (op == SLJIT_OR ? ORR_DP : EOR_DP)); if ((op == SLJIT_OR || op == SLJIT_XOR) && FAST_IS_REG(dst) && dst == src) { - EMIT_INSTRUCTION((EMIT_DATA_PROCESS_INS(ins, 0, dst, dst, SRC2_IMM | 1) & ~COND_MASK) | cc); + FAIL_IF(push_inst(compiler, (EMIT_DATA_PROCESS_INS(ins, 0, dst, dst, SRC2_IMM | 1) & ~COND_MASK) | cc)); /* The condition must always be set, even if the ORR/EOR is not executed above. */ return (flags & SLJIT_SET_E) ? push_inst(compiler, EMIT_DATA_PROCESS_INS(MOV_DP, SET_FLAGS, TMP_REG1, SLJIT_UNUSED, RM(dst))) : SLJIT_SUCCESS; } @@ -2478,8 +2505,8 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_si sljit_emit_op_flags(struct sljit_compiler *com srcw = 0; } - EMIT_INSTRUCTION((EMIT_DATA_PROCESS_INS(ins, 0, dst_r, src, SRC2_IMM | 1) & ~COND_MASK) | cc); - EMIT_INSTRUCTION((EMIT_DATA_PROCESS_INS(ins, 0, dst_r, src, SRC2_IMM | 0) & ~COND_MASK) | (cc ^ 0x10000000)); + FAIL_IF(push_inst(compiler, (EMIT_DATA_PROCESS_INS(ins, 0, dst_r, src, SRC2_IMM | 1) & ~COND_MASK) | cc)); + FAIL_IF(push_inst(compiler, (EMIT_DATA_PROCESS_INS(ins, 0, dst_r, src, SRC2_IMM | 0) & ~COND_MASK) | (cc ^ 0x10000000))); if (dst_r == TMP_REG2) FAIL_IF(emit_op_mem2(compiler, WORD_DATA, TMP_REG2, dst, dstw, 0, 0)); @@ -2492,7 +2519,7 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_const* sljit_emit_const(struct sljit_compi sljit_si reg; CHECK_ERROR_PTR(); - check_sljit_emit_const(compiler, dst, dstw, init_value); + CHECK_PTR(check_sljit_emit_const(compiler, dst, dstw, init_value)); ADJUST_LOCAL_OFFSET(dst, dstw); const_ = (struct sljit_const*)ensure_abuf(compiler, sizeof(struct sljit_const)); |