From 612599587bef71e1deca9a77a8e2e16e69b045de Mon Sep 17 00:00:00 2001 From: Damien George Date: Mon, 28 Nov 2016 09:24:50 +1100 Subject: py: Factor out common code from assemblers into asmbase.[ch]. All assemblers should "derive" from mp_asm_base_t. --- py/asmx86.h | 24 +++++++++--------------- 1 file changed, 9 insertions(+), 15 deletions(-) (limited to 'py/asmx86.h') diff --git a/py/asmx86.h b/py/asmx86.h index e0c57722a..85352fff3 100644 --- a/py/asmx86.h +++ b/py/asmx86.h @@ -28,6 +28,7 @@ #include "py/mpconfig.h" #include "py/misc.h" +#include "py/asmbase.h" // x86 cdecl calling convention is: // - args passed on the stack in reverse order @@ -42,9 +43,6 @@ // NOTE: this is a change from the old convention used in this file and // some functions still use the old (reverse) convention. -#define ASM_X86_PASS_COMPUTE (1) -#define ASM_X86_PASS_EMIT (2) - #define ASM_X86_REG_EAX (0) #define ASM_X86_REG_ECX (1) #define ASM_X86_REG_EDX (2) @@ -75,18 +73,15 @@ #define ASM_X86_CC_JLE (0xe) // less or equal, signed #define ASM_X86_CC_JG (0xf) // greater, signed -typedef struct _asm_x86_t asm_x86_t; - -asm_x86_t* asm_x86_new(mp_uint_t max_num_labels); -void asm_x86_free(asm_x86_t* as, bool free_code); -void asm_x86_start_pass(asm_x86_t *as, mp_uint_t pass); -void asm_x86_end_pass(asm_x86_t *as); -mp_uint_t asm_x86_get_code_pos(asm_x86_t *as); -mp_uint_t asm_x86_get_code_size(asm_x86_t* as); -void* asm_x86_get_code(asm_x86_t* as); +typedef struct _asm_x86_t { + mp_asm_base_t base; + byte dummy_data[4]; // in addition to dummy_data in base + int num_locals; +} asm_x86_t; -void asm_x86_align(asm_x86_t *as, mp_uint_t align); -void asm_x86_data(asm_x86_t *as, mp_uint_t bytesize, mp_uint_t val); +static inline void asm_x86_end_pass(asm_x86_t *as) { + (void)as; +} void asm_x86_mov_r32_r32(asm_x86_t* as, int dest_r32, int src_r32); void asm_x86_mov_i32_to_r32(asm_x86_t *as, int32_t src_i32, int dest_r32); @@ -108,7 +103,6 @@ void asm_x86_mul_r32_r32(asm_x86_t* as, int dest_r32, int src_r32); void asm_x86_cmp_r32_with_r32(asm_x86_t* as, int src_r32_a, int src_r32_b); void asm_x86_test_r8_with_r8(asm_x86_t* as, int src_r32_a, int src_r32_b); void asm_x86_setcc_r8(asm_x86_t* as, mp_uint_t jcc_type, int dest_r8); -void asm_x86_label_assign(asm_x86_t* as, mp_uint_t label); void asm_x86_jmp_label(asm_x86_t* as, mp_uint_t label); void asm_x86_jcc_label(asm_x86_t* as, mp_uint_t jcc_type, mp_uint_t label); void asm_x86_entry(asm_x86_t* as, mp_uint_t num_locals); -- cgit v1.2.3 From c8746e1e7243f27f2824a67c2ac9921839784857 Mon Sep 17 00:00:00 2001 From: Damien George Date: Thu, 8 Dec 2016 17:47:17 +1100 Subject: py: Move arch-specific assembler macros from emitnative to asmXXX.h. --- py/asmarm.h | 80 +++++++++++++++ py/asmthumb.h | 81 +++++++++++++++ py/asmx64.h | 83 +++++++++++++++ py/asmx86.h | 83 +++++++++++++++ py/emitnative.c | 311 +------------------------------------------------------- 5 files changed, 331 insertions(+), 307 deletions(-) (limited to 'py/asmx86.h') diff --git a/py/asmarm.h b/py/asmarm.h index 263721057..e273b98d7 100644 --- a/py/asmarm.h +++ b/py/asmarm.h @@ -122,4 +122,84 @@ void asm_arm_bcc_label(asm_arm_t *as, int cond, uint label); void asm_arm_b_label(asm_arm_t *as, uint label); void asm_arm_bl_ind(asm_arm_t *as, void *fun_ptr, uint fun_id, uint reg_temp); +#if GENERIC_ASM_API + +// The following macros provide a (mostly) arch-independent API to +// generate native code, and are used by the native emitter. + +#define ASM_WORD_SIZE (4) + +#define REG_RET ASM_ARM_REG_R0 +#define REG_ARG_1 ASM_ARM_REG_R0 +#define REG_ARG_2 ASM_ARM_REG_R1 +#define REG_ARG_3 ASM_ARM_REG_R2 +#define REG_ARG_4 ASM_ARM_REG_R3 + +#define REG_TEMP0 ASM_ARM_REG_R0 +#define REG_TEMP1 ASM_ARM_REG_R1 +#define REG_TEMP2 ASM_ARM_REG_R2 + +#define REG_LOCAL_1 ASM_ARM_REG_R4 +#define REG_LOCAL_2 ASM_ARM_REG_R5 +#define REG_LOCAL_3 ASM_ARM_REG_R6 +#define REG_LOCAL_NUM (3) + +#define ASM_T asm_arm_t +#define ASM_END_PASS asm_arm_end_pass +#define ASM_ENTRY asm_arm_entry +#define ASM_EXIT asm_arm_exit + +#define ASM_JUMP asm_arm_b_label +#define ASM_JUMP_IF_REG_ZERO(as, reg, label) \ + do { \ + asm_arm_cmp_reg_i8(as, reg, 0); \ + asm_arm_bcc_label(as, ASM_ARM_CC_EQ, label); \ + } while (0) +#define ASM_JUMP_IF_REG_NONZERO(as, reg, label) \ + do { \ + asm_arm_cmp_reg_i8(as, reg, 0); \ + asm_arm_bcc_label(as, ASM_ARM_CC_NE, label); \ + } while (0) +#define ASM_JUMP_IF_REG_EQ(as, reg1, reg2, label) \ + do { \ + asm_arm_cmp_reg_reg(as, reg1, reg2); \ + asm_arm_bcc_label(as, ASM_ARM_CC_EQ, label); \ + } while (0) +#define ASM_CALL_IND(as, ptr, idx) asm_arm_bl_ind(as, ptr, idx, ASM_ARM_REG_R3) + +#define ASM_MOV_REG_TO_LOCAL(as, reg, local_num) asm_arm_mov_local_reg(as, (local_num), (reg)) +#define ASM_MOV_IMM_TO_REG(as, imm, reg) asm_arm_mov_reg_i32(as, (reg), (imm)) +#define ASM_MOV_ALIGNED_IMM_TO_REG(as, imm, reg) asm_arm_mov_reg_i32(as, (reg), (imm)) +#define ASM_MOV_IMM_TO_LOCAL_USING(as, imm, local_num, reg_temp) \ + do { \ + asm_arm_mov_reg_i32(as, (reg_temp), (imm)); \ + asm_arm_mov_local_reg(as, (local_num), (reg_temp)); \ + } while (false) +#define ASM_MOV_LOCAL_TO_REG(as, local_num, reg) asm_arm_mov_reg_local(as, (reg), (local_num)) +#define ASM_MOV_REG_REG(as, reg_dest, reg_src) asm_arm_mov_reg_reg((as), (reg_dest), (reg_src)) +#define ASM_MOV_LOCAL_ADDR_TO_REG(as, local_num, reg) asm_arm_mov_reg_local_addr(as, (reg), (local_num)) + +#define ASM_LSL_REG_REG(as, reg_dest, reg_shift) asm_arm_lsl_reg_reg((as), (reg_dest), (reg_shift)) +#define ASM_ASR_REG_REG(as, reg_dest, reg_shift) asm_arm_asr_reg_reg((as), (reg_dest), (reg_shift)) +#define ASM_OR_REG_REG(as, reg_dest, reg_src) asm_arm_orr_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) +#define ASM_XOR_REG_REG(as, reg_dest, reg_src) asm_arm_eor_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) +#define ASM_AND_REG_REG(as, reg_dest, reg_src) asm_arm_and_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) +#define ASM_ADD_REG_REG(as, reg_dest, reg_src) asm_arm_add_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) +#define ASM_SUB_REG_REG(as, reg_dest, reg_src) asm_arm_sub_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) +#define ASM_MUL_REG_REG(as, reg_dest, reg_src) asm_arm_mul_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) + +#define ASM_LOAD_REG_REG(as, reg_dest, reg_base) asm_arm_ldr_reg_reg((as), (reg_dest), (reg_base), 0) +#define ASM_LOAD_REG_REG_OFFSET(as, reg_dest, reg_base, word_offset) asm_arm_ldr_reg_reg((as), (reg_dest), (reg_base), 4 * (word_offset)) +#define ASM_LOAD8_REG_REG(as, reg_dest, reg_base) asm_arm_ldrb_reg_reg((as), (reg_dest), (reg_base)) +#define ASM_LOAD16_REG_REG(as, reg_dest, reg_base) asm_arm_ldrh_reg_reg((as), (reg_dest), (reg_base)) +#define ASM_LOAD32_REG_REG(as, reg_dest, reg_base) asm_arm_ldr_reg_reg((as), (reg_dest), (reg_base), 0) + +#define ASM_STORE_REG_REG(as, reg_value, reg_base) asm_arm_str_reg_reg((as), (reg_value), (reg_base), 0) +#define ASM_STORE_REG_REG_OFFSET(as, reg_dest, reg_base, word_offset) asm_arm_str_reg_reg((as), (reg_dest), (reg_base), 4 * (word_offset)) +#define ASM_STORE8_REG_REG(as, reg_value, reg_base) asm_arm_strb_reg_reg((as), (reg_value), (reg_base)) +#define ASM_STORE16_REG_REG(as, reg_value, reg_base) asm_arm_strh_reg_reg((as), (reg_value), (reg_base)) +#define ASM_STORE32_REG_REG(as, reg_value, reg_base) asm_arm_str_reg_reg((as), (reg_value), (reg_base), 0) + +#endif // GENERIC_ASM_API + #endif // __MICROPY_INCLUDED_PY_ASMARM_H__ diff --git a/py/asmthumb.h b/py/asmthumb.h index c03b00da3..52e663b3b 100644 --- a/py/asmthumb.h +++ b/py/asmthumb.h @@ -237,4 +237,85 @@ void asm_thumb_b_label(asm_thumb_t *as, uint label); // convenience: picks narro void asm_thumb_bcc_label(asm_thumb_t *as, int cc, uint label); // convenience: picks narrow or wide branch void asm_thumb_bl_ind(asm_thumb_t *as, void *fun_ptr, uint fun_id, uint reg_temp); // convenience +#if GENERIC_ASM_API + +// The following macros provide a (mostly) arch-independent API to +// generate native code, and are used by the native emitter. + +#define ASM_WORD_SIZE (4) + +#define REG_RET ASM_THUMB_REG_R0 +#define REG_ARG_1 ASM_THUMB_REG_R0 +#define REG_ARG_2 ASM_THUMB_REG_R1 +#define REG_ARG_3 ASM_THUMB_REG_R2 +#define REG_ARG_4 ASM_THUMB_REG_R3 +// rest of args go on stack + +#define REG_TEMP0 ASM_THUMB_REG_R0 +#define REG_TEMP1 ASM_THUMB_REG_R1 +#define REG_TEMP2 ASM_THUMB_REG_R2 + +#define REG_LOCAL_1 ASM_THUMB_REG_R4 +#define REG_LOCAL_2 ASM_THUMB_REG_R5 +#define REG_LOCAL_3 ASM_THUMB_REG_R6 +#define REG_LOCAL_NUM (3) + +#define ASM_T asm_thumb_t +#define ASM_END_PASS asm_thumb_end_pass +#define ASM_ENTRY asm_thumb_entry +#define ASM_EXIT asm_thumb_exit + +#define ASM_JUMP asm_thumb_b_label +#define ASM_JUMP_IF_REG_ZERO(as, reg, label) \ + do { \ + asm_thumb_cmp_rlo_i8(as, reg, 0); \ + asm_thumb_bcc_label(as, ASM_THUMB_CC_EQ, label); \ + } while (0) +#define ASM_JUMP_IF_REG_NONZERO(as, reg, label) \ + do { \ + asm_thumb_cmp_rlo_i8(as, reg, 0); \ + asm_thumb_bcc_label(as, ASM_THUMB_CC_NE, label); \ + } while (0) +#define ASM_JUMP_IF_REG_EQ(as, reg1, reg2, label) \ + do { \ + asm_thumb_cmp_rlo_rlo(as, reg1, reg2); \ + asm_thumb_bcc_label(as, ASM_THUMB_CC_EQ, label); \ + } while (0) +#define ASM_CALL_IND(as, ptr, idx) asm_thumb_bl_ind(as, ptr, idx, ASM_THUMB_REG_R3) + +#define ASM_MOV_REG_TO_LOCAL(as, reg, local_num) asm_thumb_mov_local_reg(as, (local_num), (reg)) +#define ASM_MOV_IMM_TO_REG(as, imm, reg) asm_thumb_mov_reg_i32_optimised(as, (reg), (imm)) +#define ASM_MOV_ALIGNED_IMM_TO_REG(as, imm, reg) asm_thumb_mov_reg_i32_aligned(as, (reg), (imm)) +#define ASM_MOV_IMM_TO_LOCAL_USING(as, imm, local_num, reg_temp) \ + do { \ + asm_thumb_mov_reg_i32_optimised(as, (reg_temp), (imm)); \ + asm_thumb_mov_local_reg(as, (local_num), (reg_temp)); \ + } while (false) +#define ASM_MOV_LOCAL_TO_REG(as, local_num, reg) asm_thumb_mov_reg_local(as, (reg), (local_num)) +#define ASM_MOV_REG_REG(as, reg_dest, reg_src) asm_thumb_mov_reg_reg((as), (reg_dest), (reg_src)) +#define ASM_MOV_LOCAL_ADDR_TO_REG(as, local_num, reg) asm_thumb_mov_reg_local_addr(as, (reg), (local_num)) + +#define ASM_LSL_REG_REG(as, reg_dest, reg_shift) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_LSL, (reg_dest), (reg_shift)) +#define ASM_ASR_REG_REG(as, reg_dest, reg_shift) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_ASR, (reg_dest), (reg_shift)) +#define ASM_OR_REG_REG(as, reg_dest, reg_src) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_ORR, (reg_dest), (reg_src)) +#define ASM_XOR_REG_REG(as, reg_dest, reg_src) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_EOR, (reg_dest), (reg_src)) +#define ASM_AND_REG_REG(as, reg_dest, reg_src) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_AND, (reg_dest), (reg_src)) +#define ASM_ADD_REG_REG(as, reg_dest, reg_src) asm_thumb_add_rlo_rlo_rlo((as), (reg_dest), (reg_dest), (reg_src)) +#define ASM_SUB_REG_REG(as, reg_dest, reg_src) asm_thumb_sub_rlo_rlo_rlo((as), (reg_dest), (reg_dest), (reg_src)) +#define ASM_MUL_REG_REG(as, reg_dest, reg_src) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_MUL, (reg_dest), (reg_src)) + +#define ASM_LOAD_REG_REG(as, reg_dest, reg_base) asm_thumb_ldr_rlo_rlo_i5((as), (reg_dest), (reg_base), 0) +#define ASM_LOAD_REG_REG_OFFSET(as, reg_dest, reg_base, word_offset) asm_thumb_ldr_rlo_rlo_i5((as), (reg_dest), (reg_base), (word_offset)) +#define ASM_LOAD8_REG_REG(as, reg_dest, reg_base) asm_thumb_ldrb_rlo_rlo_i5((as), (reg_dest), (reg_base), 0) +#define ASM_LOAD16_REG_REG(as, reg_dest, reg_base) asm_thumb_ldrh_rlo_rlo_i5((as), (reg_dest), (reg_base), 0) +#define ASM_LOAD32_REG_REG(as, reg_dest, reg_base) asm_thumb_ldr_rlo_rlo_i5((as), (reg_dest), (reg_base), 0) + +#define ASM_STORE_REG_REG(as, reg_src, reg_base) asm_thumb_str_rlo_rlo_i5((as), (reg_src), (reg_base), 0) +#define ASM_STORE_REG_REG_OFFSET(as, reg_src, reg_base, word_offset) asm_thumb_str_rlo_rlo_i5((as), (reg_src), (reg_base), (word_offset)) +#define ASM_STORE8_REG_REG(as, reg_src, reg_base) asm_thumb_strb_rlo_rlo_i5((as), (reg_src), (reg_base), 0) +#define ASM_STORE16_REG_REG(as, reg_src, reg_base) asm_thumb_strh_rlo_rlo_i5((as), (reg_src), (reg_base), 0) +#define ASM_STORE32_REG_REG(as, reg_src, reg_base) asm_thumb_str_rlo_rlo_i5((as), (reg_src), (reg_base), 0) + +#endif // GENERIC_ASM_API + #endif // __MICROPY_INCLUDED_PY_ASMTHUMB_H__ diff --git a/py/asmx64.h b/py/asmx64.h index daf5b3d0b..f80c8da8b 100644 --- a/py/asmx64.h +++ b/py/asmx64.h @@ -115,4 +115,87 @@ void asm_x64_mov_r64_to_local(asm_x64_t* as, int src_r64, int dest_local_num); void asm_x64_mov_local_addr_to_r64(asm_x64_t* as, int local_num, int dest_r64); void asm_x64_call_ind(asm_x64_t* as, void* ptr, int temp_r32); +#if GENERIC_ASM_API + +// The following macros provide a (mostly) arch-independent API to +// generate native code, and are used by the native emitter. + +#define ASM_WORD_SIZE (8) + +#define REG_RET ASM_X64_REG_RAX +#define REG_ARG_1 ASM_X64_REG_RDI +#define REG_ARG_2 ASM_X64_REG_RSI +#define REG_ARG_3 ASM_X64_REG_RDX +#define REG_ARG_4 ASM_X64_REG_RCX +#define REG_ARG_5 ASM_X64_REG_R08 + +// caller-save +#define REG_TEMP0 ASM_X64_REG_RAX +#define REG_TEMP1 ASM_X64_REG_RDI +#define REG_TEMP2 ASM_X64_REG_RSI + +// callee-save +#define REG_LOCAL_1 ASM_X64_REG_RBX +#define REG_LOCAL_2 ASM_X64_REG_R12 +#define REG_LOCAL_3 ASM_X64_REG_R13 +#define REG_LOCAL_NUM (3) + +#define ASM_T asm_x64_t +#define ASM_END_PASS asm_x64_end_pass +#define ASM_ENTRY asm_x64_entry +#define ASM_EXIT asm_x64_exit + +#define ASM_JUMP asm_x64_jmp_label +#define ASM_JUMP_IF_REG_ZERO(as, reg, label) \ + do { \ + asm_x64_test_r8_with_r8(as, reg, reg); \ + asm_x64_jcc_label(as, ASM_X64_CC_JZ, label); \ + } while (0) +#define ASM_JUMP_IF_REG_NONZERO(as, reg, label) \ + do { \ + asm_x64_test_r8_with_r8(as, reg, reg); \ + asm_x64_jcc_label(as, ASM_X64_CC_JNZ, label); \ + } while (0) +#define ASM_JUMP_IF_REG_EQ(as, reg1, reg2, label) \ + do { \ + asm_x64_cmp_r64_with_r64(as, reg1, reg2); \ + asm_x64_jcc_label(as, ASM_X64_CC_JE, label); \ + } while (0) +#define ASM_CALL_IND(as, ptr, idx) asm_x64_call_ind(as, ptr, ASM_X64_REG_RAX) + +#define ASM_MOV_REG_TO_LOCAL asm_x64_mov_r64_to_local +#define ASM_MOV_IMM_TO_REG asm_x64_mov_i64_to_r64_optimised +#define ASM_MOV_ALIGNED_IMM_TO_REG asm_x64_mov_i64_to_r64_aligned +#define ASM_MOV_IMM_TO_LOCAL_USING(as, imm, local_num, reg_temp) \ + do { \ + asm_x64_mov_i64_to_r64_optimised(as, (imm), (reg_temp)); \ + asm_x64_mov_r64_to_local(as, (reg_temp), (local_num)); \ + } while (false) +#define ASM_MOV_LOCAL_TO_REG asm_x64_mov_local_to_r64 +#define ASM_MOV_REG_REG(as, reg_dest, reg_src) asm_x64_mov_r64_r64((as), (reg_dest), (reg_src)) +#define ASM_MOV_LOCAL_ADDR_TO_REG asm_x64_mov_local_addr_to_r64 + +#define ASM_LSL_REG(as, reg) asm_x64_shl_r64_cl((as), (reg)) +#define ASM_ASR_REG(as, reg) asm_x64_sar_r64_cl((as), (reg)) +#define ASM_OR_REG_REG(as, reg_dest, reg_src) asm_x64_or_r64_r64((as), (reg_dest), (reg_src)) +#define ASM_XOR_REG_REG(as, reg_dest, reg_src) asm_x64_xor_r64_r64((as), (reg_dest), (reg_src)) +#define ASM_AND_REG_REG(as, reg_dest, reg_src) asm_x64_and_r64_r64((as), (reg_dest), (reg_src)) +#define ASM_ADD_REG_REG(as, reg_dest, reg_src) asm_x64_add_r64_r64((as), (reg_dest), (reg_src)) +#define ASM_SUB_REG_REG(as, reg_dest, reg_src) asm_x64_sub_r64_r64((as), (reg_dest), (reg_src)) +#define ASM_MUL_REG_REG(as, reg_dest, reg_src) asm_x64_mul_r64_r64((as), (reg_dest), (reg_src)) + +#define ASM_LOAD_REG_REG(as, reg_dest, reg_base) asm_x64_mov_mem64_to_r64((as), (reg_base), 0, (reg_dest)) +#define ASM_LOAD_REG_REG_OFFSET(as, reg_dest, reg_base, word_offset) asm_x64_mov_mem64_to_r64((as), (reg_base), 8 * (word_offset), (reg_dest)) +#define ASM_LOAD8_REG_REG(as, reg_dest, reg_base) asm_x64_mov_mem8_to_r64zx((as), (reg_base), 0, (reg_dest)) +#define ASM_LOAD16_REG_REG(as, reg_dest, reg_base) asm_x64_mov_mem16_to_r64zx((as), (reg_base), 0, (reg_dest)) +#define ASM_LOAD32_REG_REG(as, reg_dest, reg_base) asm_x64_mov_mem32_to_r64zx((as), (reg_base), 0, (reg_dest)) + +#define ASM_STORE_REG_REG(as, reg_src, reg_base) asm_x64_mov_r64_to_mem64((as), (reg_src), (reg_base), 0) +#define ASM_STORE_REG_REG_OFFSET(as, reg_src, reg_base, word_offset) asm_x64_mov_r64_to_mem64((as), (reg_src), (reg_base), 8 * (word_offset)) +#define ASM_STORE8_REG_REG(as, reg_src, reg_base) asm_x64_mov_r8_to_mem8((as), (reg_src), (reg_base), 0) +#define ASM_STORE16_REG_REG(as, reg_src, reg_base) asm_x64_mov_r16_to_mem16((as), (reg_src), (reg_base), 0) +#define ASM_STORE32_REG_REG(as, reg_src, reg_base) asm_x64_mov_r32_to_mem32((as), (reg_src), (reg_base), 0) + +#endif // GENERIC_ASM_API + #endif // __MICROPY_INCLUDED_PY_ASMX64_H__ diff --git a/py/asmx86.h b/py/asmx86.h index 85352fff3..ac4b2ecd0 100644 --- a/py/asmx86.h +++ b/py/asmx86.h @@ -113,4 +113,87 @@ void asm_x86_mov_r32_to_local(asm_x86_t* as, int src_r32, int dest_local_num); void asm_x86_mov_local_addr_to_r32(asm_x86_t* as, int local_num, int dest_r32); void asm_x86_call_ind(asm_x86_t* as, void* ptr, mp_uint_t n_args, int temp_r32); +#if GENERIC_ASM_API + +// The following macros provide a (mostly) arch-independent API to +// generate native code, and are used by the native emitter. + +#define ASM_WORD_SIZE (4) + +#define REG_RET ASM_X86_REG_EAX +#define REG_ARG_1 ASM_X86_REG_ARG_1 +#define REG_ARG_2 ASM_X86_REG_ARG_2 +#define REG_ARG_3 ASM_X86_REG_ARG_3 +#define REG_ARG_4 ASM_X86_REG_ARG_4 +#define REG_ARG_5 ASM_X86_REG_ARG_5 + +// caller-save, so can be used as temporaries +#define REG_TEMP0 ASM_X86_REG_EAX +#define REG_TEMP1 ASM_X86_REG_ECX +#define REG_TEMP2 ASM_X86_REG_EDX + +// callee-save, so can be used as locals +#define REG_LOCAL_1 ASM_X86_REG_EBX +#define REG_LOCAL_2 ASM_X86_REG_ESI +#define REG_LOCAL_3 ASM_X86_REG_EDI +#define REG_LOCAL_NUM (3) + +#define ASM_T asm_x86_t +#define ASM_END_PASS asm_x86_end_pass +#define ASM_ENTRY asm_x86_entry +#define ASM_EXIT asm_x86_exit + +#define ASM_JUMP asm_x86_jmp_label +#define ASM_JUMP_IF_REG_ZERO(as, reg, label) \ + do { \ + asm_x86_test_r8_with_r8(as, reg, reg); \ + asm_x86_jcc_label(as, ASM_X86_CC_JZ, label); \ + } while (0) +#define ASM_JUMP_IF_REG_NONZERO(as, reg, label) \ + do { \ + asm_x86_test_r8_with_r8(as, reg, reg); \ + asm_x86_jcc_label(as, ASM_X86_CC_JNZ, label); \ + } while (0) +#define ASM_JUMP_IF_REG_EQ(as, reg1, reg2, label) \ + do { \ + asm_x86_cmp_r32_with_r32(as, reg1, reg2); \ + asm_x86_jcc_label(as, ASM_X86_CC_JE, label); \ + } while (0) +#define ASM_CALL_IND(as, ptr, idx) asm_x86_call_ind(as, ptr, mp_f_n_args[idx], ASM_X86_REG_EAX) + +#define ASM_MOV_REG_TO_LOCAL asm_x86_mov_r32_to_local +#define ASM_MOV_IMM_TO_REG asm_x86_mov_i32_to_r32 +#define ASM_MOV_ALIGNED_IMM_TO_REG asm_x86_mov_i32_to_r32_aligned +#define ASM_MOV_IMM_TO_LOCAL_USING(as, imm, local_num, reg_temp) \ + do { \ + asm_x86_mov_i32_to_r32(as, (imm), (reg_temp)); \ + asm_x86_mov_r32_to_local(as, (reg_temp), (local_num)); \ + } while (false) +#define ASM_MOV_LOCAL_TO_REG asm_x86_mov_local_to_r32 +#define ASM_MOV_REG_REG(as, reg_dest, reg_src) asm_x86_mov_r32_r32((as), (reg_dest), (reg_src)) +#define ASM_MOV_LOCAL_ADDR_TO_REG asm_x86_mov_local_addr_to_r32 + +#define ASM_LSL_REG(as, reg) asm_x86_shl_r32_cl((as), (reg)) +#define ASM_ASR_REG(as, reg) asm_x86_sar_r32_cl((as), (reg)) +#define ASM_OR_REG_REG(as, reg_dest, reg_src) asm_x86_or_r32_r32((as), (reg_dest), (reg_src)) +#define ASM_XOR_REG_REG(as, reg_dest, reg_src) asm_x86_xor_r32_r32((as), (reg_dest), (reg_src)) +#define ASM_AND_REG_REG(as, reg_dest, reg_src) asm_x86_and_r32_r32((as), (reg_dest), (reg_src)) +#define ASM_ADD_REG_REG(as, reg_dest, reg_src) asm_x86_add_r32_r32((as), (reg_dest), (reg_src)) +#define ASM_SUB_REG_REG(as, reg_dest, reg_src) asm_x86_sub_r32_r32((as), (reg_dest), (reg_src)) +#define ASM_MUL_REG_REG(as, reg_dest, reg_src) asm_x86_mul_r32_r32((as), (reg_dest), (reg_src)) + +#define ASM_LOAD_REG_REG(as, reg_dest, reg_base) asm_x86_mov_mem32_to_r32((as), (reg_base), 0, (reg_dest)) +#define ASM_LOAD_REG_REG_OFFSET(as, reg_dest, reg_base, word_offset) asm_x86_mov_mem32_to_r32((as), (reg_base), 4 * (word_offset), (reg_dest)) +#define ASM_LOAD8_REG_REG(as, reg_dest, reg_base) asm_x86_mov_mem8_to_r32zx((as), (reg_base), 0, (reg_dest)) +#define ASM_LOAD16_REG_REG(as, reg_dest, reg_base) asm_x86_mov_mem16_to_r32zx((as), (reg_base), 0, (reg_dest)) +#define ASM_LOAD32_REG_REG(as, reg_dest, reg_base) asm_x86_mov_mem32_to_r32((as), (reg_base), 0, (reg_dest)) + +#define ASM_STORE_REG_REG(as, reg_src, reg_base) asm_x86_mov_r32_to_mem32((as), (reg_src), (reg_base), 0) +#define ASM_STORE_REG_REG_OFFSET(as, reg_src, reg_base, word_offset) asm_x86_mov_r32_to_mem32((as), (reg_src), (reg_base), 4 * (word_offset)) +#define ASM_STORE8_REG_REG(as, reg_src, reg_base) asm_x86_mov_r8_to_mem8((as), (reg_src), (reg_base), 0) +#define ASM_STORE16_REG_REG(as, reg_src, reg_base) asm_x86_mov_r16_to_mem16((as), (reg_src), (reg_base), 0) +#define ASM_STORE32_REG_REG(as, reg_src, reg_base) asm_x86_mov_r32_to_mem32((as), (reg_src), (reg_base), 0) + +#endif // GENERIC_ASM_API + #endif // __MICROPY_INCLUDED_PY_ASMX86_H__ diff --git a/py/emitnative.c b/py/emitnative.c index d19dfe467..5a7e0ea34 100644 --- a/py/emitnative.c +++ b/py/emitnative.c @@ -63,96 +63,19 @@ || (MICROPY_EMIT_THUMB && N_THUMB) \ || (MICROPY_EMIT_ARM && N_ARM) +// this is defined so that the assembler exports generic assembler API macros +#define GENERIC_ASM_API (1) + #if N_X64 // x64 specific stuff - #include "py/asmx64.h" - #define EXPORT_FUN(name) emit_native_x64_##name -#define ASM_WORD_SIZE (8) - -#define REG_RET ASM_X64_REG_RAX -#define REG_ARG_1 ASM_X64_REG_RDI -#define REG_ARG_2 ASM_X64_REG_RSI -#define REG_ARG_3 ASM_X64_REG_RDX -#define REG_ARG_4 ASM_X64_REG_RCX -#define REG_ARG_5 ASM_X64_REG_R08 - -// caller-save -#define REG_TEMP0 ASM_X64_REG_RAX -#define REG_TEMP1 ASM_X64_REG_RDI -#define REG_TEMP2 ASM_X64_REG_RSI - -// callee-save -#define REG_LOCAL_1 ASM_X64_REG_RBX -#define REG_LOCAL_2 ASM_X64_REG_R12 -#define REG_LOCAL_3 ASM_X64_REG_R13 -#define REG_LOCAL_NUM (3) - -#define ASM_T asm_x64_t -#define ASM_END_PASS asm_x64_end_pass -#define ASM_ENTRY asm_x64_entry -#define ASM_EXIT asm_x64_exit - -#define ASM_JUMP asm_x64_jmp_label -#define ASM_JUMP_IF_REG_ZERO(as, reg, label) \ - do { \ - asm_x64_test_r8_with_r8(as, reg, reg); \ - asm_x64_jcc_label(as, ASM_X64_CC_JZ, label); \ - } while (0) -#define ASM_JUMP_IF_REG_NONZERO(as, reg, label) \ - do { \ - asm_x64_test_r8_with_r8(as, reg, reg); \ - asm_x64_jcc_label(as, ASM_X64_CC_JNZ, label); \ - } while (0) -#define ASM_JUMP_IF_REG_EQ(as, reg1, reg2, label) \ - do { \ - asm_x64_cmp_r64_with_r64(as, reg1, reg2); \ - asm_x64_jcc_label(as, ASM_X64_CC_JE, label); \ - } while (0) -#define ASM_CALL_IND(as, ptr, idx) asm_x64_call_ind(as, ptr, ASM_X64_REG_RAX) - -#define ASM_MOV_REG_TO_LOCAL asm_x64_mov_r64_to_local -#define ASM_MOV_IMM_TO_REG asm_x64_mov_i64_to_r64_optimised -#define ASM_MOV_ALIGNED_IMM_TO_REG asm_x64_mov_i64_to_r64_aligned -#define ASM_MOV_IMM_TO_LOCAL_USING(as, imm, local_num, reg_temp) \ - do { \ - asm_x64_mov_i64_to_r64_optimised(as, (imm), (reg_temp)); \ - asm_x64_mov_r64_to_local(as, (reg_temp), (local_num)); \ - } while (false) -#define ASM_MOV_LOCAL_TO_REG asm_x64_mov_local_to_r64 -#define ASM_MOV_REG_REG(as, reg_dest, reg_src) asm_x64_mov_r64_r64((as), (reg_dest), (reg_src)) -#define ASM_MOV_LOCAL_ADDR_TO_REG asm_x64_mov_local_addr_to_r64 - -#define ASM_LSL_REG(as, reg) asm_x64_shl_r64_cl((as), (reg)) -#define ASM_ASR_REG(as, reg) asm_x64_sar_r64_cl((as), (reg)) -#define ASM_OR_REG_REG(as, reg_dest, reg_src) asm_x64_or_r64_r64((as), (reg_dest), (reg_src)) -#define ASM_XOR_REG_REG(as, reg_dest, reg_src) asm_x64_xor_r64_r64((as), (reg_dest), (reg_src)) -#define ASM_AND_REG_REG(as, reg_dest, reg_src) asm_x64_and_r64_r64((as), (reg_dest), (reg_src)) -#define ASM_ADD_REG_REG(as, reg_dest, reg_src) asm_x64_add_r64_r64((as), (reg_dest), (reg_src)) -#define ASM_SUB_REG_REG(as, reg_dest, reg_src) asm_x64_sub_r64_r64((as), (reg_dest), (reg_src)) -#define ASM_MUL_REG_REG(as, reg_dest, reg_src) asm_x64_mul_r64_r64((as), (reg_dest), (reg_src)) - -#define ASM_LOAD_REG_REG(as, reg_dest, reg_base) asm_x64_mov_mem64_to_r64((as), (reg_base), 0, (reg_dest)) -#define ASM_LOAD_REG_REG_OFFSET(as, reg_dest, reg_base, word_offset) asm_x64_mov_mem64_to_r64((as), (reg_base), 8 * (word_offset), (reg_dest)) -#define ASM_LOAD8_REG_REG(as, reg_dest, reg_base) asm_x64_mov_mem8_to_r64zx((as), (reg_base), 0, (reg_dest)) -#define ASM_LOAD16_REG_REG(as, reg_dest, reg_base) asm_x64_mov_mem16_to_r64zx((as), (reg_base), 0, (reg_dest)) -#define ASM_LOAD32_REG_REG(as, reg_dest, reg_base) asm_x64_mov_mem32_to_r64zx((as), (reg_base), 0, (reg_dest)) - -#define ASM_STORE_REG_REG(as, reg_src, reg_base) asm_x64_mov_r64_to_mem64((as), (reg_src), (reg_base), 0) -#define ASM_STORE_REG_REG_OFFSET(as, reg_src, reg_base, word_offset) asm_x64_mov_r64_to_mem64((as), (reg_src), (reg_base), 8 * (word_offset)) -#define ASM_STORE8_REG_REG(as, reg_src, reg_base) asm_x64_mov_r8_to_mem8((as), (reg_src), (reg_base), 0) -#define ASM_STORE16_REG_REG(as, reg_src, reg_base) asm_x64_mov_r16_to_mem16((as), (reg_src), (reg_base), 0) -#define ASM_STORE32_REG_REG(as, reg_src, reg_base) asm_x64_mov_r32_to_mem32((as), (reg_src), (reg_base), 0) - #elif N_X86 // x86 specific stuff -#include "py/asmx86.h" - STATIC byte mp_f_n_args[MP_F_NUMBER_OF] = { [MP_F_CONVERT_OBJ_TO_NATIVE] = 2, [MP_F_CONVERT_NATIVE_TO_OBJ] = 2, @@ -201,247 +124,21 @@ STATIC byte mp_f_n_args[MP_F_NUMBER_OF] = { [MP_F_SETUP_CODE_STATE] = 5, }; +#include "py/asmx86.h" #define EXPORT_FUN(name) emit_native_x86_##name -#define ASM_WORD_SIZE (4) - -#define REG_RET ASM_X86_REG_EAX -#define REG_ARG_1 ASM_X86_REG_ARG_1 -#define REG_ARG_2 ASM_X86_REG_ARG_2 -#define REG_ARG_3 ASM_X86_REG_ARG_3 -#define REG_ARG_4 ASM_X86_REG_ARG_4 -#define REG_ARG_5 ASM_X86_REG_ARG_5 - -// caller-save, so can be used as temporaries -#define REG_TEMP0 ASM_X86_REG_EAX -#define REG_TEMP1 ASM_X86_REG_ECX -#define REG_TEMP2 ASM_X86_REG_EDX - -// callee-save, so can be used as locals -#define REG_LOCAL_1 ASM_X86_REG_EBX -#define REG_LOCAL_2 ASM_X86_REG_ESI -#define REG_LOCAL_3 ASM_X86_REG_EDI -#define REG_LOCAL_NUM (3) - -#define ASM_T asm_x86_t -#define ASM_END_PASS asm_x86_end_pass -#define ASM_ENTRY asm_x86_entry -#define ASM_EXIT asm_x86_exit - -#define ASM_JUMP asm_x86_jmp_label -#define ASM_JUMP_IF_REG_ZERO(as, reg, label) \ - do { \ - asm_x86_test_r8_with_r8(as, reg, reg); \ - asm_x86_jcc_label(as, ASM_X86_CC_JZ, label); \ - } while (0) -#define ASM_JUMP_IF_REG_NONZERO(as, reg, label) \ - do { \ - asm_x86_test_r8_with_r8(as, reg, reg); \ - asm_x86_jcc_label(as, ASM_X86_CC_JNZ, label); \ - } while (0) -#define ASM_JUMP_IF_REG_EQ(as, reg1, reg2, label) \ - do { \ - asm_x86_cmp_r32_with_r32(as, reg1, reg2); \ - asm_x86_jcc_label(as, ASM_X86_CC_JE, label); \ - } while (0) -#define ASM_CALL_IND(as, ptr, idx) asm_x86_call_ind(as, ptr, mp_f_n_args[idx], ASM_X86_REG_EAX) - -#define ASM_MOV_REG_TO_LOCAL asm_x86_mov_r32_to_local -#define ASM_MOV_IMM_TO_REG asm_x86_mov_i32_to_r32 -#define ASM_MOV_ALIGNED_IMM_TO_REG asm_x86_mov_i32_to_r32_aligned -#define ASM_MOV_IMM_TO_LOCAL_USING(as, imm, local_num, reg_temp) \ - do { \ - asm_x86_mov_i32_to_r32(as, (imm), (reg_temp)); \ - asm_x86_mov_r32_to_local(as, (reg_temp), (local_num)); \ - } while (false) -#define ASM_MOV_LOCAL_TO_REG asm_x86_mov_local_to_r32 -#define ASM_MOV_REG_REG(as, reg_dest, reg_src) asm_x86_mov_r32_r32((as), (reg_dest), (reg_src)) -#define ASM_MOV_LOCAL_ADDR_TO_REG asm_x86_mov_local_addr_to_r32 - -#define ASM_LSL_REG(as, reg) asm_x86_shl_r32_cl((as), (reg)) -#define ASM_ASR_REG(as, reg) asm_x86_sar_r32_cl((as), (reg)) -#define ASM_OR_REG_REG(as, reg_dest, reg_src) asm_x86_or_r32_r32((as), (reg_dest), (reg_src)) -#define ASM_XOR_REG_REG(as, reg_dest, reg_src) asm_x86_xor_r32_r32((as), (reg_dest), (reg_src)) -#define ASM_AND_REG_REG(as, reg_dest, reg_src) asm_x86_and_r32_r32((as), (reg_dest), (reg_src)) -#define ASM_ADD_REG_REG(as, reg_dest, reg_src) asm_x86_add_r32_r32((as), (reg_dest), (reg_src)) -#define ASM_SUB_REG_REG(as, reg_dest, reg_src) asm_x86_sub_r32_r32((as), (reg_dest), (reg_src)) -#define ASM_MUL_REG_REG(as, reg_dest, reg_src) asm_x86_mul_r32_r32((as), (reg_dest), (reg_src)) - -#define ASM_LOAD_REG_REG(as, reg_dest, reg_base) asm_x86_mov_mem32_to_r32((as), (reg_base), 0, (reg_dest)) -#define ASM_LOAD_REG_REG_OFFSET(as, reg_dest, reg_base, word_offset) asm_x86_mov_mem32_to_r32((as), (reg_base), 4 * (word_offset), (reg_dest)) -#define ASM_LOAD8_REG_REG(as, reg_dest, reg_base) asm_x86_mov_mem8_to_r32zx((as), (reg_base), 0, (reg_dest)) -#define ASM_LOAD16_REG_REG(as, reg_dest, reg_base) asm_x86_mov_mem16_to_r32zx((as), (reg_base), 0, (reg_dest)) -#define ASM_LOAD32_REG_REG(as, reg_dest, reg_base) asm_x86_mov_mem32_to_r32((as), (reg_base), 0, (reg_dest)) - -#define ASM_STORE_REG_REG(as, reg_src, reg_base) asm_x86_mov_r32_to_mem32((as), (reg_src), (reg_base), 0) -#define ASM_STORE_REG_REG_OFFSET(as, reg_src, reg_base, word_offset) asm_x86_mov_r32_to_mem32((as), (reg_src), (reg_base), 4 * (word_offset)) -#define ASM_STORE8_REG_REG(as, reg_src, reg_base) asm_x86_mov_r8_to_mem8((as), (reg_src), (reg_base), 0) -#define ASM_STORE16_REG_REG(as, reg_src, reg_base) asm_x86_mov_r16_to_mem16((as), (reg_src), (reg_base), 0) -#define ASM_STORE32_REG_REG(as, reg_src, reg_base) asm_x86_mov_r32_to_mem32((as), (reg_src), (reg_base), 0) - #elif N_THUMB // thumb specific stuff - #include "py/asmthumb.h" - #define EXPORT_FUN(name) emit_native_thumb_##name -#define ASM_WORD_SIZE (4) - -#define REG_RET ASM_THUMB_REG_R0 -#define REG_ARG_1 ASM_THUMB_REG_R0 -#define REG_ARG_2 ASM_THUMB_REG_R1 -#define REG_ARG_3 ASM_THUMB_REG_R2 -#define REG_ARG_4 ASM_THUMB_REG_R3 -// rest of args go on stack - -#define REG_TEMP0 ASM_THUMB_REG_R0 -#define REG_TEMP1 ASM_THUMB_REG_R1 -#define REG_TEMP2 ASM_THUMB_REG_R2 - -#define REG_LOCAL_1 ASM_THUMB_REG_R4 -#define REG_LOCAL_2 ASM_THUMB_REG_R5 -#define REG_LOCAL_3 ASM_THUMB_REG_R6 -#define REG_LOCAL_NUM (3) - -#define ASM_T asm_thumb_t -#define ASM_END_PASS asm_thumb_end_pass -#define ASM_ENTRY asm_thumb_entry -#define ASM_EXIT asm_thumb_exit - -#define ASM_JUMP asm_thumb_b_label -#define ASM_JUMP_IF_REG_ZERO(as, reg, label) \ - do { \ - asm_thumb_cmp_rlo_i8(as, reg, 0); \ - asm_thumb_bcc_label(as, ASM_THUMB_CC_EQ, label); \ - } while (0) -#define ASM_JUMP_IF_REG_NONZERO(as, reg, label) \ - do { \ - asm_thumb_cmp_rlo_i8(as, reg, 0); \ - asm_thumb_bcc_label(as, ASM_THUMB_CC_NE, label); \ - } while (0) -#define ASM_JUMP_IF_REG_EQ(as, reg1, reg2, label) \ - do { \ - asm_thumb_cmp_rlo_rlo(as, reg1, reg2); \ - asm_thumb_bcc_label(as, ASM_THUMB_CC_EQ, label); \ - } while (0) -#define ASM_CALL_IND(as, ptr, idx) asm_thumb_bl_ind(as, ptr, idx, ASM_THUMB_REG_R3) - -#define ASM_MOV_REG_TO_LOCAL(as, reg, local_num) asm_thumb_mov_local_reg(as, (local_num), (reg)) -#define ASM_MOV_IMM_TO_REG(as, imm, reg) asm_thumb_mov_reg_i32_optimised(as, (reg), (imm)) -#define ASM_MOV_ALIGNED_IMM_TO_REG(as, imm, reg) asm_thumb_mov_reg_i32_aligned(as, (reg), (imm)) -#define ASM_MOV_IMM_TO_LOCAL_USING(as, imm, local_num, reg_temp) \ - do { \ - asm_thumb_mov_reg_i32_optimised(as, (reg_temp), (imm)); \ - asm_thumb_mov_local_reg(as, (local_num), (reg_temp)); \ - } while (false) -#define ASM_MOV_LOCAL_TO_REG(as, local_num, reg) asm_thumb_mov_reg_local(as, (reg), (local_num)) -#define ASM_MOV_REG_REG(as, reg_dest, reg_src) asm_thumb_mov_reg_reg((as), (reg_dest), (reg_src)) -#define ASM_MOV_LOCAL_ADDR_TO_REG(as, local_num, reg) asm_thumb_mov_reg_local_addr(as, (reg), (local_num)) - -#define ASM_LSL_REG_REG(as, reg_dest, reg_shift) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_LSL, (reg_dest), (reg_shift)) -#define ASM_ASR_REG_REG(as, reg_dest, reg_shift) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_ASR, (reg_dest), (reg_shift)) -#define ASM_OR_REG_REG(as, reg_dest, reg_src) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_ORR, (reg_dest), (reg_src)) -#define ASM_XOR_REG_REG(as, reg_dest, reg_src) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_EOR, (reg_dest), (reg_src)) -#define ASM_AND_REG_REG(as, reg_dest, reg_src) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_AND, (reg_dest), (reg_src)) -#define ASM_ADD_REG_REG(as, reg_dest, reg_src) asm_thumb_add_rlo_rlo_rlo((as), (reg_dest), (reg_dest), (reg_src)) -#define ASM_SUB_REG_REG(as, reg_dest, reg_src) asm_thumb_sub_rlo_rlo_rlo((as), (reg_dest), (reg_dest), (reg_src)) -#define ASM_MUL_REG_REG(as, reg_dest, reg_src) asm_thumb_format_4((as), ASM_THUMB_FORMAT_4_MUL, (reg_dest), (reg_src)) - -#define ASM_LOAD_REG_REG(as, reg_dest, reg_base) asm_thumb_ldr_rlo_rlo_i5((as), (reg_dest), (reg_base), 0) -#define ASM_LOAD_REG_REG_OFFSET(as, reg_dest, reg_base, word_offset) asm_thumb_ldr_rlo_rlo_i5((as), (reg_dest), (reg_base), (word_offset)) -#define ASM_LOAD8_REG_REG(as, reg_dest, reg_base) asm_thumb_ldrb_rlo_rlo_i5((as), (reg_dest), (reg_base), 0) -#define ASM_LOAD16_REG_REG(as, reg_dest, reg_base) asm_thumb_ldrh_rlo_rlo_i5((as), (reg_dest), (reg_base), 0) -#define ASM_LOAD32_REG_REG(as, reg_dest, reg_base) asm_thumb_ldr_rlo_rlo_i5((as), (reg_dest), (reg_base), 0) - -#define ASM_STORE_REG_REG(as, reg_src, reg_base) asm_thumb_str_rlo_rlo_i5((as), (reg_src), (reg_base), 0) -#define ASM_STORE_REG_REG_OFFSET(as, reg_src, reg_base, word_offset) asm_thumb_str_rlo_rlo_i5((as), (reg_src), (reg_base), (word_offset)) -#define ASM_STORE8_REG_REG(as, reg_src, reg_base) asm_thumb_strb_rlo_rlo_i5((as), (reg_src), (reg_base), 0) -#define ASM_STORE16_REG_REG(as, reg_src, reg_base) asm_thumb_strh_rlo_rlo_i5((as), (reg_src), (reg_base), 0) -#define ASM_STORE32_REG_REG(as, reg_src, reg_base) asm_thumb_str_rlo_rlo_i5((as), (reg_src), (reg_base), 0) - #elif N_ARM // ARM specific stuff - #include "py/asmarm.h" - -#define ASM_WORD_SIZE (4) - #define EXPORT_FUN(name) emit_native_arm_##name -#define REG_RET ASM_ARM_REG_R0 -#define REG_ARG_1 ASM_ARM_REG_R0 -#define REG_ARG_2 ASM_ARM_REG_R1 -#define REG_ARG_3 ASM_ARM_REG_R2 -#define REG_ARG_4 ASM_ARM_REG_R3 - -#define REG_TEMP0 ASM_ARM_REG_R0 -#define REG_TEMP1 ASM_ARM_REG_R1 -#define REG_TEMP2 ASM_ARM_REG_R2 - -#define REG_LOCAL_1 ASM_ARM_REG_R4 -#define REG_LOCAL_2 ASM_ARM_REG_R5 -#define REG_LOCAL_3 ASM_ARM_REG_R6 -#define REG_LOCAL_NUM (3) - -#define ASM_T asm_arm_t -#define ASM_END_PASS asm_arm_end_pass -#define ASM_ENTRY asm_arm_entry -#define ASM_EXIT asm_arm_exit - -#define ASM_JUMP asm_arm_b_label -#define ASM_JUMP_IF_REG_ZERO(as, reg, label) \ - do { \ - asm_arm_cmp_reg_i8(as, reg, 0); \ - asm_arm_bcc_label(as, ASM_ARM_CC_EQ, label); \ - } while (0) -#define ASM_JUMP_IF_REG_NONZERO(as, reg, label) \ - do { \ - asm_arm_cmp_reg_i8(as, reg, 0); \ - asm_arm_bcc_label(as, ASM_ARM_CC_NE, label); \ - } while (0) -#define ASM_JUMP_IF_REG_EQ(as, reg1, reg2, label) \ - do { \ - asm_arm_cmp_reg_reg(as, reg1, reg2); \ - asm_arm_bcc_label(as, ASM_ARM_CC_EQ, label); \ - } while (0) -#define ASM_CALL_IND(as, ptr, idx) asm_arm_bl_ind(as, ptr, idx, ASM_ARM_REG_R3) - -#define ASM_MOV_REG_TO_LOCAL(as, reg, local_num) asm_arm_mov_local_reg(as, (local_num), (reg)) -#define ASM_MOV_IMM_TO_REG(as, imm, reg) asm_arm_mov_reg_i32(as, (reg), (imm)) -#define ASM_MOV_ALIGNED_IMM_TO_REG(as, imm, reg) asm_arm_mov_reg_i32(as, (reg), (imm)) -#define ASM_MOV_IMM_TO_LOCAL_USING(as, imm, local_num, reg_temp) \ - do { \ - asm_arm_mov_reg_i32(as, (reg_temp), (imm)); \ - asm_arm_mov_local_reg(as, (local_num), (reg_temp)); \ - } while (false) -#define ASM_MOV_LOCAL_TO_REG(as, local_num, reg) asm_arm_mov_reg_local(as, (reg), (local_num)) -#define ASM_MOV_REG_REG(as, reg_dest, reg_src) asm_arm_mov_reg_reg((as), (reg_dest), (reg_src)) -#define ASM_MOV_LOCAL_ADDR_TO_REG(as, local_num, reg) asm_arm_mov_reg_local_addr(as, (reg), (local_num)) - -#define ASM_LSL_REG_REG(as, reg_dest, reg_shift) asm_arm_lsl_reg_reg((as), (reg_dest), (reg_shift)) -#define ASM_ASR_REG_REG(as, reg_dest, reg_shift) asm_arm_asr_reg_reg((as), (reg_dest), (reg_shift)) -#define ASM_OR_REG_REG(as, reg_dest, reg_src) asm_arm_orr_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) -#define ASM_XOR_REG_REG(as, reg_dest, reg_src) asm_arm_eor_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) -#define ASM_AND_REG_REG(as, reg_dest, reg_src) asm_arm_and_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) -#define ASM_ADD_REG_REG(as, reg_dest, reg_src) asm_arm_add_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) -#define ASM_SUB_REG_REG(as, reg_dest, reg_src) asm_arm_sub_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) -#define ASM_MUL_REG_REG(as, reg_dest, reg_src) asm_arm_mul_reg_reg_reg((as), (reg_dest), (reg_dest), (reg_src)) - -#define ASM_LOAD_REG_REG(as, reg_dest, reg_base) asm_arm_ldr_reg_reg((as), (reg_dest), (reg_base), 0) -#define ASM_LOAD_REG_REG_OFFSET(as, reg_dest, reg_base, word_offset) asm_arm_ldr_reg_reg((as), (reg_dest), (reg_base), 4 * (word_offset)) -#define ASM_LOAD8_REG_REG(as, reg_dest, reg_base) asm_arm_ldrb_reg_reg((as), (reg_dest), (reg_base)) -#define ASM_LOAD16_REG_REG(as, reg_dest, reg_base) asm_arm_ldrh_reg_reg((as), (reg_dest), (reg_base)) -#define ASM_LOAD32_REG_REG(as, reg_dest, reg_base) asm_arm_ldr_reg_reg((as), (reg_dest), (reg_base), 0) - -#define ASM_STORE_REG_REG(as, reg_value, reg_base) asm_arm_str_reg_reg((as), (reg_value), (reg_base), 0) -#define ASM_STORE_REG_REG_OFFSET(as, reg_dest, reg_base, word_offset) asm_arm_str_reg_reg((as), (reg_dest), (reg_base), 4 * (word_offset)) -#define ASM_STORE8_REG_REG(as, reg_value, reg_base) asm_arm_strb_reg_reg((as), (reg_value), (reg_base)) -#define ASM_STORE16_REG_REG(as, reg_value, reg_base) asm_arm_strh_reg_reg((as), (reg_value), (reg_base)) -#define ASM_STORE32_REG_REG(as, reg_value, reg_base) asm_arm_str_reg_reg((as), (reg_value), (reg_base), 0) - #else #error unknown native emitter -- cgit v1.2.3 From 155fdc74d5864266441887d6c88111159f401a62 Mon Sep 17 00:00:00 2001 From: Damien George Date: Fri, 9 Dec 2016 22:50:58 +1100 Subject: py/asm: Remove need for dummy_data when doing initial assembler passes. For all but the last pass the assembler only needs to count how much space is needed for the machine code, it doesn't actually need to emit anything. The dummy_data just uses unnecessary RAM and without it the code is not any more complex (and code size does not increase for Thumb and Xtensa archs). --- py/asmarm.c | 5 ++++- py/asmbase.c | 18 ++++++++---------- py/asmbase.h | 3 --- py/asmthumb.c | 20 ++++++++++++-------- py/asmx64.c | 46 ++++++++++++++++++++++++++++------------------ py/asmx64.h | 1 - py/asmx86.c | 36 ++++++++++++++++++++++-------------- py/asmx86.h | 1 - py/asmxtensa.c | 16 ++++++++++------ 9 files changed, 84 insertions(+), 62 deletions(-) (limited to 'py/asmx86.h') diff --git a/py/asmarm.c b/py/asmarm.c index 663f86156..63963d2c0 100644 --- a/py/asmarm.c +++ b/py/asmarm.c @@ -55,7 +55,10 @@ void asm_arm_end_pass(asm_arm_t *as) { // Insert word into instruction flow STATIC void emit(asm_arm_t *as, uint op) { - *(uint*)asm_arm_get_cur_to_write_bytes(as, 4) = op; + uint8_t *c = mp_asm_base_get_cur_to_write_bytes(&as->base, 4); + if (c != NULL) { + *(uint32_t*)c = op; + } } // Insert word into instruction flow, add "ALWAYS" condition code diff --git a/py/asmbase.c b/py/asmbase.c index 848730593..c941e917b 100644 --- a/py/asmbase.c +++ b/py/asmbase.c @@ -59,17 +59,16 @@ void mp_asm_base_start_pass(mp_asm_base_t *as, int pass) { } // all functions must go through this one to emit bytes -// if as->pass < MP_ASM_PASS_EMIT, then this function returns dummy_data +// if as->pass < MP_ASM_PASS_EMIT, then this function just counts the number +// of bytes needed and returns NULL, and callers should not store any data uint8_t *mp_asm_base_get_cur_to_write_bytes(mp_asm_base_t *as, size_t num_bytes_to_write) { - if (as->pass < MP_ASM_PASS_EMIT) { - as->code_offset += num_bytes_to_write; - return as->dummy_data; - } else { + uint8_t *c = NULL; + if (as->pass == MP_ASM_PASS_EMIT) { assert(as->code_offset + num_bytes_to_write <= as->code_size); - uint8_t *c = as->code_base + as->code_offset; - as->code_offset += num_bytes_to_write; - return c; + c = as->code_base + as->code_offset; } + as->code_offset += num_bytes_to_write; + return c; } void mp_asm_base_label_assign(mp_asm_base_t *as, size_t label) { @@ -92,8 +91,7 @@ void mp_asm_base_align(mp_asm_base_t* as, unsigned int align) { // this function assumes a little endian machine void mp_asm_base_data(mp_asm_base_t* as, unsigned int bytesize, uintptr_t val) { uint8_t *c = mp_asm_base_get_cur_to_write_bytes(as, bytesize); - // only write to the buffer in the emit pass (otherwise we may overflow dummy_data) - if (as->pass == MP_ASM_PASS_EMIT) { + if (c != NULL) { for (unsigned int i = 0; i < bytesize; i++) { *c++ = val; val >>= 8; diff --git a/py/asmbase.h b/py/asmbase.h index 06fdd4b90..d2b403893 100644 --- a/py/asmbase.h +++ b/py/asmbase.h @@ -40,9 +40,6 @@ typedef struct _mp_asm_base_t { size_t max_num_labels; size_t *label_offsets; - - // must be last in struct - uint8_t dummy_data[4]; } mp_asm_base_t; void mp_asm_base_init(mp_asm_base_t *as, size_t max_num_labels); diff --git a/py/asmthumb.c b/py/asmthumb.c index 82a226b62..749c1e405 100644 --- a/py/asmthumb.c +++ b/py/asmthumb.c @@ -162,18 +162,22 @@ STATIC mp_uint_t get_label_dest(asm_thumb_t *as, uint label) { void asm_thumb_op16(asm_thumb_t *as, uint op) { byte *c = asm_thumb_get_cur_to_write_bytes(as, 2); - // little endian - c[0] = op; - c[1] = op >> 8; + if (c != NULL) { + // little endian + c[0] = op; + c[1] = op >> 8; + } } void asm_thumb_op32(asm_thumb_t *as, uint op1, uint op2) { byte *c = asm_thumb_get_cur_to_write_bytes(as, 4); - // little endian, op1 then op2 - c[0] = op1; - c[1] = op1 >> 8; - c[2] = op2; - c[3] = op2 >> 8; + if (c != NULL) { + // little endian, op1 then op2 + c[0] = op1; + c[1] = op1 >> 8; + c[2] = op2; + c[3] = op2 >> 8; + } } #define OP_FORMAT_4(op, rlo_dest, rlo_src) ((op) | ((rlo_src) << 3) | (rlo_dest)) diff --git a/py/asmx64.c b/py/asmx64.c index c9dad2a66..cf1a86b3f 100644 --- a/py/asmx64.c +++ b/py/asmx64.c @@ -122,40 +122,50 @@ static inline byte *asm_x64_get_cur_to_write_bytes(asm_x64_t *as, int n) { STATIC void asm_x64_write_byte_1(asm_x64_t *as, byte b1) { byte* c = asm_x64_get_cur_to_write_bytes(as, 1); - c[0] = b1; + if (c != NULL) { + c[0] = b1; + } } STATIC void asm_x64_write_byte_2(asm_x64_t *as, byte b1, byte b2) { byte* c = asm_x64_get_cur_to_write_bytes(as, 2); - c[0] = b1; - c[1] = b2; + if (c != NULL) { + c[0] = b1; + c[1] = b2; + } } STATIC void asm_x64_write_byte_3(asm_x64_t *as, byte b1, byte b2, byte b3) { byte* c = asm_x64_get_cur_to_write_bytes(as, 3); - c[0] = b1; - c[1] = b2; - c[2] = b3; + if (c != NULL) { + c[0] = b1; + c[1] = b2; + c[2] = b3; + } } STATIC void asm_x64_write_word32(asm_x64_t *as, int w32) { byte* c = asm_x64_get_cur_to_write_bytes(as, 4); - c[0] = IMM32_L0(w32); - c[1] = IMM32_L1(w32); - c[2] = IMM32_L2(w32); - c[3] = IMM32_L3(w32); + if (c != NULL) { + c[0] = IMM32_L0(w32); + c[1] = IMM32_L1(w32); + c[2] = IMM32_L2(w32); + c[3] = IMM32_L3(w32); + } } STATIC void asm_x64_write_word64(asm_x64_t *as, int64_t w64) { byte* c = asm_x64_get_cur_to_write_bytes(as, 8); - c[0] = IMM32_L0(w64); - c[1] = IMM32_L1(w64); - c[2] = IMM32_L2(w64); - c[3] = IMM32_L3(w64); - c[4] = IMM64_L4(w64); - c[5] = IMM64_L5(w64); - c[6] = IMM64_L6(w64); - c[7] = IMM64_L7(w64); + if (c != NULL) { + c[0] = IMM32_L0(w64); + c[1] = IMM32_L1(w64); + c[2] = IMM32_L2(w64); + c[3] = IMM32_L3(w64); + c[4] = IMM64_L4(w64); + c[5] = IMM64_L5(w64); + c[6] = IMM64_L6(w64); + c[7] = IMM64_L7(w64); + } } /* unused diff --git a/py/asmx64.h b/py/asmx64.h index f80c8da8b..4499c53c3 100644 --- a/py/asmx64.h +++ b/py/asmx64.h @@ -72,7 +72,6 @@ typedef struct _asm_x64_t { mp_asm_base_t base; - byte dummy_data[4]; // in addition to dummy_data in base int num_locals; } asm_x64_t; diff --git a/py/asmx86.c b/py/asmx86.c index cb9b30d40..ef315bf43 100644 --- a/py/asmx86.c +++ b/py/asmx86.c @@ -101,29 +101,37 @@ #define SIGNED_FIT8(x) (((x) & 0xffffff80) == 0) || (((x) & 0xffffff80) == 0xffffff80) STATIC void asm_x86_write_byte_1(asm_x86_t *as, byte b1) { - byte* c = asm_x86_get_cur_to_write_bytes(as, 1); - c[0] = b1; + byte* c = mp_asm_base_get_cur_to_write_bytes(&as->base, 1); + if (c != NULL) { + c[0] = b1; + } } STATIC void asm_x86_write_byte_2(asm_x86_t *as, byte b1, byte b2) { - byte* c = asm_x86_get_cur_to_write_bytes(as, 2); - c[0] = b1; - c[1] = b2; + byte* c = mp_asm_base_get_cur_to_write_bytes(&as->base, 2); + if (c != NULL) { + c[0] = b1; + c[1] = b2; + } } STATIC void asm_x86_write_byte_3(asm_x86_t *as, byte b1, byte b2, byte b3) { - byte* c = asm_x86_get_cur_to_write_bytes(as, 3); - c[0] = b1; - c[1] = b2; - c[2] = b3; + byte* c = mp_asm_base_get_cur_to_write_bytes(&as->base, 3); + if (c != NULL) { + c[0] = b1; + c[1] = b2; + c[2] = b3; + } } STATIC void asm_x86_write_word32(asm_x86_t *as, int w32) { - byte* c = asm_x86_get_cur_to_write_bytes(as, 4); - c[0] = IMM32_L0(w32); - c[1] = IMM32_L1(w32); - c[2] = IMM32_L2(w32); - c[3] = IMM32_L3(w32); + byte* c = mp_asm_base_get_cur_to_write_bytes(&as->base, 4); + if (c != NULL) { + c[0] = IMM32_L0(w32); + c[1] = IMM32_L1(w32); + c[2] = IMM32_L2(w32); + c[3] = IMM32_L3(w32); + } } STATIC void asm_x86_write_r32_disp(asm_x86_t *as, int r32, int disp_r32, int disp_offset) { diff --git a/py/asmx86.h b/py/asmx86.h index ac4b2ecd0..0b44af663 100644 --- a/py/asmx86.h +++ b/py/asmx86.h @@ -75,7 +75,6 @@ typedef struct _asm_x86_t { mp_asm_base_t base; - byte dummy_data[4]; // in addition to dummy_data in base int num_locals; } asm_x86_t; diff --git a/py/asmxtensa.c b/py/asmxtensa.c index 00df432ce..00448dfc5 100644 --- a/py/asmxtensa.c +++ b/py/asmxtensa.c @@ -97,15 +97,19 @@ STATIC uint32_t get_label_dest(asm_xtensa_t *as, uint label) { void asm_xtensa_op16(asm_xtensa_t *as, uint16_t op) { uint8_t *c = mp_asm_base_get_cur_to_write_bytes(&as->base, 2); - c[0] = op; - c[1] = op >> 8; + if (c != NULL) { + c[0] = op; + c[1] = op >> 8; + } } void asm_xtensa_op24(asm_xtensa_t *as, uint32_t op) { uint8_t *c = mp_asm_base_get_cur_to_write_bytes(&as->base, 3); - c[0] = op; - c[1] = op >> 8; - c[2] = op >> 16; + if (c != NULL) { + c[0] = op; + c[1] = op >> 8; + c[2] = op >> 16; + } } void asm_xtensa_j_label(asm_xtensa_t *as, uint label) { @@ -147,7 +151,7 @@ void asm_xtensa_mov_reg_i32(asm_xtensa_t *as, uint reg_dest, uint32_t i32) { // load the constant asm_xtensa_op_l32r(as, reg_dest, as->base.code_offset, 4 + as->cur_const * WORD_SIZE); // store the constant in the table - if (as->base.pass == MP_ASM_PASS_EMIT) { + if (as->const_table != NULL) { as->const_table[as->cur_const] = i32; } ++as->cur_const; -- cgit v1.2.3