diff --git a/py/asmthumb.c b/py/asmthumb.c index bb10935fda..e3558b2cf2 100644 --- a/py/asmthumb.c +++ b/py/asmthumb.c @@ -47,6 +47,7 @@ #define SIGNED_FIT12(x) (((x) & 0xfffff800) == 0) || (((x) & 0xfffff800) == 0xfffff800) #define SIGNED_FIT23(x) (((x) & 0xffc00000) == 0) || (((x) & 0xffc00000) == 0xffc00000) +#if MICROPY_EMIT_THUMB_ARMV7M // Note: these actually take an imm12 but the high-bit is not encoded here #define OP_ADD_W_RRI_HI(reg_src) (0xf200 | (reg_src)) #define OP_ADD_W_RRI_LO(reg_dest, imm11) ((imm11 << 4 & 0x7000) | reg_dest << 8 | (imm11 & 0xff)) @@ -55,6 +56,7 @@ #define OP_LDR_W_HI(reg_base) (0xf8d0 | (reg_base)) #define OP_LDR_W_LO(reg_dest, imm12) ((reg_dest) << 12 | (imm12)) +#endif static inline byte *asm_thumb_get_cur_to_write_bytes(asm_thumb_t *as, int n) { return mp_asm_base_get_cur_to_write_bytes(&as->base, n); @@ -122,7 +124,7 @@ void asm_thumb_entry(asm_thumb_t *as, int num_locals) { // If this Thumb machine code is run from ARM state then add a prelude // to switch to Thumb state for the duration of the function. - #if MICROPY_DYNAMIC_COMPILER || MICROPY_EMIT_ARM || (defined(__arm__) && !defined(__thumb2__)) + #if MICROPY_DYNAMIC_COMPILER || MICROPY_EMIT_ARM || (defined(__arm__) && !defined(__thumb2__) && !defined(__thumb__)) #if MICROPY_DYNAMIC_COMPILER if (mp_dynamic_compiler.native_arch == MP_NATIVE_ARCH_ARMV6) #endif @@ -171,11 +173,21 @@ void asm_thumb_entry(asm_thumb_t *as, int num_locals) { } asm_thumb_op16(as, OP_PUSH_RLIST_LR(reglist)); if (stack_adjust > 0) { + #if MICROPY_EMIT_THUMB_ARMV7M if (UNSIGNED_FIT7(stack_adjust)) { asm_thumb_op16(as, OP_SUB_SP(stack_adjust)); } else { asm_thumb_op32(as, OP_SUB_W_RRI_HI(ASM_THUMB_REG_SP), OP_SUB_W_RRI_LO(ASM_THUMB_REG_SP, stack_adjust * 4)); } + #else + int adj = stack_adjust; + // we don't expect the stack_adjust to be massive + while (!UNSIGNED_FIT7(adj)) { + asm_thumb_op16(as, OP_SUB_SP(127)); + adj -= 127; + } + asm_thumb_op16(as, OP_SUB_SP(adj)); + #endif } as->push_reglist = reglist; as->stack_adjust = stack_adjust; @@ -183,11 +195,21 @@ void asm_thumb_entry(asm_thumb_t *as, int num_locals) { void asm_thumb_exit(asm_thumb_t *as) { if (as->stack_adjust > 0) { + #if MICROPY_EMIT_THUMB_ARMV7M if (UNSIGNED_FIT7(as->stack_adjust)) { asm_thumb_op16(as, OP_ADD_SP(as->stack_adjust)); } else { asm_thumb_op32(as, OP_ADD_W_RRI_HI(ASM_THUMB_REG_SP), OP_ADD_W_RRI_LO(ASM_THUMB_REG_SP, as->stack_adjust * 4)); } + #else + int adj = as->stack_adjust; + // we don't expect the stack_adjust to be massive + while (!UNSIGNED_FIT7(adj)) { + asm_thumb_op16(as, OP_ADD_SP(127)); + adj -= 127; + } + asm_thumb_op16(as, OP_ADD_SP(adj)); + #endif } asm_thumb_op16(as, OP_POP_RLIST_PC(as->push_reglist)); } @@ -241,6 +263,8 @@ void asm_thumb_mov_reg_reg(asm_thumb_t *as, uint reg_dest, uint reg_src) { asm_thumb_op16(as, 0x4600 | op_lo); } +#if MICROPY_EMIT_THUMB_ARMV7M + // if loading lo half with movw, the i16 value will be zero extended into the r32 register! size_t asm_thumb_mov_reg_i16(asm_thumb_t *as, uint mov_op, uint reg_dest, int i16_src) { assert(reg_dest < ASM_THUMB_REG_R15); @@ -250,6 +274,16 @@ size_t asm_thumb_mov_reg_i16(asm_thumb_t *as, uint mov_op, uint reg_dest, int i1 return loc; } +#else + +void asm_thumb_mov_rlo_i16(asm_thumb_t *as, uint rlo_dest, int i16_src) { + asm_thumb_mov_rlo_i8(as, rlo_dest, (i16_src >> 8) & 0xff); + asm_thumb_lsl_rlo_rlo_i5(as, rlo_dest, rlo_dest, 8); + asm_thumb_add_rlo_i8(as, rlo_dest, i16_src & 0xff); +} + +#endif + #define OP_B_N(byte_offset) (0xe000 | (((byte_offset) >> 1) & 0x07ff)) bool asm_thumb_b_n_label(asm_thumb_t *as, uint label) { @@ -274,8 +308,13 @@ bool asm_thumb_bcc_nw_label(asm_thumb_t *as, int cond, uint label, bool wide) { asm_thumb_op16(as, OP_BCC_N(cond, rel)); return as->base.pass != MP_ASM_PASS_EMIT || SIGNED_FIT9(rel); } else { + #if MICROPY_EMIT_THUMB_ARMV7M asm_thumb_op32(as, OP_BCC_W_HI(cond, rel), OP_BCC_W_LO(rel)); return true; + #else + // this method should not be called for ARMV6M + return false; + #endif } } @@ -296,8 +335,30 @@ size_t asm_thumb_mov_reg_i32(asm_thumb_t *as, uint reg_dest, mp_uint_t i32) { size_t loc = mp_asm_base_get_code_pos(&as->base); + #if MICROPY_EMIT_THUMB_ARMV7M asm_thumb_mov_reg_i16(as, ASM_THUMB_OP_MOVW, reg_dest, i32); asm_thumb_mov_reg_i16(as, ASM_THUMB_OP_MOVT, reg_dest, i32 >> 16); + #else + // should only be called with lo reg for ARMV6M + assert(reg_dest < ASM_THUMB_REG_R8); + + // sanity check that generated code is aligned + assert(!as->base.code_base || !(3u & (uintptr_t)as->base.code_base)); + + // basically: + // (nop) + // ldr reg_dest, _data + // b 1f + // _data: .word i32 + // 1: + if (as->base.code_offset & 2u) { + asm_thumb_op16(as, ASM_THUMB_OP_NOP); + } + asm_thumb_ldr_rlo_pcrel_i8(as, reg_dest, 0); + asm_thumb_op16(as, OP_B_N(2)); + asm_thumb_op16(as, i32 & 0xffff); + asm_thumb_op16(as, i32 >> 16); + #endif return loc; } @@ -305,27 +366,68 @@ size_t asm_thumb_mov_reg_i32(asm_thumb_t *as, uint reg_dest, mp_uint_t i32) { void asm_thumb_mov_reg_i32_optimised(asm_thumb_t *as, uint reg_dest, int i32) { if (reg_dest < 8 && UNSIGNED_FIT8(i32)) { asm_thumb_mov_rlo_i8(as, reg_dest, i32); - } else if (UNSIGNED_FIT16(i32)) { - asm_thumb_mov_reg_i16(as, ASM_THUMB_OP_MOVW, reg_dest, i32); } else { - asm_thumb_mov_reg_i32(as, reg_dest, i32); + #if MICROPY_EMIT_THUMB_ARMV7M + if (UNSIGNED_FIT16(i32)) { + asm_thumb_mov_reg_i16(as, ASM_THUMB_OP_MOVW, reg_dest, i32); + } else { + asm_thumb_mov_reg_i32(as, reg_dest, i32); + } + #else + uint rlo_dest = reg_dest; + assert(rlo_dest < ASM_THUMB_REG_R8); // should never be called for ARMV6M + + bool negate = i32 < 0 && ((i32 + i32) & 0xffffffffu); // don't negate 0x80000000 + if (negate) { + i32 = -i32; + } + + uint clz = __builtin_clz(i32); + uint ctz = i32 ? __builtin_ctz(i32) : 0; + assert(clz + ctz <= 32); + if (clz + ctz >= 24) { + asm_thumb_mov_rlo_i8(as, rlo_dest, (i32 >> ctz) & 0xff); + asm_thumb_lsl_rlo_rlo_i5(as, rlo_dest, rlo_dest, ctz); + } else if (UNSIGNED_FIT16(i32)) { + asm_thumb_mov_rlo_i16(as, rlo_dest, i32); + } else { + if (negate) { + // no point in negating if we're storing in 32 bit anyway + negate = false; + i32 = -i32; + } + asm_thumb_mov_reg_i32(as, rlo_dest, i32); + } + if (negate) { + asm_thumb_neg_rlo_rlo(as, rlo_dest, rlo_dest); + } + #endif } } #define OP_STR_TO_SP_OFFSET(rlo_dest, word_offset) (0x9000 | ((rlo_dest) << 8) | ((word_offset) & 0x00ff)) #define OP_LDR_FROM_SP_OFFSET(rlo_dest, word_offset) (0x9800 | ((rlo_dest) << 8) | ((word_offset) & 0x00ff)) +static void asm_thumb_mov_local_check(asm_thumb_t *as, int word_offset) { + if (as->base.pass >= MP_ASM_PASS_EMIT) { + assert(word_offset >= 0); + if (!UNSIGNED_FIT8(word_offset)) { + mp_raise_NotImplementedError(MP_ERROR_TEXT("too many locals for native method")); + } + } +} + void asm_thumb_mov_local_reg(asm_thumb_t *as, int local_num, uint rlo_src) { assert(rlo_src < ASM_THUMB_REG_R8); int word_offset = local_num; - assert(as->base.pass < MP_ASM_PASS_EMIT || word_offset >= 0); + asm_thumb_mov_local_check(as, word_offset); asm_thumb_op16(as, OP_STR_TO_SP_OFFSET(rlo_src, word_offset)); } void asm_thumb_mov_reg_local(asm_thumb_t *as, uint rlo_dest, int local_num) { assert(rlo_dest < ASM_THUMB_REG_R8); int word_offset = local_num; - assert(as->base.pass < MP_ASM_PASS_EMIT || word_offset >= 0); + asm_thumb_mov_local_check(as, word_offset); asm_thumb_op16(as, OP_LDR_FROM_SP_OFFSET(rlo_dest, word_offset)); } @@ -341,21 +443,63 @@ void asm_thumb_mov_reg_local_addr(asm_thumb_t *as, uint rlo_dest, int local_num) void asm_thumb_mov_reg_pcrel(asm_thumb_t *as, uint rlo_dest, uint label) { mp_uint_t dest = get_label_dest(as, label); mp_int_t rel = dest - as->base.code_offset; - rel -= 4 + 4; // adjust for mov_reg_i16 and then PC+4 prefetch of add_reg_reg rel |= 1; // to stay in Thumb state when jumping to this address + #if MICROPY_EMIT_THUMB_ARMV7M + rel -= 4 + 4; // adjust for mov_reg_i16 and then PC+4 prefetch of add_reg_reg asm_thumb_mov_reg_i16(as, ASM_THUMB_OP_MOVW, rlo_dest, rel); // 4 bytes + #else + rel -= 8 + 4; // adjust for four instructions and then PC+4 prefetch of add_reg_reg + // 6 bytes + asm_thumb_mov_rlo_i16(as, rlo_dest, rel); + // 2 bytes - not always needed, but we want to keep the size the same + asm_thumb_sxth_rlo_rlo(as, rlo_dest, rlo_dest); + #endif asm_thumb_add_reg_reg(as, rlo_dest, ASM_THUMB_REG_R15); // 2 bytes } +#if MICROPY_EMIT_THUMB_ARMV7M static inline void asm_thumb_ldr_reg_reg_i12(asm_thumb_t *as, uint reg_dest, uint reg_base, uint word_offset) { asm_thumb_op32(as, OP_LDR_W_HI(reg_base), OP_LDR_W_LO(reg_dest, word_offset * 4)); } +#endif void asm_thumb_ldr_reg_reg_i12_optimised(asm_thumb_t *as, uint reg_dest, uint reg_base, uint word_offset) { if (reg_dest < ASM_THUMB_REG_R8 && reg_base < ASM_THUMB_REG_R8 && UNSIGNED_FIT5(word_offset)) { asm_thumb_ldr_rlo_rlo_i5(as, reg_dest, reg_base, word_offset); } else { + #if MICROPY_EMIT_THUMB_ARMV7M asm_thumb_ldr_reg_reg_i12(as, reg_dest, reg_base, word_offset); + #else + word_offset -= 31; + if (reg_dest < ASM_THUMB_REG_R8 && reg_base < ASM_THUMB_REG_R8) { + if (UNSIGNED_FIT8(word_offset) && (word_offset < 64 || reg_dest != reg_base)) { + if (word_offset < 64) { + if (reg_dest != reg_base) { + asm_thumb_mov_reg_reg(as, reg_dest, reg_base); + } + asm_thumb_add_rlo_i8(as, reg_dest, word_offset * 4); + } else { + asm_thumb_mov_rlo_i8(as, reg_dest, word_offset); + asm_thumb_lsl_rlo_rlo_i5(as, reg_dest, reg_dest, 2); + asm_thumb_add_rlo_rlo_rlo(as, reg_dest, reg_dest, reg_base); + } + } else { + if (reg_dest != reg_base) { + asm_thumb_mov_rlo_i16(as, reg_dest, word_offset * 4); + asm_thumb_add_rlo_rlo_rlo(as, reg_dest, reg_dest, reg_dest); + } else { + uint reg_other = reg_dest ^ 7; + asm_thumb_op16(as, OP_PUSH_RLIST((1 << reg_other))); + asm_thumb_mov_rlo_i16(as, reg_other, word_offset * 4); + asm_thumb_add_rlo_rlo_rlo(as, reg_dest, reg_dest, reg_other); + asm_thumb_op16(as, OP_POP_RLIST((1 << reg_other))); + } + } + } else { + assert(0); // should never be called for ARMV6M + } + asm_thumb_ldr_rlo_rlo_i5(as, reg_dest, reg_dest, 31); + #endif } } @@ -378,7 +522,20 @@ void asm_thumb_b_label(asm_thumb_t *as, uint label) { } else { // is a forwards jump, so need to assume it's large large_jump: + #if MICROPY_EMIT_THUMB_ARMV7M asm_thumb_op32(as, OP_BW_HI(rel), OP_BW_LO(rel)); + #else + if (SIGNED_FIT12(rel)) { + // this code path has to be the same number of instructions irrespective of rel + asm_thumb_op16(as, OP_B_N(rel)); + } else { + asm_thumb_op16(as, ASM_THUMB_OP_NOP); + if (dest != (mp_uint_t)-1) { + // we have an actual branch > 12 bits; this is not handled yet + mp_raise_NotImplementedError(MP_ERROR_TEXT("native method too big")); + } + } + #endif } } @@ -397,7 +554,13 @@ void asm_thumb_bcc_label(asm_thumb_t *as, int cond, uint label) { } else { // is a forwards jump, so need to assume it's large large_jump: + #if MICROPY_EMIT_THUMB_ARMV7M asm_thumb_op32(as, OP_BCC_W_HI(cond, rel), OP_BCC_W_LO(rel)); + #else + // reverse the sense of the branch to jump over a longer branch + asm_thumb_op16(as, OP_BCC_N(cond ^ 1, 0)); + asm_thumb_b_label(as, label); + #endif } } diff --git a/py/asmthumb.h b/py/asmthumb.h index 17b694a74d..3c45336186 100644 --- a/py/asmthumb.h +++ b/py/asmthumb.h @@ -157,6 +157,7 @@ static inline void asm_thumb_sub_rlo_rlo_i3(asm_thumb_t *as, uint rlo_dest, uint #define ASM_THUMB_FORMAT_3_CMP (0x2800) #define ASM_THUMB_FORMAT_3_ADD (0x3000) #define ASM_THUMB_FORMAT_3_SUB (0x3800) +#define ASM_THUMB_FORMAT_3_LDR (0x4800) #define ASM_THUMB_FORMAT_3_ENCODE(op, rlo, i8) ((op) | ((rlo) << 8) | (i8)) @@ -177,6 +178,9 @@ static inline void asm_thumb_add_rlo_i8(asm_thumb_t *as, uint rlo, int i8) { static inline void asm_thumb_sub_rlo_i8(asm_thumb_t *as, uint rlo, int i8) { asm_thumb_format_3(as, ASM_THUMB_FORMAT_3_SUB, rlo, i8); } +static inline void asm_thumb_ldr_rlo_pcrel_i8(asm_thumb_t *as, uint rlo, uint i8) { + asm_thumb_format_3(as, ASM_THUMB_FORMAT_3_LDR, rlo, i8); +} // FORMAT 4: ALU operations @@ -202,6 +206,12 @@ void asm_thumb_format_4(asm_thumb_t *as, uint op, uint rlo_dest, uint rlo_src); static inline void asm_thumb_cmp_rlo_rlo(asm_thumb_t *as, uint rlo_dest, uint rlo_src) { asm_thumb_format_4(as, ASM_THUMB_FORMAT_4_CMP, rlo_dest, rlo_src); } +static inline void asm_thumb_mvn_rlo_rlo(asm_thumb_t *as, uint rlo_dest, uint rlo_src) { + asm_thumb_format_4(as, ASM_THUMB_FORMAT_4_MVN, rlo_dest, rlo_src); +} +static inline void asm_thumb_neg_rlo_rlo(asm_thumb_t *as, uint rlo_dest, uint rlo_src) { + asm_thumb_format_4(as, ASM_THUMB_FORMAT_4_NEG, rlo_dest, rlo_src); +} // FORMAT 5: hi register operations (add, cmp, mov, bx) // For add/cmp/mov, at least one of the args must be a high register @@ -263,6 +273,32 @@ static inline void asm_thumb_ldrb_rlo_rlo_i5(asm_thumb_t *as, uint rlo_dest, uin static inline void asm_thumb_ldrh_rlo_rlo_i5(asm_thumb_t *as, uint rlo_dest, uint rlo_base, uint byte_offset) { asm_thumb_format_9_10(as, ASM_THUMB_FORMAT_10_LDRH, rlo_dest, rlo_base, byte_offset); } +static inline void asm_thumb_lsl_rlo_rlo_i5(asm_thumb_t *as, uint rlo_dest, uint rlo_src, uint shift) { + asm_thumb_format_1(as, ASM_THUMB_FORMAT_1_LSL, rlo_dest, rlo_src, shift); +} +static inline void asm_thumb_asr_rlo_rlo_i5(asm_thumb_t *as, uint rlo_dest, uint rlo_src, uint shift) { + asm_thumb_format_1(as, ASM_THUMB_FORMAT_1_ASR, rlo_dest, rlo_src, shift); +} + +// FORMAT 11: sign/zero extend + +#define ASM_THUMB_FORMAT_11_ENCODE(op, rlo_dest, rlo_src) \ + ((op) | ((rlo_src) << 3) | (rlo_dest)) + +#define ASM_THUMB_FORMAT_11_SXTH (0xb200) +#define ASM_THUMB_FORMAT_11_SXTB (0xb240) +#define ASM_THUMB_FORMAT_11_UXTH (0xb280) +#define ASM_THUMB_FORMAT_11_UXTB (0xb2c0) + +static inline void asm_thumb_format_11(asm_thumb_t *as, uint op, uint rlo_dest, uint rlo_src) { + assert(rlo_dest < ASM_THUMB_REG_R8); + assert(rlo_src < ASM_THUMB_REG_R8); + asm_thumb_op16(as, ASM_THUMB_FORMAT_11_ENCODE(op, rlo_dest, rlo_src)); +} + +static inline void asm_thumb_sxth_rlo_rlo(asm_thumb_t *as, uint rlo_dest, uint rlo_src) { + asm_thumb_format_11(as, ASM_THUMB_FORMAT_11_SXTH, rlo_dest, rlo_src); +} // TODO convert these to above format style @@ -270,7 +306,12 @@ static inline void asm_thumb_ldrh_rlo_rlo_i5(asm_thumb_t *as, uint rlo_dest, uin #define ASM_THUMB_OP_MOVT (0xf2c0) void asm_thumb_mov_reg_reg(asm_thumb_t *as, uint reg_dest, uint reg_src); + +#if MICROPY_EMIT_THUMB_ARMV7M size_t asm_thumb_mov_reg_i16(asm_thumb_t *as, uint mov_op, uint reg_dest, int i16_src); +#else +void asm_thumb_mov_rlo_i16(asm_thumb_t *as, uint rlo_dest, int i16_src); +#endif // these return true if the destination is in range, false otherwise bool asm_thumb_b_n_label(asm_thumb_t *as, uint label); @@ -344,7 +385,11 @@ void asm_thumb_bl_ind(asm_thumb_t *as, uint fun_id, uint reg_temp); // convenien #define ASM_MOV_LOCAL_REG(as, local_num, reg) asm_thumb_mov_local_reg((as), (local_num), (reg)) #define ASM_MOV_REG_IMM(as, reg_dest, imm) asm_thumb_mov_reg_i32_optimised((as), (reg_dest), (imm)) +#if MICROPY_EMIT_THUMB_ARMV7M #define ASM_MOV_REG_IMM_FIX_U16(as, reg_dest, imm) asm_thumb_mov_reg_i16((as), ASM_THUMB_OP_MOVW, (reg_dest), (imm)) +#else +#define ASM_MOV_REG_IMM_FIX_U16(as, reg_dest, imm) asm_thumb_mov_rlo_i16((as), (reg_dest), (imm)) +#endif #define ASM_MOV_REG_IMM_FIX_WORD(as, reg_dest, imm) asm_thumb_mov_reg_i32((as), (reg_dest), (imm)) #define ASM_MOV_REG_LOCAL(as, reg_dest, local_num) asm_thumb_mov_reg_local((as), (reg_dest), (local_num)) #define ASM_MOV_REG_REG(as, reg_dest, reg_src) asm_thumb_mov_reg_reg((as), (reg_dest), (reg_src)) diff --git a/py/emitinlinethumb.c b/py/emitinlinethumb.c index cffaa4bb89..073b88da71 100644 --- a/py/emitinlinethumb.c +++ b/py/emitinlinethumb.c @@ -573,7 +573,11 @@ STATIC void emit_inline_thumb_op(emit_inline_asm_t *emit, qstr op, mp_uint_t n_a goto unknown_op; } int label_num = get_arg_label(emit, op_str, pn_args[0]); - if (!asm_thumb_bcc_nw_label(&emit->as, cc, label_num, op_len == 5 && op_str[4] == 'w')) { + bool wide = op_len == 5 && op_str[4] == 'w'; + if (wide && !ARMV7M) { + goto unknown_op; + } + if (!asm_thumb_bcc_nw_label(&emit->as, cc, label_num, wide)) { goto branch_not_in_range; } } else if (ARMV7M && op_str[0] == 'i' && op_str[1] == 't') { diff --git a/py/mpconfig.h b/py/mpconfig.h index 854188b66b..1df2e8fe91 100644 --- a/py/mpconfig.h +++ b/py/mpconfig.h @@ -304,6 +304,11 @@ #define MICROPY_EMIT_THUMB (0) #endif +// Whether to emit ARMv7-M instruction support in thumb native code +#ifndef MICROPY_EMIT_THUMB_ARMV7M +#define MICROPY_EMIT_THUMB_ARMV7M (1) +#endif + // Whether to enable the thumb inline assembler #ifndef MICROPY_EMIT_INLINE_THUMB #define MICROPY_EMIT_INLINE_THUMB (0)