1 // SPDX-License-Identifier: LGPL-2.1-or-later
3 * Copyright (C) 2014-2021 Paul Cercueil <paul@crapouillou.net>
6 #include "blockcache.h"
8 #include "disassembler.h"
10 #include "interpreter.h"
11 #include "lightrec-config.h"
12 #include "lightning-wrapper.h"
14 #include "memmanager.h"
16 #include "recompiler.h"
18 #include "optimizer.h"
23 #if ENABLE_THREADED_COMPILER
24 #include <stdatomic.h>
33 #define GENMASK(h, l) \
34 (((uintptr_t)-1 << (l)) & ((uintptr_t)-1 >> (__WORDSIZE - 1 - (h))))
36 static struct block * lightrec_precompile_block(struct lightrec_state *state,
38 static bool lightrec_block_is_fully_tagged(const struct block *block);
40 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data);
41 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg);
43 static void lightrec_default_sb(struct lightrec_state *state, u32 opcode,
44 void *host, u32 addr, u8 data)
48 if (!state->invalidate_from_dma_only)
49 lightrec_invalidate(state, addr, 1);
52 static void lightrec_default_sh(struct lightrec_state *state, u32 opcode,
53 void *host, u32 addr, u16 data)
55 *(u16 *)host = HTOLE16(data);
57 if (!state->invalidate_from_dma_only)
58 lightrec_invalidate(state, addr, 2);
61 static void lightrec_default_sw(struct lightrec_state *state, u32 opcode,
62 void *host, u32 addr, u32 data)
64 *(u32 *)host = HTOLE32(data);
66 if (!state->invalidate_from_dma_only)
67 lightrec_invalidate(state, addr, 4);
70 static u8 lightrec_default_lb(struct lightrec_state *state,
71 u32 opcode, void *host, u32 addr)
76 static u16 lightrec_default_lh(struct lightrec_state *state,
77 u32 opcode, void *host, u32 addr)
79 return LE16TOH(*(u16 *)host);
82 static u32 lightrec_default_lw(struct lightrec_state *state,
83 u32 opcode, void *host, u32 addr)
85 return LE32TOH(*(u32 *)host);
88 static const struct lightrec_mem_map_ops lightrec_default_ops = {
89 .sb = lightrec_default_sb,
90 .sh = lightrec_default_sh,
91 .sw = lightrec_default_sw,
92 .lb = lightrec_default_lb,
93 .lh = lightrec_default_lh,
94 .lw = lightrec_default_lw,
97 static void __segfault_cb(struct lightrec_state *state, u32 addr,
98 const struct block *block)
100 lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
101 pr_err("Segmentation fault in recompiled code: invalid "
102 "load/store at address 0x%08x\n", addr);
104 pr_err("Was executing block PC 0x%08x\n", block->pc);
107 static void lightrec_swl(struct lightrec_state *state,
108 const struct lightrec_mem_map_ops *ops,
109 u32 opcode, void *host, u32 addr, u32 data)
111 unsigned int shift = addr & 0x3;
112 unsigned int mask = GENMASK(31, (shift + 1) * 8);
115 /* Align to 32 bits */
117 host = (void *)((uintptr_t)host & ~3);
119 old_data = ops->lw(state, opcode, host, addr);
121 data = (data >> ((3 - shift) * 8)) | (old_data & mask);
123 ops->sw(state, opcode, host, addr, data);
126 static void lightrec_swr(struct lightrec_state *state,
127 const struct lightrec_mem_map_ops *ops,
128 u32 opcode, void *host, u32 addr, u32 data)
130 unsigned int shift = addr & 0x3;
131 unsigned int mask = (1 << (shift * 8)) - 1;
134 /* Align to 32 bits */
136 host = (void *)((uintptr_t)host & ~3);
138 old_data = ops->lw(state, opcode, host, addr);
140 data = (data << (shift * 8)) | (old_data & mask);
142 ops->sw(state, opcode, host, addr, data);
145 static void lightrec_swc2(struct lightrec_state *state, union code op,
146 const struct lightrec_mem_map_ops *ops,
147 void *host, u32 addr)
149 u32 data = lightrec_mfc2(state, op.i.rt);
151 ops->sw(state, op.opcode, host, addr, data);
154 static u32 lightrec_lwl(struct lightrec_state *state,
155 const struct lightrec_mem_map_ops *ops,
156 u32 opcode, void *host, u32 addr, u32 data)
158 unsigned int shift = addr & 0x3;
159 unsigned int mask = (1 << (24 - shift * 8)) - 1;
162 /* Align to 32 bits */
164 host = (void *)((uintptr_t)host & ~3);
166 old_data = ops->lw(state, opcode, host, addr);
168 return (data & mask) | (old_data << (24 - shift * 8));
171 static u32 lightrec_lwr(struct lightrec_state *state,
172 const struct lightrec_mem_map_ops *ops,
173 u32 opcode, void *host, u32 addr, u32 data)
175 unsigned int shift = addr & 0x3;
176 unsigned int mask = GENMASK(31, 32 - shift * 8);
179 /* Align to 32 bits */
181 host = (void *)((uintptr_t)host & ~3);
183 old_data = ops->lw(state, opcode, host, addr);
185 return (data & mask) | (old_data >> (shift * 8));
188 static void lightrec_lwc2(struct lightrec_state *state, union code op,
189 const struct lightrec_mem_map_ops *ops,
190 void *host, u32 addr)
192 u32 data = ops->lw(state, op.opcode, host, addr);
194 lightrec_mtc2(state, op.i.rt, data);
197 static void lightrec_invalidate_map(struct lightrec_state *state,
198 const struct lightrec_mem_map *map, u32 addr, u32 len)
200 if (map == &state->maps[PSX_MAP_KERNEL_USER_RAM]) {
201 memset(&state->code_lut[lut_offset(addr)], 0,
202 ((len + 3) / 4) * sizeof(void *));
206 const struct lightrec_mem_map *
207 lightrec_get_map(struct lightrec_state *state, void **host, u32 kaddr)
209 const struct lightrec_mem_map *map;
213 for (i = 0; i < state->nb_maps; i++) {
214 const struct lightrec_mem_map *mapi = &state->maps[i];
216 if (kaddr >= mapi->pc && kaddr < mapi->pc + mapi->length) {
222 if (i == state->nb_maps)
225 addr = kaddr - map->pc;
227 while (map->mirror_of)
228 map = map->mirror_of;
231 *host = map->address + addr;
236 u32 lightrec_rw(struct lightrec_state *state, union code op,
237 u32 addr, u32 data, u16 *flags, struct block *block)
239 const struct lightrec_mem_map *map;
240 const struct lightrec_mem_map_ops *ops;
241 u32 opcode = op.opcode;
244 addr += (s16) op.i.imm;
246 map = lightrec_get_map(state, &host, kunseg(addr));
248 __segfault_cb(state, addr, block);
252 if (unlikely(map->ops)) {
254 *flags |= LIGHTREC_HW_IO;
259 *flags |= LIGHTREC_DIRECT_IO;
261 ops = &lightrec_default_ops;
266 ops->sb(state, opcode, host, addr, (u8) data);
269 ops->sh(state, opcode, host, addr, (u16) data);
272 lightrec_swl(state, ops, opcode, host, addr, data);
275 lightrec_swr(state, ops, opcode, host, addr, data);
278 ops->sw(state, opcode, host, addr, data);
281 lightrec_swc2(state, op, ops, host, addr);
284 return (s32) (s8) ops->lb(state, opcode, host, addr);
286 return ops->lb(state, opcode, host, addr);
288 return (s32) (s16) ops->lh(state, opcode, host, addr);
290 return ops->lh(state, opcode, host, addr);
292 lightrec_lwc2(state, op, ops, host, addr);
295 return lightrec_lwl(state, ops, opcode, host, addr, data);
297 return lightrec_lwr(state, ops, opcode, host, addr, data);
300 return ops->lw(state, opcode, host, addr);
304 static void lightrec_rw_helper(struct lightrec_state *state,
305 union code op, u16 *flags,
308 u32 ret = lightrec_rw(state, op, state->regs.gpr[op.i.rs],
309 state->regs.gpr[op.i.rt], flags, block);
320 state->regs.gpr[op.i.rt] = ret;
321 default: /* fall-through */
326 static void lightrec_rw_cb(struct lightrec_state *state, union code op)
328 lightrec_rw_helper(state, op, NULL, NULL);
331 static void lightrec_rw_generic_cb(struct lightrec_state *state, u32 arg)
336 u16 offset = (u16)arg;
338 block = lightrec_find_block_from_lut(state->block_cache,
339 arg >> 16, state->next_pc);
340 if (unlikely(!block)) {
341 pr_err("rw_generic: No block found in LUT for PC 0x%x offset 0x%x\n",
342 state->next_pc, offset);
346 op = &block->opcode_list[offset];
347 was_tagged = op->flags & (LIGHTREC_HW_IO | LIGHTREC_DIRECT_IO);
349 lightrec_rw_helper(state, op->c, &op->flags, block);
352 pr_debug("Opcode of block at PC 0x%08x has been tagged - flag "
353 "for recompilation\n", block->pc);
355 block->flags |= BLOCK_SHOULD_RECOMPILE;
359 static u32 clamp_s32(s32 val, s32 min, s32 max)
361 return val < min ? min : val > max ? max : val;
364 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg)
366 s16 gteir1, gteir2, gteir3;
376 return (s32)(s16) state->regs.cp2d[reg];
382 return (u16) state->regs.cp2d[reg];
385 gteir1 = (s16) state->regs.cp2d[9];
386 gteir2 = (s16) state->regs.cp2d[10];
387 gteir3 = (s16) state->regs.cp2d[11];
389 return clamp_s32(gteir1 >> 7, 0, 0x1f) << 0 |
390 clamp_s32(gteir2 >> 7, 0, 0x1f) << 5 |
391 clamp_s32(gteir3 >> 7, 0, 0x1f) << 10;
394 default: /* fall-through */
395 return state->regs.cp2d[reg];
399 u32 lightrec_mfc(struct lightrec_state *state, union code op)
401 if (op.i.op == OP_CP0)
402 return state->regs.cp0[op.r.rd];
403 else if (op.r.rs == OP_CP2_BASIC_MFC2)
404 return lightrec_mfc2(state, op.r.rd);
406 return state->regs.cp2c[op.r.rd];
409 static void lightrec_mfc_cb(struct lightrec_state *state, union code op)
411 u32 rt = lightrec_mfc(state, op);
414 state->regs.gpr[op.r.rt] = rt;
417 static void lightrec_mtc0(struct lightrec_state *state, u8 reg, u32 data)
427 /* Those registers are read-only */
429 default: /* fall-through */
434 status = state->regs.cp0[12];
436 if (status & ~data & BIT(16)) {
437 state->ops.enable_ram(state, true);
438 lightrec_invalidate_all(state);
439 } else if (~status & data & BIT(16)) {
440 state->ops.enable_ram(state, false);
444 state->regs.cp0[reg] = data;
446 if (reg == 12 || reg == 13) {
447 cause = state->regs.cp0[13];
448 status = state->regs.cp0[12];
450 if (!!(status & cause & 0x300) & status)
451 lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
455 static u32 count_leading_bits(s32 data)
460 #if __has_builtin(__builtin_clrsb)
461 return 1 + __builtin_clrsb(data);
465 data = (data ^ (data >> 31)) << 1;
475 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data)
479 state->regs.cp2d[12] = state->regs.cp2d[13];
480 state->regs.cp2d[13] = state->regs.cp2d[14];
481 state->regs.cp2d[14] = data;
484 state->regs.cp2d[9] = (data << 7) & 0xf80;
485 state->regs.cp2d[10] = (data << 2) & 0xf80;
486 state->regs.cp2d[11] = (data >> 3) & 0xf80;
491 state->regs.cp2d[31] = count_leading_bits((s32) data);
492 default: /* fall-through */
493 state->regs.cp2d[reg] = data;
498 static void lightrec_ctc2(struct lightrec_state *state, u8 reg, u32 data)
508 data = (s32)(s16) data;
511 data = (data & 0x7ffff000) | !!(data & 0x7f87e000) << 31;
512 default: /* fall-through */
516 state->regs.cp2c[reg] = data;
519 void lightrec_mtc(struct lightrec_state *state, union code op, u32 data)
521 if (op.i.op == OP_CP0)
522 lightrec_mtc0(state, op.r.rd, data);
523 else if (op.r.rs == OP_CP2_BASIC_CTC2)
524 lightrec_ctc2(state, op.r.rd, data);
526 lightrec_mtc2(state, op.r.rd, data);
529 static void lightrec_mtc_cb(struct lightrec_state *state, union code op)
531 lightrec_mtc(state, op, state->regs.gpr[op.r.rt]);
534 void lightrec_rfe(struct lightrec_state *state)
538 /* Read CP0 Status register (r12) */
539 status = state->regs.cp0[12];
541 /* Switch the bits */
542 status = ((status & 0x3c) >> 2) | (status & ~0xf);
545 lightrec_mtc0(state, 12, status);
548 void lightrec_cp(struct lightrec_state *state, union code op)
550 if (op.i.op == OP_CP0) {
551 pr_err("Invalid CP opcode to coprocessor #0\n");
555 (*state->ops.cop2_op)(state, op.opcode);
558 static void lightrec_syscall_cb(struct lightrec_state *state, union code op)
560 lightrec_set_exit_flags(state, LIGHTREC_EXIT_SYSCALL);
563 static void lightrec_break_cb(struct lightrec_state *state, union code op)
565 lightrec_set_exit_flags(state, LIGHTREC_EXIT_BREAK);
568 struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
570 struct block *block = lightrec_find_block(state->block_cache, pc);
572 if (block && lightrec_block_is_outdated(state, block)) {
573 pr_debug("Block at PC 0x%08x is outdated!\n", block->pc);
575 /* Make sure the recompiler isn't processing the block we'll
577 if (ENABLE_THREADED_COMPILER)
578 lightrec_recompiler_remove(state->rec, block);
580 lightrec_unregister_block(state->block_cache, block);
581 remove_from_code_lut(state->block_cache, block);
582 lightrec_free_block(state, block);
587 block = lightrec_precompile_block(state, pc);
589 pr_err("Unable to recompile block at PC 0x%x\n", pc);
590 lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
594 lightrec_register_block(state->block_cache, block);
600 static void * get_next_block_func(struct lightrec_state *state, u32 pc)
603 bool should_recompile;
607 func = state->code_lut[lut_offset(pc)];
608 if (func && func != state->get_next_block)
611 block = lightrec_get_block(state, pc);
613 if (unlikely(!block))
616 if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET)) {
617 func = state->memset_func;
621 should_recompile = block->flags & BLOCK_SHOULD_RECOMPILE &&
622 !(block->flags & BLOCK_IS_DEAD);
624 if (unlikely(should_recompile)) {
625 pr_debug("Block at PC 0x%08x should recompile\n", pc);
627 lightrec_unregister(MEM_FOR_CODE, block->code_size);
629 if (ENABLE_THREADED_COMPILER)
630 lightrec_recompiler_add(state->rec, block);
632 lightrec_compile_block(state->cstate, block);
635 if (ENABLE_THREADED_COMPILER && likely(!should_recompile))
636 func = lightrec_recompiler_run_first_pass(state, block, &pc);
638 func = block->function;
643 if (unlikely(block->flags & BLOCK_NEVER_COMPILE)) {
644 pc = lightrec_emulate_block(state, block, pc);
646 } else if (!ENABLE_THREADED_COMPILER) {
647 /* Block wasn't compiled yet - run the interpreter */
648 if (block->flags & BLOCK_FULLY_TAGGED)
649 pr_debug("Block fully tagged, skipping first pass\n");
650 else if (ENABLE_FIRST_PASS && likely(!should_recompile))
651 pc = lightrec_emulate_block(state, block, pc);
653 /* Then compile it using the profiled data */
654 lightrec_compile_block(state->cstate, block);
656 lightrec_recompiler_add(state->rec, block);
659 if (state->exit_flags != LIGHTREC_EXIT_NORMAL ||
660 state->current_cycle >= state->target_cycle)
668 static s32 c_function_wrapper(struct lightrec_state *state, s32 cycles_delta,
669 void (*f)(struct lightrec_state *, u32 d),
672 state->current_cycle = state->target_cycle - cycles_delta;
676 return state->target_cycle - state->current_cycle;
679 static struct block * generate_wrapper(struct lightrec_state *state)
685 jit_word_t code_size;
686 jit_node_t *to_tramp, *to_fn_epilog;
688 block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
692 _jit = jit_new_state();
696 jit_name("RW wrapper");
697 jit_note(__FILE__, __LINE__);
699 /* Wrapper entry point */
702 stack_ptr = jit_allocai(sizeof(uintptr_t) * NUM_TEMPS);
704 for (i = 0; i < NUM_TEMPS; i++)
705 jit_stxi(stack_ptr + i * sizeof(uintptr_t), JIT_FP, JIT_R(i));
707 /* Jump to the trampoline */
708 to_tramp = jit_jmpi();
710 /* The trampoline will jump back here */
711 to_fn_epilog = jit_label();
713 for (i = 0; i < NUM_TEMPS; i++)
714 jit_ldxi(JIT_R(i), JIT_FP, stack_ptr + i * sizeof(uintptr_t));
719 /* Trampoline entry point.
720 * The sole purpose of the trampoline is to cheese Lightning not to
721 * save/restore the callee-saved register LIGHTREC_REG_CYCLE, since we
722 * do want to return to the caller with this register modified. */
728 jit_pushargr(LIGHTREC_REG_STATE);
729 jit_pushargr(LIGHTREC_REG_CYCLE);
730 jit_pushargr(JIT_R0);
731 jit_pushargr(JIT_R1);
732 jit_finishi(c_function_wrapper);
733 jit_retval_i(LIGHTREC_REG_CYCLE);
735 jit_patch_at(jit_jmpi(), to_fn_epilog);
739 block->function = jit_emit();
740 block->opcode_list = NULL;
744 jit_get_code(&code_size);
745 lightrec_register(MEM_FOR_CODE, code_size);
747 block->code_size = code_size;
749 if (ENABLE_DISASSEMBLER) {
750 pr_debug("Wrapper block:\n");
758 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
760 pr_err("Unable to compile wrapper: Out of memory\n");
764 static u32 lightrec_memset(struct lightrec_state *state)
766 u32 kunseg_pc = kunseg(state->regs.gpr[4]);
768 const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg_pc);
769 u32 length = state->regs.gpr[5] * 4;
772 pr_err("Unable to find memory map for memset target address "
773 "0x%x\n", kunseg_pc);
777 pr_debug("Calling host memset, PC 0x%x (host address 0x%" PRIxPTR ") for %u bytes\n",
778 kunseg_pc, (uintptr_t)host, length);
779 memset(host, 0, length);
781 if (!state->invalidate_from_dma_only)
782 lightrec_invalidate_map(state, map, kunseg_pc, length);
784 /* Rough estimation of the number of cycles consumed */
785 return 8 + 5 * (length + 3 / 4);
788 static struct block * generate_dispatcher(struct lightrec_state *state)
792 jit_node_t *to_end, *to_c, *loop, *addr, *addr2, *addr3;
795 jit_word_t code_size;
797 block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
801 _jit = jit_new_state();
805 jit_name("dispatcher");
806 jit_note(__FILE__, __LINE__);
811 jit_getarg(JIT_R0, jit_arg());
812 jit_getarg_i(LIGHTREC_REG_CYCLE, jit_arg());
814 /* Force all callee-saved registers to be pushed on the stack */
815 for (i = 0; i < NUM_REGS; i++)
816 jit_movr(JIT_V(i), JIT_V(i));
818 /* Pass lightrec_state structure to blocks, using the last callee-saved
819 * register that Lightning provides */
820 jit_movi(LIGHTREC_REG_STATE, (intptr_t) state);
824 /* Call the block's code */
827 if (OPT_REPLACE_MEMSET) {
828 /* Blocks will jump here when they need to call
829 * lightrec_memset() */
830 addr3 = jit_indirect();
833 jit_pushargr(LIGHTREC_REG_STATE);
834 jit_finishi(lightrec_memset);
836 jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
837 offsetof(struct lightrec_state, regs.gpr[31]));
840 jit_subr(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, JIT_R0);
843 /* The block will jump here, with the number of cycles remaining in
844 * LIGHTREC_REG_CYCLE */
845 addr2 = jit_indirect();
847 /* Store back the next_pc to the lightrec_state structure */
848 offset = offsetof(struct lightrec_state, next_pc);
849 jit_stxi_i(offset, LIGHTREC_REG_STATE, JIT_V0);
851 /* Jump to end if state->target_cycle < state->current_cycle */
852 to_end = jit_blei(LIGHTREC_REG_CYCLE, 0);
854 /* Convert next PC to KUNSEG and avoid mirrors */
855 ram_len = state->maps[PSX_MAP_KERNEL_USER_RAM].length;
856 jit_andi(JIT_R0, JIT_V0, 0x10000000 | (ram_len - 1));
857 to_c = jit_bgei(JIT_R0, ram_len);
859 /* Fast path: code is running from RAM, use the code LUT */
860 if (__WORDSIZE == 64)
861 jit_lshi(JIT_R0, JIT_R0, 1);
862 jit_addr(JIT_R0, JIT_R0, LIGHTREC_REG_STATE);
863 jit_ldxi(JIT_R0, JIT_R0, offsetof(struct lightrec_state, code_lut));
865 /* If we get non-NULL, loop */
866 jit_patch_at(jit_bnei(JIT_R0, 0), loop);
868 /* Slow path: call C function get_next_block_func() */
871 if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
872 /* We may call the interpreter - update state->current_cycle */
873 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
874 offsetof(struct lightrec_state, target_cycle));
875 jit_subr(JIT_R1, JIT_R2, LIGHTREC_REG_CYCLE);
876 jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
877 LIGHTREC_REG_STATE, JIT_R1);
880 /* The code LUT will be set to this address when the block at the target
881 * PC has been preprocessed but not yet compiled by the threaded
883 addr = jit_indirect();
885 /* Get the next block */
887 jit_pushargr(LIGHTREC_REG_STATE);
888 jit_pushargr(JIT_V0);
889 jit_finishi(&get_next_block_func);
892 if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
893 /* The interpreter may have updated state->current_cycle and
894 * state->target_cycle - recalc the delta */
895 jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
896 offsetof(struct lightrec_state, current_cycle));
897 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
898 offsetof(struct lightrec_state, target_cycle));
899 jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
902 /* If we get non-NULL, loop */
903 jit_patch_at(jit_bnei(JIT_R0, 0), loop);
905 /* When exiting, the recompiled code will jump to that address */
906 jit_note(__FILE__, __LINE__);
909 jit_retr(LIGHTREC_REG_CYCLE);
913 block->function = jit_emit();
914 block->opcode_list = NULL;
918 jit_get_code(&code_size);
919 lightrec_register(MEM_FOR_CODE, code_size);
921 block->code_size = code_size;
923 state->eob_wrapper_func = jit_address(addr2);
924 if (OPT_REPLACE_MEMSET)
925 state->memset_func = jit_address(addr3);
926 state->get_next_block = jit_address(addr);
928 if (ENABLE_DISASSEMBLER) {
929 pr_debug("Dispatcher block:\n");
938 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
940 pr_err("Unable to compile dispatcher: Out of memory\n");
944 union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
948 lightrec_get_map(state, &host, kunseg(pc));
950 const u32 *code = (u32 *)host;
951 return (union code) *code;
954 unsigned int lightrec_cycles_of_opcode(union code code)
959 void lightrec_free_opcode_list(struct lightrec_state *state, struct block *block)
961 lightrec_free(state, MEM_FOR_IR,
962 sizeof(*block->opcode_list) * block->nb_ops,
966 static unsigned int lightrec_get_mips_block_len(const u32 *src)
972 c.opcode = LE32TOH(*src++);
977 if (is_unconditional_jump(c))
982 static struct opcode * lightrec_disassemble(struct lightrec_state *state,
983 const u32 *src, unsigned int *len)
986 unsigned int i, length;
988 length = lightrec_get_mips_block_len(src);
990 list = lightrec_malloc(state, MEM_FOR_IR, sizeof(*list) * length);
992 pr_err("Unable to allocate memory\n");
996 for (i = 0; i < length; i++) {
997 list[i].opcode = LE32TOH(src[i]);
1001 *len = length * sizeof(u32);
1006 static struct block * lightrec_precompile_block(struct lightrec_state *state,
1009 struct opcode *list;
1010 struct block *block;
1012 const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg(pc));
1013 const u32 *code = (u32 *) host;
1014 unsigned int length;
1020 block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1022 pr_err("Unable to recompile block: Out of memory\n");
1026 list = lightrec_disassemble(state, code, &length);
1028 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1034 block->function = NULL;
1035 block->opcode_list = list;
1039 block->code_size = 0;
1040 #if ENABLE_THREADED_COMPILER
1041 block->op_list_freed = (atomic_flag)ATOMIC_FLAG_INIT;
1043 block->nb_ops = length / sizeof(u32);
1045 lightrec_optimize(state, block);
1047 length = block->nb_ops * sizeof(u32);
1049 lightrec_register(MEM_FOR_MIPS_CODE, length);
1051 if (ENABLE_DISASSEMBLER) {
1052 pr_debug("Disassembled block at PC: 0x%08x\n", block->pc);
1053 lightrec_print_disassembly(block, code);
1056 pr_debug("Block size: %hu opcodes\n", block->nb_ops);
1058 /* If the first opcode is an 'impossible' branch, never compile the
1060 if (should_emulate(block->opcode_list))
1061 block->flags |= BLOCK_NEVER_COMPILE;
1063 fully_tagged = lightrec_block_is_fully_tagged(block);
1065 block->flags |= BLOCK_FULLY_TAGGED;
1067 if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET))
1068 state->code_lut[lut_offset(pc)] = state->memset_func;
1070 block->hash = lightrec_calculate_block_hash(block);
1072 pr_debug("Recompile count: %u\n", state->nb_precompile++);
1077 static bool lightrec_block_is_fully_tagged(const struct block *block)
1079 const struct opcode *op;
1082 for (i = 0; i < block->nb_ops; i++) {
1083 op = &block->opcode_list[i];
1085 /* Verify that all load/stores of the opcode list
1086 * Check all loads/stores of the opcode list and mark the
1087 * block as fully compiled if they all have been tagged. */
1088 switch (op->c.i.op) {
1103 if (!(op->flags & (LIGHTREC_DIRECT_IO |
1106 default: /* fall-through */
1114 static void lightrec_reap_block(struct lightrec_state *state, void *data)
1116 struct block *block = data;
1118 pr_debug("Reap dead block at PC 0x%08x\n", block->pc);
1119 lightrec_unregister_block(state->block_cache, block);
1120 lightrec_free_block(state, block);
1123 static void lightrec_reap_jit(struct lightrec_state *state, void *data)
1125 _jit_destroy_state(data);
1128 int lightrec_compile_block(struct lightrec_cstate *cstate,
1129 struct block *block)
1131 struct lightrec_state *state = cstate->state;
1132 struct lightrec_branch_target *target;
1133 bool op_list_freed = false, fully_tagged = false;
1134 struct block *block2;
1136 jit_state_t *_jit, *oldjit;
1137 jit_node_t *start_of_block;
1138 bool skip_next = false;
1139 jit_word_t code_size;
1143 fully_tagged = lightrec_block_is_fully_tagged(block);
1145 block->flags |= BLOCK_FULLY_TAGGED;
1147 _jit = jit_new_state();
1151 oldjit = block->_jit;
1154 lightrec_regcache_reset(cstate->reg_cache);
1156 cstate->nb_branches = 0;
1157 cstate->nb_local_branches = 0;
1158 cstate->nb_targets = 0;
1163 start_of_block = jit_label();
1165 for (i = 0; i < block->nb_ops; i++) {
1166 elm = &block->opcode_list[i];
1173 cstate->cycles += lightrec_cycles_of_opcode(elm->c);
1175 if (should_emulate(elm)) {
1176 pr_debug("Branch at offset 0x%x will be emulated\n",
1179 lightrec_emit_eob(cstate, block, i, false);
1180 skip_next = !(elm->flags & LIGHTREC_NO_DS);
1182 lightrec_rec_opcode(cstate, block, i);
1183 skip_next = has_delay_slot(elm->c) &&
1184 !(elm->flags & LIGHTREC_NO_DS);
1186 /* FIXME: GNU Lightning on Windows seems to use our
1187 * mapped registers as temporaries. Until the actual bug
1188 * is found and fixed, unconditionally mark our
1189 * registers as live here. */
1190 lightrec_regcache_mark_live(cstate->reg_cache, _jit);
1195 for (i = 0; i < cstate->nb_branches; i++)
1196 jit_patch(cstate->branches[i]);
1198 for (i = 0; i < cstate->nb_local_branches; i++) {
1199 struct lightrec_branch *branch = &cstate->local_branches[i];
1201 pr_debug("Patch local branch to offset 0x%x\n",
1202 branch->target << 2);
1204 if (branch->target == 0) {
1205 jit_patch_at(branch->branch, start_of_block);
1209 for (j = 0; j < cstate->nb_targets; j++) {
1210 if (cstate->targets[j].offset == branch->target) {
1211 jit_patch_at(branch->branch,
1212 cstate->targets[j].label);
1217 if (j == cstate->nb_targets)
1218 pr_err("Unable to find branch target\n");
1221 jit_ldxi(JIT_R0, LIGHTREC_REG_STATE,
1222 offsetof(struct lightrec_state, eob_wrapper_func));
1229 block->function = jit_emit();
1230 block->flags &= ~BLOCK_SHOULD_RECOMPILE;
1232 /* Add compiled function to the LUT */
1233 state->code_lut[lut_offset(block->pc)] = block->function;
1235 if (ENABLE_THREADED_COMPILER) {
1236 /* Since we might try to reap the same block multiple times,
1237 * we need the reaper to wait until everything has been
1238 * submitted, so that the duplicate entries can be dropped. */
1239 lightrec_reaper_pause(state->reaper);
1242 /* Detect old blocks that have been covered by the new one */
1243 for (i = 0; i < cstate->nb_targets; i++) {
1244 target = &cstate->targets[i];
1246 if (!target->offset)
1249 offset = block->pc + target->offset * sizeof(u32);
1250 block2 = lightrec_find_block(state->block_cache, offset);
1252 /* No need to check if block2 is compilable - it must
1253 * be, otherwise block wouldn't be compilable either */
1255 /* Set the "block dead" flag to prevent the dynarec from
1256 * recompiling this block */
1257 block2->flags |= BLOCK_IS_DEAD;
1259 /* If block2 was pending for compilation, cancel it.
1260 * If it's being compiled right now, wait until it
1262 if (ENABLE_THREADED_COMPILER)
1263 lightrec_recompiler_remove(state->rec, block2);
1265 /* We know from now on that block2 isn't going to be
1266 * compiled. We can override the LUT entry with our
1267 * new block's entry point. */
1268 offset = lut_offset(block->pc) + target->offset;
1269 state->code_lut[offset] = jit_address(target->label);
1271 pr_debug("Reap block 0x%08x as it's covered by block "
1272 "0x%08x\n", block2->pc, block->pc);
1274 /* Finally, reap the block. */
1275 if (ENABLE_THREADED_COMPILER) {
1276 lightrec_reaper_add(state->reaper,
1277 lightrec_reap_block,
1280 lightrec_unregister_block(state->block_cache, block2);
1281 lightrec_free_block(state, block2);
1286 if (ENABLE_THREADED_COMPILER)
1287 lightrec_reaper_continue(state->reaper);
1289 jit_get_code(&code_size);
1290 lightrec_register(MEM_FOR_CODE, code_size);
1292 block->code_size = code_size;
1294 if (ENABLE_DISASSEMBLER) {
1295 pr_debug("Compiling block at PC: 0x%08x\n", block->pc);
1301 #if ENABLE_THREADED_COMPILER
1303 op_list_freed = atomic_flag_test_and_set(&block->op_list_freed);
1305 if (fully_tagged && !op_list_freed) {
1306 pr_debug("Block PC 0x%08x is fully tagged"
1307 " - free opcode list\n", block->pc);
1308 lightrec_free_opcode_list(state, block);
1309 block->opcode_list = NULL;
1313 pr_debug("Block 0x%08x recompiled, reaping old jit context.\n",
1316 if (ENABLE_THREADED_COMPILER)
1317 lightrec_reaper_add(state->reaper,
1318 lightrec_reap_jit, oldjit);
1320 _jit_destroy_state(oldjit);
1326 static void lightrec_print_info(struct lightrec_state *state)
1328 if ((state->current_cycle & ~0xfffffff) != state->old_cycle_counter) {
1329 pr_info("Lightrec RAM usage: IR %u KiB, CODE %u KiB, "
1330 "MIPS %u KiB, TOTAL %u KiB, avg. IPI %f\n",
1331 lightrec_get_mem_usage(MEM_FOR_IR) / 1024,
1332 lightrec_get_mem_usage(MEM_FOR_CODE) / 1024,
1333 lightrec_get_mem_usage(MEM_FOR_MIPS_CODE) / 1024,
1334 lightrec_get_total_mem_usage() / 1024,
1335 lightrec_get_average_ipi());
1336 state->old_cycle_counter = state->current_cycle & ~0xfffffff;
1340 u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
1342 s32 (*func)(void *, s32) = (void *)state->dispatcher->function;
1346 state->exit_flags = LIGHTREC_EXIT_NORMAL;
1348 /* Handle the cycle counter overflowing */
1349 if (unlikely(target_cycle < state->current_cycle))
1350 target_cycle = UINT_MAX;
1352 state->target_cycle = target_cycle;
1353 state->next_pc = pc;
1355 block_trace = get_next_block_func(state, pc);
1357 cycles_delta = state->target_cycle - state->current_cycle;
1359 cycles_delta = (*func)(block_trace, cycles_delta);
1361 state->current_cycle = state->target_cycle - cycles_delta;
1364 if (ENABLE_THREADED_COMPILER)
1365 lightrec_reaper_reap(state->reaper);
1367 if (LOG_LEVEL >= INFO_L)
1368 lightrec_print_info(state);
1370 return state->next_pc;
1373 u32 lightrec_execute_one(struct lightrec_state *state, u32 pc)
1375 return lightrec_execute(state, pc, state->current_cycle);
1378 u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc)
1380 struct block *block = lightrec_get_block(state, pc);
1384 state->exit_flags = LIGHTREC_EXIT_NORMAL;
1386 pc = lightrec_emulate_block(state, block, pc);
1388 if (LOG_LEVEL >= INFO_L)
1389 lightrec_print_info(state);
1394 void lightrec_free_block(struct lightrec_state *state, struct block *block)
1396 lightrec_unregister(MEM_FOR_MIPS_CODE, block->nb_ops * sizeof(u32));
1397 if (block->opcode_list)
1398 lightrec_free_opcode_list(state, block);
1400 _jit_destroy_state(block->_jit);
1401 lightrec_unregister(MEM_FOR_CODE, block->code_size);
1402 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1405 struct lightrec_cstate * lightrec_create_cstate(struct lightrec_state *state)
1407 struct lightrec_cstate *cstate;
1409 cstate = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*cstate));
1413 cstate->reg_cache = lightrec_regcache_init(state);
1414 if (!cstate->reg_cache) {
1415 lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1419 cstate->state = state;
1424 void lightrec_free_cstate(struct lightrec_cstate *cstate)
1426 lightrec_free_regcache(cstate->reg_cache);
1427 lightrec_free(cstate->state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1430 struct lightrec_state * lightrec_init(char *argv0,
1431 const struct lightrec_mem_map *map,
1433 const struct lightrec_ops *ops)
1435 struct lightrec_state *state;
1437 /* Sanity-check ops */
1438 if (!ops || !ops->cop2_op || !ops->enable_ram) {
1439 pr_err("Missing callbacks in lightrec_ops structure\n");
1445 state = calloc(1, sizeof(*state) +
1446 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1448 goto err_finish_jit;
1450 lightrec_register(MEM_FOR_LIGHTREC, sizeof(*state) +
1451 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1454 state->tinymm = tinymm_init(malloc, free, 4096);
1456 goto err_free_state;
1459 state->block_cache = lightrec_blockcache_init(state);
1460 if (!state->block_cache)
1461 goto err_free_tinymm;
1463 if (ENABLE_THREADED_COMPILER) {
1464 state->rec = lightrec_recompiler_init(state);
1466 goto err_free_block_cache;
1468 state->reaper = lightrec_reaper_init(state);
1470 goto err_free_recompiler;
1472 state->cstate = lightrec_create_cstate(state);
1474 goto err_free_block_cache;
1477 state->nb_maps = nb;
1480 memcpy(&state->ops, ops, sizeof(*ops));
1482 state->dispatcher = generate_dispatcher(state);
1483 if (!state->dispatcher)
1484 goto err_free_reaper;
1486 state->c_wrapper_block = generate_wrapper(state);
1487 if (!state->c_wrapper_block)
1488 goto err_free_dispatcher;
1490 state->c_wrapper = state->c_wrapper_block->function;
1492 state->c_wrappers[C_WRAPPER_RW] = lightrec_rw_cb;
1493 state->c_wrappers[C_WRAPPER_RW_GENERIC] = lightrec_rw_generic_cb;
1494 state->c_wrappers[C_WRAPPER_MFC] = lightrec_mfc_cb;
1495 state->c_wrappers[C_WRAPPER_MTC] = lightrec_mtc_cb;
1496 state->c_wrappers[C_WRAPPER_CP] = lightrec_cp;
1497 state->c_wrappers[C_WRAPPER_SYSCALL] = lightrec_syscall_cb;
1498 state->c_wrappers[C_WRAPPER_BREAK] = lightrec_break_cb;
1500 map = &state->maps[PSX_MAP_BIOS];
1501 state->offset_bios = (uintptr_t)map->address - map->pc;
1503 map = &state->maps[PSX_MAP_SCRATCH_PAD];
1504 state->offset_scratch = (uintptr_t)map->address - map->pc;
1506 map = &state->maps[PSX_MAP_KERNEL_USER_RAM];
1507 state->offset_ram = (uintptr_t)map->address - map->pc;
1509 if (state->maps[PSX_MAP_MIRROR1].address == map->address + 0x200000 &&
1510 state->maps[PSX_MAP_MIRROR2].address == map->address + 0x400000 &&
1511 state->maps[PSX_MAP_MIRROR3].address == map->address + 0x600000)
1512 state->mirrors_mapped = true;
1514 if (state->offset_bios == 0 &&
1515 state->offset_scratch == 0 &&
1516 state->offset_ram == 0 &&
1517 state->mirrors_mapped) {
1518 pr_info("Memory map is perfect. Emitted code will be best.\n");
1520 pr_info("Memory map is sub-par. Emitted code will be slow.\n");
1525 err_free_dispatcher:
1526 lightrec_free_block(state, state->dispatcher);
1528 if (ENABLE_THREADED_COMPILER)
1529 lightrec_reaper_destroy(state->reaper);
1530 err_free_recompiler:
1531 if (ENABLE_THREADED_COMPILER)
1532 lightrec_free_recompiler(state->rec);
1534 lightrec_free_cstate(state->cstate);
1535 err_free_block_cache:
1536 lightrec_free_block_cache(state->block_cache);
1539 tinymm_shutdown(state->tinymm);
1542 lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1543 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1550 void lightrec_destroy(struct lightrec_state *state)
1552 /* Force a print info on destroy*/
1553 state->current_cycle = ~state->current_cycle;
1554 lightrec_print_info(state);
1556 if (ENABLE_THREADED_COMPILER) {
1557 lightrec_free_recompiler(state->rec);
1558 lightrec_reaper_destroy(state->reaper);
1560 lightrec_free_cstate(state->cstate);
1563 lightrec_free_block_cache(state->block_cache);
1564 lightrec_free_block(state, state->dispatcher);
1565 lightrec_free_block(state, state->c_wrapper_block);
1569 tinymm_shutdown(state->tinymm);
1571 lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1572 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1576 void lightrec_invalidate(struct lightrec_state *state, u32 addr, u32 len)
1578 u32 kaddr = kunseg(addr & ~0x3);
1579 const struct lightrec_mem_map *map = lightrec_get_map(state, NULL, kaddr);
1582 if (map != &state->maps[PSX_MAP_KERNEL_USER_RAM])
1585 /* Handle mirrors */
1586 kaddr &= (state->maps[PSX_MAP_KERNEL_USER_RAM].length - 1);
1588 lightrec_invalidate_map(state, map, kaddr, len);
1592 void lightrec_invalidate_all(struct lightrec_state *state)
1594 memset(state->code_lut, 0, sizeof(*state->code_lut) * CODE_LUT_SIZE);
1597 void lightrec_set_invalidate_mode(struct lightrec_state *state, bool dma_only)
1599 if (state->invalidate_from_dma_only != dma_only)
1600 lightrec_invalidate_all(state);
1602 state->invalidate_from_dma_only = dma_only;
1605 void lightrec_set_exit_flags(struct lightrec_state *state, u32 flags)
1607 if (flags != LIGHTREC_EXIT_NORMAL) {
1608 state->exit_flags |= flags;
1609 state->target_cycle = state->current_cycle;
1613 u32 lightrec_exit_flags(struct lightrec_state *state)
1615 return state->exit_flags;
1618 u32 lightrec_current_cycle_count(const struct lightrec_state *state)
1620 return state->current_cycle;
1623 void lightrec_reset_cycle_count(struct lightrec_state *state, u32 cycles)
1625 state->current_cycle = cycles;
1627 if (state->target_cycle < cycles)
1628 state->target_cycle = cycles;
1631 void lightrec_set_target_cycle_count(struct lightrec_state *state, u32 cycles)
1633 if (state->exit_flags == LIGHTREC_EXIT_NORMAL) {
1634 if (cycles < state->current_cycle)
1635 cycles = state->current_cycle;
1637 state->target_cycle = cycles;
1641 struct lightrec_registers * lightrec_get_registers(struct lightrec_state *state)
1643 return &state->regs;