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)
457 #if defined(__has_builtin) && __has_builtin(__builtin_clrsb)
458 return 1 + __builtin_clrsb(data);
462 data = (data ^ (data >> 31)) << 1;
473 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data)
477 state->regs.cp2d[12] = state->regs.cp2d[13];
478 state->regs.cp2d[13] = state->regs.cp2d[14];
479 state->regs.cp2d[14] = data;
482 state->regs.cp2d[9] = (data << 7) & 0xf80;
483 state->regs.cp2d[10] = (data << 2) & 0xf80;
484 state->regs.cp2d[11] = (data >> 3) & 0xf80;
489 state->regs.cp2d[31] = count_leading_bits((s32) data);
490 default: /* fall-through */
491 state->regs.cp2d[reg] = data;
496 static void lightrec_ctc2(struct lightrec_state *state, u8 reg, u32 data)
506 data = (s32)(s16) data;
509 data = (data & 0x7ffff000) | !!(data & 0x7f87e000) << 31;
510 default: /* fall-through */
514 state->regs.cp2c[reg] = data;
517 void lightrec_mtc(struct lightrec_state *state, union code op, u32 data)
519 if (op.i.op == OP_CP0)
520 lightrec_mtc0(state, op.r.rd, data);
521 else if (op.r.rs == OP_CP2_BASIC_CTC2)
522 lightrec_ctc2(state, op.r.rd, data);
524 lightrec_mtc2(state, op.r.rd, data);
527 static void lightrec_mtc_cb(struct lightrec_state *state, union code op)
529 lightrec_mtc(state, op, state->regs.gpr[op.r.rt]);
532 void lightrec_rfe(struct lightrec_state *state)
536 /* Read CP0 Status register (r12) */
537 status = state->regs.cp0[12];
539 /* Switch the bits */
540 status = ((status & 0x3c) >> 2) | (status & ~0xf);
543 lightrec_mtc0(state, 12, status);
546 void lightrec_cp(struct lightrec_state *state, union code op)
548 if (op.i.op == OP_CP0) {
549 pr_err("Invalid CP opcode to coprocessor #0\n");
553 (*state->ops.cop2_op)(state, op.opcode);
556 static void lightrec_syscall_cb(struct lightrec_state *state, union code op)
558 lightrec_set_exit_flags(state, LIGHTREC_EXIT_SYSCALL);
561 static void lightrec_break_cb(struct lightrec_state *state, union code op)
563 lightrec_set_exit_flags(state, LIGHTREC_EXIT_BREAK);
566 struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
568 struct block *block = lightrec_find_block(state->block_cache, pc);
570 if (block && lightrec_block_is_outdated(state, block)) {
571 pr_debug("Block at PC 0x%08x is outdated!\n", block->pc);
573 /* Make sure the recompiler isn't processing the block we'll
575 if (ENABLE_THREADED_COMPILER)
576 lightrec_recompiler_remove(state->rec, block);
578 lightrec_unregister_block(state->block_cache, block);
579 remove_from_code_lut(state->block_cache, block);
580 lightrec_free_block(state, block);
585 block = lightrec_precompile_block(state, pc);
587 pr_err("Unable to recompile block at PC 0x%x\n", pc);
588 lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
592 lightrec_register_block(state->block_cache, block);
598 static void * get_next_block_func(struct lightrec_state *state, u32 pc)
601 bool should_recompile;
605 func = state->code_lut[lut_offset(pc)];
606 if (func && func != state->get_next_block)
609 block = lightrec_get_block(state, pc);
611 if (unlikely(!block))
614 if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET)) {
615 func = state->memset_func;
619 should_recompile = block->flags & BLOCK_SHOULD_RECOMPILE &&
620 !(block->flags & BLOCK_IS_DEAD);
622 if (unlikely(should_recompile)) {
623 pr_debug("Block at PC 0x%08x should recompile\n", pc);
625 lightrec_unregister(MEM_FOR_CODE, block->code_size);
627 if (ENABLE_THREADED_COMPILER)
628 lightrec_recompiler_add(state->rec, block);
630 lightrec_compile_block(state->cstate, block);
633 if (ENABLE_THREADED_COMPILER && likely(!should_recompile))
634 func = lightrec_recompiler_run_first_pass(state, block, &pc);
636 func = block->function;
641 if (unlikely(block->flags & BLOCK_NEVER_COMPILE)) {
642 pc = lightrec_emulate_block(state, block, pc);
644 } else if (!ENABLE_THREADED_COMPILER) {
645 /* Block wasn't compiled yet - run the interpreter */
646 if (block->flags & BLOCK_FULLY_TAGGED)
647 pr_debug("Block fully tagged, skipping first pass\n");
648 else if (ENABLE_FIRST_PASS && likely(!should_recompile))
649 pc = lightrec_emulate_block(state, block, pc);
651 /* Then compile it using the profiled data */
652 lightrec_compile_block(state->cstate, block);
654 lightrec_recompiler_add(state->rec, block);
657 if (state->exit_flags != LIGHTREC_EXIT_NORMAL ||
658 state->current_cycle >= state->target_cycle)
666 static s32 c_function_wrapper(struct lightrec_state *state, s32 cycles_delta,
667 void (*f)(struct lightrec_state *, u32 d),
670 state->current_cycle = state->target_cycle - cycles_delta;
674 return state->target_cycle - state->current_cycle;
677 static struct block * generate_wrapper(struct lightrec_state *state)
683 jit_word_t code_size;
684 jit_node_t *to_tramp, *to_fn_epilog;
686 block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
690 _jit = jit_new_state();
694 jit_name("RW wrapper");
695 jit_note(__FILE__, __LINE__);
697 /* Wrapper entry point */
700 stack_ptr = jit_allocai(sizeof(uintptr_t) * NUM_TEMPS);
702 for (i = 0; i < NUM_TEMPS; i++)
703 jit_stxi(stack_ptr + i * sizeof(uintptr_t), JIT_FP, JIT_R(i));
705 /* Jump to the trampoline */
706 to_tramp = jit_jmpi();
708 /* The trampoline will jump back here */
709 to_fn_epilog = jit_label();
711 for (i = 0; i < NUM_TEMPS; i++)
712 jit_ldxi(JIT_R(i), JIT_FP, stack_ptr + i * sizeof(uintptr_t));
717 /* Trampoline entry point.
718 * The sole purpose of the trampoline is to cheese Lightning not to
719 * save/restore the callee-saved register LIGHTREC_REG_CYCLE, since we
720 * do want to return to the caller with this register modified. */
726 jit_pushargr(LIGHTREC_REG_STATE);
727 jit_pushargr(LIGHTREC_REG_CYCLE);
728 jit_pushargr(JIT_R0);
729 jit_pushargr(JIT_R1);
730 jit_finishi(c_function_wrapper);
731 jit_retval_i(LIGHTREC_REG_CYCLE);
733 jit_patch_at(jit_jmpi(), to_fn_epilog);
737 block->function = jit_emit();
738 block->opcode_list = NULL;
742 jit_get_code(&code_size);
743 lightrec_register(MEM_FOR_CODE, code_size);
745 block->code_size = code_size;
747 if (ENABLE_DISASSEMBLER) {
748 pr_debug("Wrapper block:\n");
756 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
758 pr_err("Unable to compile wrapper: Out of memory\n");
762 static u32 lightrec_memset(struct lightrec_state *state)
764 u32 kunseg_pc = kunseg(state->regs.gpr[4]);
766 const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg_pc);
767 u32 length = state->regs.gpr[5] * 4;
770 pr_err("Unable to find memory map for memset target address "
771 "0x%x\n", kunseg_pc);
775 pr_debug("Calling host memset, PC 0x%x (host address 0x%" PRIxPTR ") for %u bytes\n",
776 kunseg_pc, (uintptr_t)host, length);
777 memset(host, 0, length);
779 if (!state->invalidate_from_dma_only)
780 lightrec_invalidate_map(state, map, kunseg_pc, length);
782 /* Rough estimation of the number of cycles consumed */
783 return 8 + 5 * (length + 3 / 4);
786 static struct block * generate_dispatcher(struct lightrec_state *state)
790 jit_node_t *to_end, *to_c, *loop, *addr, *addr2, *addr3;
793 jit_word_t code_size;
795 block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
799 _jit = jit_new_state();
803 jit_name("dispatcher");
804 jit_note(__FILE__, __LINE__);
809 jit_getarg(JIT_R0, jit_arg());
810 jit_getarg_i(LIGHTREC_REG_CYCLE, jit_arg());
812 /* Force all callee-saved registers to be pushed on the stack */
813 for (i = 0; i < NUM_REGS; i++)
814 jit_movr(JIT_V(i), JIT_V(i));
816 /* Pass lightrec_state structure to blocks, using the last callee-saved
817 * register that Lightning provides */
818 jit_movi(LIGHTREC_REG_STATE, (intptr_t) state);
822 /* Call the block's code */
825 if (OPT_REPLACE_MEMSET) {
826 /* Blocks will jump here when they need to call
827 * lightrec_memset() */
828 addr3 = jit_indirect();
831 jit_pushargr(LIGHTREC_REG_STATE);
832 jit_finishi(lightrec_memset);
834 jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
835 offsetof(struct lightrec_state, regs.gpr[31]));
838 jit_subr(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, JIT_R0);
841 /* The block will jump here, with the number of cycles remaining in
842 * LIGHTREC_REG_CYCLE */
843 addr2 = jit_indirect();
845 /* Store back the next_pc to the lightrec_state structure */
846 offset = offsetof(struct lightrec_state, next_pc);
847 jit_stxi_i(offset, LIGHTREC_REG_STATE, JIT_V0);
849 /* Jump to end if state->target_cycle < state->current_cycle */
850 to_end = jit_blei(LIGHTREC_REG_CYCLE, 0);
852 /* Convert next PC to KUNSEG and avoid mirrors */
853 ram_len = state->maps[PSX_MAP_KERNEL_USER_RAM].length;
854 jit_andi(JIT_R0, JIT_V0, 0x10000000 | (ram_len - 1));
855 to_c = jit_bgei(JIT_R0, ram_len);
857 /* Fast path: code is running from RAM, use the code LUT */
858 if (__WORDSIZE == 64)
859 jit_lshi(JIT_R0, JIT_R0, 1);
860 jit_addr(JIT_R0, JIT_R0, LIGHTREC_REG_STATE);
861 jit_ldxi(JIT_R0, JIT_R0, offsetof(struct lightrec_state, code_lut));
863 /* If we get non-NULL, loop */
864 jit_patch_at(jit_bnei(JIT_R0, 0), loop);
866 /* Slow path: call C function get_next_block_func() */
869 if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
870 /* We may call the interpreter - update state->current_cycle */
871 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
872 offsetof(struct lightrec_state, target_cycle));
873 jit_subr(JIT_R1, JIT_R2, LIGHTREC_REG_CYCLE);
874 jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
875 LIGHTREC_REG_STATE, JIT_R1);
878 /* The code LUT will be set to this address when the block at the target
879 * PC has been preprocessed but not yet compiled by the threaded
881 addr = jit_indirect();
883 /* Get the next block */
885 jit_pushargr(LIGHTREC_REG_STATE);
886 jit_pushargr(JIT_V0);
887 jit_finishi(&get_next_block_func);
890 if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
891 /* The interpreter may have updated state->current_cycle and
892 * state->target_cycle - recalc the delta */
893 jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
894 offsetof(struct lightrec_state, current_cycle));
895 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
896 offsetof(struct lightrec_state, target_cycle));
897 jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
900 /* If we get non-NULL, loop */
901 jit_patch_at(jit_bnei(JIT_R0, 0), loop);
903 /* When exiting, the recompiled code will jump to that address */
904 jit_note(__FILE__, __LINE__);
907 jit_retr(LIGHTREC_REG_CYCLE);
911 block->function = jit_emit();
912 block->opcode_list = NULL;
916 jit_get_code(&code_size);
917 lightrec_register(MEM_FOR_CODE, code_size);
919 block->code_size = code_size;
921 state->eob_wrapper_func = jit_address(addr2);
922 if (OPT_REPLACE_MEMSET)
923 state->memset_func = jit_address(addr3);
924 state->get_next_block = jit_address(addr);
926 if (ENABLE_DISASSEMBLER) {
927 pr_debug("Dispatcher block:\n");
936 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
938 pr_err("Unable to compile dispatcher: Out of memory\n");
942 union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
946 lightrec_get_map(state, &host, kunseg(pc));
948 const u32 *code = (u32 *)host;
949 return (union code) *code;
952 unsigned int lightrec_cycles_of_opcode(union code code)
957 void lightrec_free_opcode_list(struct lightrec_state *state, struct block *block)
959 lightrec_free(state, MEM_FOR_IR,
960 sizeof(*block->opcode_list) * block->nb_ops,
964 static unsigned int lightrec_get_mips_block_len(const u32 *src)
970 c.opcode = LE32TOH(*src++);
975 if (is_unconditional_jump(c))
980 static struct opcode * lightrec_disassemble(struct lightrec_state *state,
981 const u32 *src, unsigned int *len)
984 unsigned int i, length;
986 length = lightrec_get_mips_block_len(src);
988 list = lightrec_malloc(state, MEM_FOR_IR, sizeof(*list) * length);
990 pr_err("Unable to allocate memory\n");
994 for (i = 0; i < length; i++) {
995 list[i].opcode = LE32TOH(src[i]);
999 *len = length * sizeof(u32);
1004 static struct block * lightrec_precompile_block(struct lightrec_state *state,
1007 struct opcode *list;
1008 struct block *block;
1010 const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg(pc));
1011 const u32 *code = (u32 *) host;
1012 unsigned int length;
1018 block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1020 pr_err("Unable to recompile block: Out of memory\n");
1024 list = lightrec_disassemble(state, code, &length);
1026 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1032 block->function = NULL;
1033 block->opcode_list = list;
1037 block->code_size = 0;
1038 #if ENABLE_THREADED_COMPILER
1039 block->op_list_freed = (atomic_flag)ATOMIC_FLAG_INIT;
1041 block->nb_ops = length / sizeof(u32);
1043 lightrec_optimize(state, block);
1045 length = block->nb_ops * sizeof(u32);
1047 lightrec_register(MEM_FOR_MIPS_CODE, length);
1049 if (ENABLE_DISASSEMBLER) {
1050 pr_debug("Disassembled block at PC: 0x%08x\n", block->pc);
1051 lightrec_print_disassembly(block, code);
1054 pr_debug("Block size: %hu opcodes\n", block->nb_ops);
1056 /* If the first opcode is an 'impossible' branch, never compile the
1058 if (should_emulate(block->opcode_list))
1059 block->flags |= BLOCK_NEVER_COMPILE;
1061 fully_tagged = lightrec_block_is_fully_tagged(block);
1063 block->flags |= BLOCK_FULLY_TAGGED;
1065 if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET))
1066 state->code_lut[lut_offset(pc)] = state->memset_func;
1068 block->hash = lightrec_calculate_block_hash(block);
1070 pr_debug("Recompile count: %u\n", state->nb_precompile++);
1075 static bool lightrec_block_is_fully_tagged(const struct block *block)
1077 const struct opcode *op;
1080 for (i = 0; i < block->nb_ops; i++) {
1081 op = &block->opcode_list[i];
1083 /* Verify that all load/stores of the opcode list
1084 * Check all loads/stores of the opcode list and mark the
1085 * block as fully compiled if they all have been tagged. */
1086 switch (op->c.i.op) {
1101 if (!(op->flags & (LIGHTREC_DIRECT_IO |
1104 default: /* fall-through */
1112 static void lightrec_reap_block(struct lightrec_state *state, void *data)
1114 struct block *block = data;
1116 pr_debug("Reap dead block at PC 0x%08x\n", block->pc);
1117 lightrec_unregister_block(state->block_cache, block);
1118 lightrec_free_block(state, block);
1121 static void lightrec_reap_jit(struct lightrec_state *state, void *data)
1123 _jit_destroy_state(data);
1126 int lightrec_compile_block(struct lightrec_cstate *cstate,
1127 struct block *block)
1129 struct lightrec_state *state = cstate->state;
1130 struct lightrec_branch_target *target;
1131 bool op_list_freed = false, fully_tagged = false;
1132 struct block *block2;
1134 jit_state_t *_jit, *oldjit;
1135 jit_node_t *start_of_block;
1136 bool skip_next = false;
1137 jit_word_t code_size;
1141 fully_tagged = lightrec_block_is_fully_tagged(block);
1143 block->flags |= BLOCK_FULLY_TAGGED;
1145 _jit = jit_new_state();
1149 oldjit = block->_jit;
1152 lightrec_regcache_reset(cstate->reg_cache);
1154 cstate->nb_branches = 0;
1155 cstate->nb_local_branches = 0;
1156 cstate->nb_targets = 0;
1161 start_of_block = jit_label();
1163 for (i = 0; i < block->nb_ops; i++) {
1164 elm = &block->opcode_list[i];
1171 cstate->cycles += lightrec_cycles_of_opcode(elm->c);
1173 if (should_emulate(elm)) {
1174 pr_debug("Branch at offset 0x%x will be emulated\n",
1177 lightrec_emit_eob(cstate, block, i, false);
1178 skip_next = !(elm->flags & LIGHTREC_NO_DS);
1180 lightrec_rec_opcode(cstate, block, i);
1181 skip_next = has_delay_slot(elm->c) &&
1182 !(elm->flags & LIGHTREC_NO_DS);
1184 /* FIXME: GNU Lightning on Windows seems to use our
1185 * mapped registers as temporaries. Until the actual bug
1186 * is found and fixed, unconditionally mark our
1187 * registers as live here. */
1188 lightrec_regcache_mark_live(cstate->reg_cache, _jit);
1193 for (i = 0; i < cstate->nb_branches; i++)
1194 jit_patch(cstate->branches[i]);
1196 for (i = 0; i < cstate->nb_local_branches; i++) {
1197 struct lightrec_branch *branch = &cstate->local_branches[i];
1199 pr_debug("Patch local branch to offset 0x%x\n",
1200 branch->target << 2);
1202 if (branch->target == 0) {
1203 jit_patch_at(branch->branch, start_of_block);
1207 for (j = 0; j < cstate->nb_targets; j++) {
1208 if (cstate->targets[j].offset == branch->target) {
1209 jit_patch_at(branch->branch,
1210 cstate->targets[j].label);
1215 if (j == cstate->nb_targets)
1216 pr_err("Unable to find branch target\n");
1219 jit_ldxi(JIT_R0, LIGHTREC_REG_STATE,
1220 offsetof(struct lightrec_state, eob_wrapper_func));
1227 block->function = jit_emit();
1228 block->flags &= ~BLOCK_SHOULD_RECOMPILE;
1230 /* Add compiled function to the LUT */
1231 state->code_lut[lut_offset(block->pc)] = block->function;
1233 if (ENABLE_THREADED_COMPILER) {
1234 /* Since we might try to reap the same block multiple times,
1235 * we need the reaper to wait until everything has been
1236 * submitted, so that the duplicate entries can be dropped. */
1237 lightrec_reaper_pause(state->reaper);
1240 /* Detect old blocks that have been covered by the new one */
1241 for (i = 0; i < cstate->nb_targets; i++) {
1242 target = &cstate->targets[i];
1244 if (!target->offset)
1247 offset = block->pc + target->offset * sizeof(u32);
1248 block2 = lightrec_find_block(state->block_cache, offset);
1250 /* No need to check if block2 is compilable - it must
1251 * be, otherwise block wouldn't be compilable either */
1253 /* Set the "block dead" flag to prevent the dynarec from
1254 * recompiling this block */
1255 block2->flags |= BLOCK_IS_DEAD;
1257 /* If block2 was pending for compilation, cancel it.
1258 * If it's being compiled right now, wait until it
1260 if (ENABLE_THREADED_COMPILER)
1261 lightrec_recompiler_remove(state->rec, block2);
1263 /* We know from now on that block2 isn't going to be
1264 * compiled. We can override the LUT entry with our
1265 * new block's entry point. */
1266 offset = lut_offset(block->pc) + target->offset;
1267 state->code_lut[offset] = jit_address(target->label);
1269 pr_debug("Reap block 0x%08x as it's covered by block "
1270 "0x%08x\n", block2->pc, block->pc);
1272 /* Finally, reap the block. */
1273 if (ENABLE_THREADED_COMPILER) {
1274 lightrec_reaper_add(state->reaper,
1275 lightrec_reap_block,
1278 lightrec_unregister_block(state->block_cache, block2);
1279 lightrec_free_block(state, block2);
1284 if (ENABLE_DISASSEMBLER)
1285 lightrec_reaper_continue(state->reaper);
1287 jit_get_code(&code_size);
1288 lightrec_register(MEM_FOR_CODE, code_size);
1290 block->code_size = code_size;
1292 if (ENABLE_DISASSEMBLER) {
1293 pr_debug("Compiling block at PC: 0x%08x\n", block->pc);
1299 #if ENABLE_THREADED_COMPILER
1301 op_list_freed = atomic_flag_test_and_set(&block->op_list_freed);
1303 if (fully_tagged && !op_list_freed) {
1304 pr_debug("Block PC 0x%08x is fully tagged"
1305 " - free opcode list\n", block->pc);
1306 lightrec_free_opcode_list(state, block);
1307 block->opcode_list = NULL;
1311 pr_debug("Block 0x%08x recompiled, reaping old jit context.\n",
1314 if (ENABLE_THREADED_COMPILER)
1315 lightrec_reaper_add(state->reaper,
1316 lightrec_reap_jit, oldjit);
1318 _jit_destroy_state(oldjit);
1324 static void lightrec_print_info(struct lightrec_state *state)
1326 if ((state->current_cycle & ~0xfffffff) != state->old_cycle_counter) {
1327 pr_info("Lightrec RAM usage: IR %u KiB, CODE %u KiB, "
1328 "MIPS %u KiB, TOTAL %u KiB, avg. IPI %f\n",
1329 lightrec_get_mem_usage(MEM_FOR_IR) / 1024,
1330 lightrec_get_mem_usage(MEM_FOR_CODE) / 1024,
1331 lightrec_get_mem_usage(MEM_FOR_MIPS_CODE) / 1024,
1332 lightrec_get_total_mem_usage() / 1024,
1333 lightrec_get_average_ipi());
1334 state->old_cycle_counter = state->current_cycle & ~0xfffffff;
1338 u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
1340 s32 (*func)(void *, s32) = (void *)state->dispatcher->function;
1344 state->exit_flags = LIGHTREC_EXIT_NORMAL;
1346 /* Handle the cycle counter overflowing */
1347 if (unlikely(target_cycle < state->current_cycle))
1348 target_cycle = UINT_MAX;
1350 state->target_cycle = target_cycle;
1351 state->next_pc = pc;
1353 block_trace = get_next_block_func(state, pc);
1355 cycles_delta = state->target_cycle - state->current_cycle;
1357 cycles_delta = (*func)(block_trace, cycles_delta);
1359 state->current_cycle = state->target_cycle - cycles_delta;
1362 if (ENABLE_THREADED_COMPILER)
1363 lightrec_reaper_reap(state->reaper);
1365 if (LOG_LEVEL >= INFO_L)
1366 lightrec_print_info(state);
1368 return state->next_pc;
1371 u32 lightrec_execute_one(struct lightrec_state *state, u32 pc)
1373 return lightrec_execute(state, pc, state->current_cycle);
1376 u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc)
1378 struct block *block = lightrec_get_block(state, pc);
1382 state->exit_flags = LIGHTREC_EXIT_NORMAL;
1384 pc = lightrec_emulate_block(state, block, pc);
1386 if (LOG_LEVEL >= INFO_L)
1387 lightrec_print_info(state);
1392 void lightrec_free_block(struct lightrec_state *state, struct block *block)
1394 lightrec_unregister(MEM_FOR_MIPS_CODE, block->nb_ops * sizeof(u32));
1395 if (block->opcode_list)
1396 lightrec_free_opcode_list(state, block);
1398 _jit_destroy_state(block->_jit);
1399 lightrec_unregister(MEM_FOR_CODE, block->code_size);
1400 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1403 struct lightrec_cstate * lightrec_create_cstate(struct lightrec_state *state)
1405 struct lightrec_cstate *cstate;
1407 cstate = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*cstate));
1411 cstate->reg_cache = lightrec_regcache_init(state);
1412 if (!cstate->reg_cache) {
1413 lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1417 cstate->state = state;
1422 void lightrec_free_cstate(struct lightrec_cstate *cstate)
1424 lightrec_free_regcache(cstate->reg_cache);
1425 lightrec_free(cstate->state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1428 struct lightrec_state * lightrec_init(char *argv0,
1429 const struct lightrec_mem_map *map,
1431 const struct lightrec_ops *ops)
1433 struct lightrec_state *state;
1435 /* Sanity-check ops */
1436 if (!ops || !ops->cop2_op || !ops->enable_ram) {
1437 pr_err("Missing callbacks in lightrec_ops structure\n");
1443 state = calloc(1, sizeof(*state) +
1444 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1446 goto err_finish_jit;
1448 lightrec_register(MEM_FOR_LIGHTREC, sizeof(*state) +
1449 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1452 state->tinymm = tinymm_init(malloc, free, 4096);
1454 goto err_free_state;
1457 state->block_cache = lightrec_blockcache_init(state);
1458 if (!state->block_cache)
1459 goto err_free_tinymm;
1461 if (ENABLE_THREADED_COMPILER) {
1462 state->rec = lightrec_recompiler_init(state);
1464 goto err_free_block_cache;
1466 state->reaper = lightrec_reaper_init(state);
1468 goto err_free_recompiler;
1470 state->cstate = lightrec_create_cstate(state);
1472 goto err_free_block_cache;
1475 state->nb_maps = nb;
1478 memcpy(&state->ops, ops, sizeof(*ops));
1480 state->dispatcher = generate_dispatcher(state);
1481 if (!state->dispatcher)
1482 goto err_free_reaper;
1484 state->c_wrapper_block = generate_wrapper(state);
1485 if (!state->c_wrapper_block)
1486 goto err_free_dispatcher;
1488 state->c_wrapper = state->c_wrapper_block->function;
1490 state->c_wrappers[C_WRAPPER_RW] = lightrec_rw_cb;
1491 state->c_wrappers[C_WRAPPER_RW_GENERIC] = lightrec_rw_generic_cb;
1492 state->c_wrappers[C_WRAPPER_MFC] = lightrec_mfc_cb;
1493 state->c_wrappers[C_WRAPPER_MTC] = lightrec_mtc_cb;
1494 state->c_wrappers[C_WRAPPER_CP] = lightrec_cp;
1495 state->c_wrappers[C_WRAPPER_SYSCALL] = lightrec_syscall_cb;
1496 state->c_wrappers[C_WRAPPER_BREAK] = lightrec_break_cb;
1498 map = &state->maps[PSX_MAP_BIOS];
1499 state->offset_bios = (uintptr_t)map->address - map->pc;
1501 map = &state->maps[PSX_MAP_SCRATCH_PAD];
1502 state->offset_scratch = (uintptr_t)map->address - map->pc;
1504 map = &state->maps[PSX_MAP_KERNEL_USER_RAM];
1505 state->offset_ram = (uintptr_t)map->address - map->pc;
1507 if (state->maps[PSX_MAP_MIRROR1].address == map->address + 0x200000 &&
1508 state->maps[PSX_MAP_MIRROR2].address == map->address + 0x400000 &&
1509 state->maps[PSX_MAP_MIRROR3].address == map->address + 0x600000)
1510 state->mirrors_mapped = true;
1512 if (state->offset_bios == 0 &&
1513 state->offset_scratch == 0 &&
1514 state->offset_ram == 0 &&
1515 state->mirrors_mapped) {
1516 pr_info("Memory map is perfect. Emitted code will be best.\n");
1518 pr_info("Memory map is sub-par. Emitted code will be slow.\n");
1523 err_free_dispatcher:
1524 lightrec_free_block(state, state->dispatcher);
1526 if (ENABLE_THREADED_COMPILER)
1527 lightrec_reaper_destroy(state->reaper);
1528 err_free_recompiler:
1529 if (ENABLE_THREADED_COMPILER)
1530 lightrec_free_recompiler(state->rec);
1532 lightrec_free_cstate(state->cstate);
1533 err_free_block_cache:
1534 lightrec_free_block_cache(state->block_cache);
1537 tinymm_shutdown(state->tinymm);
1540 lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1541 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1548 void lightrec_destroy(struct lightrec_state *state)
1550 /* Force a print info on destroy*/
1551 state->current_cycle = ~state->current_cycle;
1552 lightrec_print_info(state);
1554 if (ENABLE_THREADED_COMPILER) {
1555 lightrec_free_recompiler(state->rec);
1556 lightrec_reaper_destroy(state->reaper);
1558 lightrec_free_cstate(state->cstate);
1561 lightrec_free_block_cache(state->block_cache);
1562 lightrec_free_block(state, state->dispatcher);
1563 lightrec_free_block(state, state->c_wrapper_block);
1567 tinymm_shutdown(state->tinymm);
1569 lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1570 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1574 void lightrec_invalidate(struct lightrec_state *state, u32 addr, u32 len)
1576 u32 kaddr = kunseg(addr & ~0x3);
1577 const struct lightrec_mem_map *map = lightrec_get_map(state, NULL, kaddr);
1580 if (map != &state->maps[PSX_MAP_KERNEL_USER_RAM])
1583 /* Handle mirrors */
1584 kaddr &= (state->maps[PSX_MAP_KERNEL_USER_RAM].length - 1);
1586 lightrec_invalidate_map(state, map, kaddr, len);
1590 void lightrec_invalidate_all(struct lightrec_state *state)
1592 memset(state->code_lut, 0, sizeof(*state->code_lut) * CODE_LUT_SIZE);
1595 void lightrec_set_invalidate_mode(struct lightrec_state *state, bool dma_only)
1597 if (state->invalidate_from_dma_only != dma_only)
1598 lightrec_invalidate_all(state);
1600 state->invalidate_from_dma_only = dma_only;
1603 void lightrec_set_exit_flags(struct lightrec_state *state, u32 flags)
1605 if (flags != LIGHTREC_EXIT_NORMAL) {
1606 state->exit_flags |= flags;
1607 state->target_cycle = state->current_cycle;
1611 u32 lightrec_exit_flags(struct lightrec_state *state)
1613 return state->exit_flags;
1616 u32 lightrec_current_cycle_count(const struct lightrec_state *state)
1618 return state->current_cycle;
1621 void lightrec_reset_cycle_count(struct lightrec_state *state, u32 cycles)
1623 state->current_cycle = cycles;
1625 if (state->target_cycle < cycles)
1626 state->target_cycle = cycles;
1629 void lightrec_set_target_cycle_count(struct lightrec_state *state, u32 cycles)
1631 if (state->exit_flags == LIGHTREC_EXIT_NORMAL) {
1632 if (cycles < state->current_cycle)
1633 cycles = state->current_cycle;
1635 state->target_cycle = cycles;
1639 struct lightrec_registers * lightrec_get_registers(struct lightrec_state *state)
1641 return &state->regs;