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_mtc0(struct lightrec_state *state, u8 reg, u32 data)
411 u32 status, oldstatus, cause;
419 /* Those registers are read-only */
426 status = state->regs.cp0[12];
429 if (status & ~data & BIT(16)) {
430 state->ops.enable_ram(state, true);
431 lightrec_invalidate_all(state);
432 } else if (~status & data & BIT(16)) {
433 state->ops.enable_ram(state, false);
438 state->regs.cp0[13] &= ~0x300;
439 state->regs.cp0[13] |= data & 0x300;
441 state->regs.cp0[reg] = data;
444 if (reg == 12 || reg == 13) {
445 cause = state->regs.cp0[13];
446 status = state->regs.cp0[12];
448 /* Handle software interrupts */
449 if (!!(status & cause & 0x300) & status)
450 lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
452 /* Handle hardware interrupts */
453 if (reg == 12 && !(~status & 0x401) && (~oldstatus & 0x401))
454 lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
458 static u32 count_leading_bits(s32 data)
463 #if __has_builtin(__builtin_clrsb)
464 return 1 + __builtin_clrsb(data);
468 data = (data ^ (data >> 31)) << 1;
478 static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data)
482 state->regs.cp2d[12] = state->regs.cp2d[13];
483 state->regs.cp2d[13] = state->regs.cp2d[14];
484 state->regs.cp2d[14] = data;
487 state->regs.cp2d[9] = (data << 7) & 0xf80;
488 state->regs.cp2d[10] = (data << 2) & 0xf80;
489 state->regs.cp2d[11] = (data >> 3) & 0xf80;
494 state->regs.cp2d[31] = count_leading_bits((s32) data);
495 default: /* fall-through */
496 state->regs.cp2d[reg] = data;
501 static void lightrec_ctc2(struct lightrec_state *state, u8 reg, u32 data)
511 data = (s32)(s16) data;
514 data = (data & 0x7ffff000) | !!(data & 0x7f87e000) << 31;
515 default: /* fall-through */
519 state->regs.cp2c[reg] = data;
522 void lightrec_mtc(struct lightrec_state *state, union code op, u32 data)
524 if (op.i.op == OP_CP0)
525 lightrec_mtc0(state, op.r.rd, data);
526 else if (op.r.rs == OP_CP2_BASIC_CTC2)
527 lightrec_ctc2(state, op.r.rd, data);
529 lightrec_mtc2(state, op.r.rd, data);
532 static void lightrec_mtc_cb(struct lightrec_state *state, union code op)
534 lightrec_mtc(state, op, state->regs.gpr[op.r.rt]);
537 void lightrec_rfe(struct lightrec_state *state)
541 /* Read CP0 Status register (r12) */
542 status = state->regs.cp0[12];
544 /* Switch the bits */
545 status = ((status & 0x3c) >> 2) | (status & ~0xf);
548 lightrec_mtc0(state, 12, status);
551 void lightrec_cp(struct lightrec_state *state, union code op)
553 if (op.i.op == OP_CP0) {
554 pr_err("Invalid CP opcode to coprocessor #0\n");
558 (*state->ops.cop2_op)(state, op.opcode);
561 static void lightrec_syscall_cb(struct lightrec_state *state, union code op)
563 lightrec_set_exit_flags(state, LIGHTREC_EXIT_SYSCALL);
566 static void lightrec_break_cb(struct lightrec_state *state, union code op)
568 lightrec_set_exit_flags(state, LIGHTREC_EXIT_BREAK);
571 struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
573 struct block *block = lightrec_find_block(state->block_cache, pc);
575 if (block && lightrec_block_is_outdated(state, block)) {
576 pr_debug("Block at PC 0x%08x is outdated!\n", block->pc);
578 /* Make sure the recompiler isn't processing the block we'll
580 if (ENABLE_THREADED_COMPILER)
581 lightrec_recompiler_remove(state->rec, block);
583 lightrec_unregister_block(state->block_cache, block);
584 remove_from_code_lut(state->block_cache, block);
585 lightrec_free_block(state, block);
590 block = lightrec_precompile_block(state, pc);
592 pr_err("Unable to recompile block at PC 0x%x\n", pc);
593 lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
597 lightrec_register_block(state->block_cache, block);
603 static void * get_next_block_func(struct lightrec_state *state, u32 pc)
606 bool should_recompile;
610 func = state->code_lut[lut_offset(pc)];
611 if (func && func != state->get_next_block)
614 block = lightrec_get_block(state, pc);
616 if (unlikely(!block))
619 if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET)) {
620 func = state->memset_func;
624 should_recompile = block->flags & BLOCK_SHOULD_RECOMPILE &&
625 !(block->flags & BLOCK_IS_DEAD);
627 if (unlikely(should_recompile)) {
628 pr_debug("Block at PC 0x%08x should recompile\n", pc);
630 lightrec_unregister(MEM_FOR_CODE, block->code_size);
632 if (ENABLE_THREADED_COMPILER)
633 lightrec_recompiler_add(state->rec, block);
635 lightrec_compile_block(state->cstate, block);
638 if (ENABLE_THREADED_COMPILER && likely(!should_recompile))
639 func = lightrec_recompiler_run_first_pass(state, block, &pc);
641 func = block->function;
646 if (unlikely(block->flags & BLOCK_NEVER_COMPILE)) {
647 pc = lightrec_emulate_block(state, block, pc);
649 } else if (!ENABLE_THREADED_COMPILER) {
650 /* Block wasn't compiled yet - run the interpreter */
651 if (block->flags & BLOCK_FULLY_TAGGED)
652 pr_debug("Block fully tagged, skipping first pass\n");
653 else if (ENABLE_FIRST_PASS && likely(!should_recompile))
654 pc = lightrec_emulate_block(state, block, pc);
656 /* Then compile it using the profiled data */
657 lightrec_compile_block(state->cstate, block);
659 lightrec_recompiler_add(state->rec, block);
662 if (state->exit_flags != LIGHTREC_EXIT_NORMAL ||
663 state->current_cycle >= state->target_cycle)
671 static s32 c_function_wrapper(struct lightrec_state *state, s32 cycles_delta,
672 void (*f)(struct lightrec_state *, u32 d),
675 state->current_cycle = state->target_cycle - cycles_delta;
679 return state->target_cycle - state->current_cycle;
682 static struct block * generate_wrapper(struct lightrec_state *state)
688 jit_word_t code_size;
689 jit_node_t *to_tramp, *to_fn_epilog;
690 jit_node_t *addr[C_WRAPPERS_COUNT - 1];
692 block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
696 _jit = jit_new_state();
700 jit_name("RW wrapper");
701 jit_note(__FILE__, __LINE__);
703 /* Wrapper entry point */
707 /* Add entry points; separate them by opcodes that increment
708 * LIGHTREC_REG_STATE (since we cannot touch other registers).
709 * The difference will then tell us which C function to call. */
710 for (i = C_WRAPPERS_COUNT - 1; i > 0; i--) {
711 jit_addi(LIGHTREC_REG_STATE, LIGHTREC_REG_STATE, __WORDSIZE / 8);
712 addr[i - 1] = jit_indirect();
718 stack_ptr = jit_allocai(sizeof(uintptr_t) * NUM_TEMPS);
720 /* Save all temporaries on stack */
721 for (i = 0; i < NUM_TEMPS; i++)
722 jit_stxi(stack_ptr + i * sizeof(uintptr_t), JIT_FP, JIT_R(i));
724 /* Jump to the trampoline */
725 to_tramp = jit_jmpi();
727 /* The trampoline will jump back here */
728 to_fn_epilog = jit_label();
730 /* Restore temporaries from stack */
731 for (i = 0; i < NUM_TEMPS; i++)
732 jit_ldxi(JIT_R(i), JIT_FP, stack_ptr + i * sizeof(uintptr_t));
737 /* Trampoline entry point.
738 * The sole purpose of the trampoline is to cheese Lightning not to
739 * save/restore the callee-saved register LIGHTREC_REG_CYCLE, since we
740 * do want to return to the caller with this register modified. */
745 /* Retrieve the wrapper function */
746 jit_ldxi(JIT_R0, LIGHTREC_REG_STATE,
747 offsetof(struct lightrec_state, c_wrappers));
749 /* Restore LIGHTREC_REG_STATE to its correct value */
750 jit_movi(LIGHTREC_REG_STATE, (uintptr_t) state);
753 jit_pushargr(LIGHTREC_REG_STATE);
754 jit_pushargr(LIGHTREC_REG_CYCLE);
755 jit_pushargr(JIT_R0);
756 jit_pushargr(JIT_R1);
757 jit_finishi(c_function_wrapper);
758 jit_retval_i(LIGHTREC_REG_CYCLE);
760 jit_patch_at(jit_jmpi(), to_fn_epilog);
764 block->function = jit_emit();
765 block->opcode_list = NULL;
769 state->wrappers_eps[C_WRAPPERS_COUNT - 1] = block->function;
771 for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
772 state->wrappers_eps[i] = jit_address(addr[i]);
774 jit_get_code(&code_size);
775 lightrec_register(MEM_FOR_CODE, code_size);
777 block->code_size = code_size;
779 if (ENABLE_DISASSEMBLER) {
780 pr_debug("Wrapper block:\n");
788 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
790 pr_err("Unable to compile wrapper: Out of memory\n");
794 static u32 lightrec_memset(struct lightrec_state *state)
796 u32 kunseg_pc = kunseg(state->regs.gpr[4]);
798 const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg_pc);
799 u32 length = state->regs.gpr[5] * 4;
802 pr_err("Unable to find memory map for memset target address "
803 "0x%x\n", kunseg_pc);
807 pr_debug("Calling host memset, PC 0x%x (host address 0x%" PRIxPTR ") for %u bytes\n",
808 kunseg_pc, (uintptr_t)host, length);
809 memset(host, 0, length);
811 if (!state->invalidate_from_dma_only)
812 lightrec_invalidate_map(state, map, kunseg_pc, length);
814 /* Rough estimation of the number of cycles consumed */
815 return 8 + 5 * (length + 3 / 4);
818 static struct block * generate_dispatcher(struct lightrec_state *state)
822 jit_node_t *to_end, *to_c, *loop, *addr, *addr2, *addr3;
825 jit_word_t code_size;
827 block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
831 _jit = jit_new_state();
835 jit_name("dispatcher");
836 jit_note(__FILE__, __LINE__);
841 jit_getarg(JIT_R0, jit_arg());
842 jit_getarg_i(LIGHTREC_REG_CYCLE, jit_arg());
844 /* Force all callee-saved registers to be pushed on the stack */
845 for (i = 0; i < NUM_REGS; i++)
846 jit_movr(JIT_V(i), JIT_V(i));
848 /* Pass lightrec_state structure to blocks, using the last callee-saved
849 * register that Lightning provides */
850 jit_movi(LIGHTREC_REG_STATE, (intptr_t) state);
854 /* Call the block's code */
857 if (OPT_REPLACE_MEMSET) {
858 /* Blocks will jump here when they need to call
859 * lightrec_memset() */
860 addr3 = jit_indirect();
863 jit_pushargr(LIGHTREC_REG_STATE);
864 jit_finishi(lightrec_memset);
866 jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
867 offsetof(struct lightrec_state, regs.gpr[31]));
870 jit_subr(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, JIT_R0);
873 /* The block will jump here, with the number of cycles remaining in
874 * LIGHTREC_REG_CYCLE */
875 addr2 = jit_indirect();
877 /* Store back the next_pc to the lightrec_state structure */
878 offset = offsetof(struct lightrec_state, next_pc);
879 jit_stxi_i(offset, LIGHTREC_REG_STATE, JIT_V0);
881 /* Jump to end if state->target_cycle < state->current_cycle */
882 to_end = jit_blei(LIGHTREC_REG_CYCLE, 0);
884 /* Convert next PC to KUNSEG and avoid mirrors */
885 ram_len = state->maps[PSX_MAP_KERNEL_USER_RAM].length;
886 jit_andi(JIT_R0, JIT_V0, 0x10000000 | (ram_len - 1));
887 to_c = jit_bgei(JIT_R0, ram_len);
889 /* Fast path: code is running from RAM, use the code LUT */
890 if (__WORDSIZE == 64)
891 jit_lshi(JIT_R0, JIT_R0, 1);
892 jit_addr(JIT_R0, JIT_R0, LIGHTREC_REG_STATE);
893 jit_ldxi(JIT_R0, JIT_R0, offsetof(struct lightrec_state, code_lut));
895 /* If we get non-NULL, loop */
896 jit_patch_at(jit_bnei(JIT_R0, 0), loop);
898 /* Slow path: call C function get_next_block_func() */
901 if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
902 /* We may call the interpreter - update state->current_cycle */
903 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
904 offsetof(struct lightrec_state, target_cycle));
905 jit_subr(JIT_R1, JIT_R2, LIGHTREC_REG_CYCLE);
906 jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
907 LIGHTREC_REG_STATE, JIT_R1);
910 /* The code LUT will be set to this address when the block at the target
911 * PC has been preprocessed but not yet compiled by the threaded
913 addr = jit_indirect();
915 /* Get the next block */
917 jit_pushargr(LIGHTREC_REG_STATE);
918 jit_pushargr(JIT_V0);
919 jit_finishi(&get_next_block_func);
922 if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
923 /* The interpreter may have updated state->current_cycle and
924 * state->target_cycle - recalc the delta */
925 jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
926 offsetof(struct lightrec_state, current_cycle));
927 jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
928 offsetof(struct lightrec_state, target_cycle));
929 jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
932 /* If we get non-NULL, loop */
933 jit_patch_at(jit_bnei(JIT_R0, 0), loop);
935 /* When exiting, the recompiled code will jump to that address */
936 jit_note(__FILE__, __LINE__);
939 jit_retr(LIGHTREC_REG_CYCLE);
943 block->function = jit_emit();
944 block->opcode_list = NULL;
948 jit_get_code(&code_size);
949 lightrec_register(MEM_FOR_CODE, code_size);
951 block->code_size = code_size;
953 state->eob_wrapper_func = jit_address(addr2);
954 if (OPT_REPLACE_MEMSET)
955 state->memset_func = jit_address(addr3);
956 state->get_next_block = jit_address(addr);
958 if (ENABLE_DISASSEMBLER) {
959 pr_debug("Dispatcher block:\n");
968 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
970 pr_err("Unable to compile dispatcher: Out of memory\n");
974 union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
978 lightrec_get_map(state, &host, kunseg(pc));
980 const u32 *code = (u32 *)host;
981 return (union code) *code;
984 unsigned int lightrec_cycles_of_opcode(union code code)
989 void lightrec_free_opcode_list(struct lightrec_state *state, struct block *block)
991 lightrec_free(state, MEM_FOR_IR,
992 sizeof(*block->opcode_list) * block->nb_ops,
996 static unsigned int lightrec_get_mips_block_len(const u32 *src)
1001 for (i = 1; ; i++) {
1002 c.opcode = LE32TOH(*src++);
1007 if (is_unconditional_jump(c))
1012 static struct opcode * lightrec_disassemble(struct lightrec_state *state,
1013 const u32 *src, unsigned int *len)
1015 struct opcode *list;
1016 unsigned int i, length;
1018 length = lightrec_get_mips_block_len(src);
1020 list = lightrec_malloc(state, MEM_FOR_IR, sizeof(*list) * length);
1022 pr_err("Unable to allocate memory\n");
1026 for (i = 0; i < length; i++) {
1027 list[i].opcode = LE32TOH(src[i]);
1031 *len = length * sizeof(u32);
1036 static struct block * lightrec_precompile_block(struct lightrec_state *state,
1039 struct opcode *list;
1040 struct block *block;
1042 const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg(pc));
1043 const u32 *code = (u32 *) host;
1044 unsigned int length;
1050 block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
1052 pr_err("Unable to recompile block: Out of memory\n");
1056 list = lightrec_disassemble(state, code, &length);
1058 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1064 block->function = NULL;
1065 block->opcode_list = list;
1069 block->code_size = 0;
1070 #if ENABLE_THREADED_COMPILER
1071 block->op_list_freed = (atomic_flag)ATOMIC_FLAG_INIT;
1073 block->nb_ops = length / sizeof(u32);
1075 lightrec_optimize(state, block);
1077 length = block->nb_ops * sizeof(u32);
1079 lightrec_register(MEM_FOR_MIPS_CODE, length);
1081 if (ENABLE_DISASSEMBLER) {
1082 pr_debug("Disassembled block at PC: 0x%08x\n", block->pc);
1083 lightrec_print_disassembly(block, code);
1086 pr_debug("Block size: %hu opcodes\n", block->nb_ops);
1088 /* If the first opcode is an 'impossible' branch, never compile the
1090 if (should_emulate(block->opcode_list))
1091 block->flags |= BLOCK_NEVER_COMPILE;
1093 fully_tagged = lightrec_block_is_fully_tagged(block);
1095 block->flags |= BLOCK_FULLY_TAGGED;
1097 if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET))
1098 state->code_lut[lut_offset(pc)] = state->memset_func;
1100 block->hash = lightrec_calculate_block_hash(block);
1102 pr_debug("Recompile count: %u\n", state->nb_precompile++);
1107 static bool lightrec_block_is_fully_tagged(const struct block *block)
1109 const struct opcode *op;
1112 for (i = 0; i < block->nb_ops; i++) {
1113 op = &block->opcode_list[i];
1115 /* Verify that all load/stores of the opcode list
1116 * Check all loads/stores of the opcode list and mark the
1117 * block as fully compiled if they all have been tagged. */
1118 switch (op->c.i.op) {
1133 if (!(op->flags & (LIGHTREC_DIRECT_IO |
1136 default: /* fall-through */
1144 static void lightrec_reap_block(struct lightrec_state *state, void *data)
1146 struct block *block = data;
1148 pr_debug("Reap dead block at PC 0x%08x\n", block->pc);
1149 lightrec_unregister_block(state->block_cache, block);
1150 lightrec_free_block(state, block);
1153 static void lightrec_reap_jit(struct lightrec_state *state, void *data)
1155 _jit_destroy_state(data);
1158 int lightrec_compile_block(struct lightrec_cstate *cstate,
1159 struct block *block)
1161 struct lightrec_state *state = cstate->state;
1162 struct lightrec_branch_target *target;
1163 bool op_list_freed = false, fully_tagged = false;
1164 struct block *block2;
1166 jit_state_t *_jit, *oldjit;
1167 jit_node_t *start_of_block;
1168 bool skip_next = false;
1169 jit_word_t code_size;
1173 fully_tagged = lightrec_block_is_fully_tagged(block);
1175 block->flags |= BLOCK_FULLY_TAGGED;
1177 _jit = jit_new_state();
1181 oldjit = block->_jit;
1184 lightrec_regcache_reset(cstate->reg_cache);
1186 cstate->nb_branches = 0;
1187 cstate->nb_local_branches = 0;
1188 cstate->nb_targets = 0;
1193 start_of_block = jit_label();
1195 for (i = 0; i < block->nb_ops; i++) {
1196 elm = &block->opcode_list[i];
1203 cstate->cycles += lightrec_cycles_of_opcode(elm->c);
1205 if (should_emulate(elm)) {
1206 pr_debug("Branch at offset 0x%x will be emulated\n",
1209 lightrec_emit_eob(cstate, block, i, false);
1210 skip_next = !(elm->flags & LIGHTREC_NO_DS);
1212 lightrec_rec_opcode(cstate, block, i);
1213 skip_next = has_delay_slot(elm->c) &&
1214 !(elm->flags & LIGHTREC_NO_DS);
1216 /* FIXME: GNU Lightning on Windows seems to use our
1217 * mapped registers as temporaries. Until the actual bug
1218 * is found and fixed, unconditionally mark our
1219 * registers as live here. */
1220 lightrec_regcache_mark_live(cstate->reg_cache, _jit);
1225 for (i = 0; i < cstate->nb_branches; i++)
1226 jit_patch(cstate->branches[i]);
1228 for (i = 0; i < cstate->nb_local_branches; i++) {
1229 struct lightrec_branch *branch = &cstate->local_branches[i];
1231 pr_debug("Patch local branch to offset 0x%x\n",
1232 branch->target << 2);
1234 if (branch->target == 0) {
1235 jit_patch_at(branch->branch, start_of_block);
1239 for (j = 0; j < cstate->nb_targets; j++) {
1240 if (cstate->targets[j].offset == branch->target) {
1241 jit_patch_at(branch->branch,
1242 cstate->targets[j].label);
1247 if (j == cstate->nb_targets)
1248 pr_err("Unable to find branch target\n");
1251 jit_ldxi(JIT_R0, LIGHTREC_REG_STATE,
1252 offsetof(struct lightrec_state, eob_wrapper_func));
1259 block->function = jit_emit();
1260 block->flags &= ~BLOCK_SHOULD_RECOMPILE;
1262 /* Add compiled function to the LUT */
1263 state->code_lut[lut_offset(block->pc)] = block->function;
1265 if (ENABLE_THREADED_COMPILER) {
1266 /* Since we might try to reap the same block multiple times,
1267 * we need the reaper to wait until everything has been
1268 * submitted, so that the duplicate entries can be dropped. */
1269 lightrec_reaper_pause(state->reaper);
1272 /* Detect old blocks that have been covered by the new one */
1273 for (i = 0; i < cstate->nb_targets; i++) {
1274 target = &cstate->targets[i];
1276 if (!target->offset)
1279 offset = block->pc + target->offset * sizeof(u32);
1280 block2 = lightrec_find_block(state->block_cache, offset);
1282 /* No need to check if block2 is compilable - it must
1283 * be, otherwise block wouldn't be compilable either */
1285 /* Set the "block dead" flag to prevent the dynarec from
1286 * recompiling this block */
1287 block2->flags |= BLOCK_IS_DEAD;
1289 /* If block2 was pending for compilation, cancel it.
1290 * If it's being compiled right now, wait until it
1292 if (ENABLE_THREADED_COMPILER)
1293 lightrec_recompiler_remove(state->rec, block2);
1296 /* We know from now on that block2 (if present) isn't going to
1297 * be compiled. We can override the LUT entry with our new
1298 * block's entry point. */
1299 offset = lut_offset(block->pc) + target->offset;
1300 state->code_lut[offset] = jit_address(target->label);
1303 pr_debug("Reap block 0x%08x as it's covered by block "
1304 "0x%08x\n", block2->pc, block->pc);
1306 /* Finally, reap the block. */
1307 if (ENABLE_THREADED_COMPILER) {
1308 lightrec_reaper_add(state->reaper,
1309 lightrec_reap_block,
1312 lightrec_unregister_block(state->block_cache, block2);
1313 lightrec_free_block(state, block2);
1318 if (ENABLE_THREADED_COMPILER)
1319 lightrec_reaper_continue(state->reaper);
1321 jit_get_code(&code_size);
1322 lightrec_register(MEM_FOR_CODE, code_size);
1324 block->code_size = code_size;
1326 if (ENABLE_DISASSEMBLER) {
1327 pr_debug("Compiling block at PC: 0x%08x\n", block->pc);
1333 #if ENABLE_THREADED_COMPILER
1335 op_list_freed = atomic_flag_test_and_set(&block->op_list_freed);
1337 if (fully_tagged && !op_list_freed) {
1338 pr_debug("Block PC 0x%08x is fully tagged"
1339 " - free opcode list\n", block->pc);
1340 lightrec_free_opcode_list(state, block);
1341 block->opcode_list = NULL;
1345 pr_debug("Block 0x%08x recompiled, reaping old jit context.\n",
1348 if (ENABLE_THREADED_COMPILER)
1349 lightrec_reaper_add(state->reaper,
1350 lightrec_reap_jit, oldjit);
1352 _jit_destroy_state(oldjit);
1358 static void lightrec_print_info(struct lightrec_state *state)
1360 if ((state->current_cycle & ~0xfffffff) != state->old_cycle_counter) {
1361 pr_info("Lightrec RAM usage: IR %u KiB, CODE %u KiB, "
1362 "MIPS %u KiB, TOTAL %u KiB, avg. IPI %f\n",
1363 lightrec_get_mem_usage(MEM_FOR_IR) / 1024,
1364 lightrec_get_mem_usage(MEM_FOR_CODE) / 1024,
1365 lightrec_get_mem_usage(MEM_FOR_MIPS_CODE) / 1024,
1366 lightrec_get_total_mem_usage() / 1024,
1367 lightrec_get_average_ipi());
1368 state->old_cycle_counter = state->current_cycle & ~0xfffffff;
1372 u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
1374 s32 (*func)(void *, s32) = (void *)state->dispatcher->function;
1378 state->exit_flags = LIGHTREC_EXIT_NORMAL;
1380 /* Handle the cycle counter overflowing */
1381 if (unlikely(target_cycle < state->current_cycle))
1382 target_cycle = UINT_MAX;
1384 state->target_cycle = target_cycle;
1385 state->next_pc = pc;
1387 block_trace = get_next_block_func(state, pc);
1389 cycles_delta = state->target_cycle - state->current_cycle;
1391 cycles_delta = (*func)(block_trace, cycles_delta);
1393 state->current_cycle = state->target_cycle - cycles_delta;
1396 if (ENABLE_THREADED_COMPILER)
1397 lightrec_reaper_reap(state->reaper);
1399 if (LOG_LEVEL >= INFO_L)
1400 lightrec_print_info(state);
1402 return state->next_pc;
1405 u32 lightrec_execute_one(struct lightrec_state *state, u32 pc)
1407 return lightrec_execute(state, pc, state->current_cycle);
1410 u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc)
1412 struct block *block = lightrec_get_block(state, pc);
1416 state->exit_flags = LIGHTREC_EXIT_NORMAL;
1418 pc = lightrec_emulate_block(state, block, pc);
1420 if (LOG_LEVEL >= INFO_L)
1421 lightrec_print_info(state);
1426 void lightrec_free_block(struct lightrec_state *state, struct block *block)
1428 lightrec_unregister(MEM_FOR_MIPS_CODE, block->nb_ops * sizeof(u32));
1429 if (block->opcode_list)
1430 lightrec_free_opcode_list(state, block);
1432 _jit_destroy_state(block->_jit);
1433 lightrec_unregister(MEM_FOR_CODE, block->code_size);
1434 lightrec_free(state, MEM_FOR_IR, sizeof(*block), block);
1437 struct lightrec_cstate * lightrec_create_cstate(struct lightrec_state *state)
1439 struct lightrec_cstate *cstate;
1441 cstate = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*cstate));
1445 cstate->reg_cache = lightrec_regcache_init(state);
1446 if (!cstate->reg_cache) {
1447 lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1451 cstate->state = state;
1456 void lightrec_free_cstate(struct lightrec_cstate *cstate)
1458 lightrec_free_regcache(cstate->reg_cache);
1459 lightrec_free(cstate->state, MEM_FOR_LIGHTREC, sizeof(*cstate), cstate);
1462 struct lightrec_state * lightrec_init(char *argv0,
1463 const struct lightrec_mem_map *map,
1465 const struct lightrec_ops *ops)
1467 struct lightrec_state *state;
1469 /* Sanity-check ops */
1470 if (!ops || !ops->cop2_op || !ops->enable_ram) {
1471 pr_err("Missing callbacks in lightrec_ops structure\n");
1477 state = calloc(1, sizeof(*state) +
1478 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1480 goto err_finish_jit;
1482 lightrec_register(MEM_FOR_LIGHTREC, sizeof(*state) +
1483 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1486 state->tinymm = tinymm_init(malloc, free, 4096);
1488 goto err_free_state;
1491 state->block_cache = lightrec_blockcache_init(state);
1492 if (!state->block_cache)
1493 goto err_free_tinymm;
1495 if (ENABLE_THREADED_COMPILER) {
1496 state->rec = lightrec_recompiler_init(state);
1498 goto err_free_block_cache;
1500 state->reaper = lightrec_reaper_init(state);
1502 goto err_free_recompiler;
1504 state->cstate = lightrec_create_cstate(state);
1506 goto err_free_block_cache;
1509 state->nb_maps = nb;
1512 memcpy(&state->ops, ops, sizeof(*ops));
1514 state->dispatcher = generate_dispatcher(state);
1515 if (!state->dispatcher)
1516 goto err_free_reaper;
1518 state->c_wrapper_block = generate_wrapper(state);
1519 if (!state->c_wrapper_block)
1520 goto err_free_dispatcher;
1522 state->c_wrappers[C_WRAPPER_RW] = lightrec_rw_cb;
1523 state->c_wrappers[C_WRAPPER_RW_GENERIC] = lightrec_rw_generic_cb;
1524 state->c_wrappers[C_WRAPPER_MTC] = lightrec_mtc_cb;
1525 state->c_wrappers[C_WRAPPER_CP] = lightrec_cp;
1526 state->c_wrappers[C_WRAPPER_SYSCALL] = lightrec_syscall_cb;
1527 state->c_wrappers[C_WRAPPER_BREAK] = lightrec_break_cb;
1529 map = &state->maps[PSX_MAP_BIOS];
1530 state->offset_bios = (uintptr_t)map->address - map->pc;
1532 map = &state->maps[PSX_MAP_SCRATCH_PAD];
1533 state->offset_scratch = (uintptr_t)map->address - map->pc;
1535 map = &state->maps[PSX_MAP_KERNEL_USER_RAM];
1536 state->offset_ram = (uintptr_t)map->address - map->pc;
1538 if (state->maps[PSX_MAP_MIRROR1].address == map->address + 0x200000 &&
1539 state->maps[PSX_MAP_MIRROR2].address == map->address + 0x400000 &&
1540 state->maps[PSX_MAP_MIRROR3].address == map->address + 0x600000)
1541 state->mirrors_mapped = true;
1543 if (state->offset_bios == 0 &&
1544 state->offset_scratch == 0 &&
1545 state->offset_ram == 0 &&
1546 state->mirrors_mapped) {
1547 pr_info("Memory map is perfect. Emitted code will be best.\n");
1549 pr_info("Memory map is sub-par. Emitted code will be slow.\n");
1554 err_free_dispatcher:
1555 lightrec_free_block(state, state->dispatcher);
1557 if (ENABLE_THREADED_COMPILER)
1558 lightrec_reaper_destroy(state->reaper);
1559 err_free_recompiler:
1560 if (ENABLE_THREADED_COMPILER)
1561 lightrec_free_recompiler(state->rec);
1563 lightrec_free_cstate(state->cstate);
1564 err_free_block_cache:
1565 lightrec_free_block_cache(state->block_cache);
1568 tinymm_shutdown(state->tinymm);
1571 lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1572 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1579 void lightrec_destroy(struct lightrec_state *state)
1581 /* Force a print info on destroy*/
1582 state->current_cycle = ~state->current_cycle;
1583 lightrec_print_info(state);
1585 if (ENABLE_THREADED_COMPILER) {
1586 lightrec_free_recompiler(state->rec);
1587 lightrec_reaper_destroy(state->reaper);
1589 lightrec_free_cstate(state->cstate);
1592 lightrec_free_block_cache(state->block_cache);
1593 lightrec_free_block(state, state->dispatcher);
1594 lightrec_free_block(state, state->c_wrapper_block);
1598 tinymm_shutdown(state->tinymm);
1600 lightrec_unregister(MEM_FOR_LIGHTREC, sizeof(*state) +
1601 sizeof(*state->code_lut) * CODE_LUT_SIZE);
1605 void lightrec_invalidate(struct lightrec_state *state, u32 addr, u32 len)
1607 u32 kaddr = kunseg(addr & ~0x3);
1608 const struct lightrec_mem_map *map = lightrec_get_map(state, NULL, kaddr);
1611 if (map != &state->maps[PSX_MAP_KERNEL_USER_RAM])
1614 /* Handle mirrors */
1615 kaddr &= (state->maps[PSX_MAP_KERNEL_USER_RAM].length - 1);
1617 lightrec_invalidate_map(state, map, kaddr, len);
1621 void lightrec_invalidate_all(struct lightrec_state *state)
1623 memset(state->code_lut, 0, sizeof(*state->code_lut) * CODE_LUT_SIZE);
1626 void lightrec_set_invalidate_mode(struct lightrec_state *state, bool dma_only)
1628 if (state->invalidate_from_dma_only != dma_only)
1629 lightrec_invalidate_all(state);
1631 state->invalidate_from_dma_only = dma_only;
1634 void lightrec_set_exit_flags(struct lightrec_state *state, u32 flags)
1636 if (flags != LIGHTREC_EXIT_NORMAL) {
1637 state->exit_flags |= flags;
1638 state->target_cycle = state->current_cycle;
1642 u32 lightrec_exit_flags(struct lightrec_state *state)
1644 return state->exit_flags;
1647 u32 lightrec_current_cycle_count(const struct lightrec_state *state)
1649 return state->current_cycle;
1652 void lightrec_reset_cycle_count(struct lightrec_state *state, u32 cycles)
1654 state->current_cycle = cycles;
1656 if (state->target_cycle < cycles)
1657 state->target_cycle = cycles;
1660 void lightrec_set_target_cycle_count(struct lightrec_state *state, u32 cycles)
1662 if (state->exit_flags == LIGHTREC_EXIT_NORMAL) {
1663 if (cycles < state->current_cycle)
1664 cycles = state->current_cycle;
1666 state->target_cycle = cycles;
1670 struct lightrec_registers * lightrec_get_registers(struct lightrec_state *state)
1672 return &state->regs;