* Copyright (C) 2014-2021 Paul Cercueil <paul@crapouillou.net>
*/
+#include "constprop.h"
#include "lightrec-config.h"
#include "disassembler.h"
#include "lightrec.h"
for (i = offset; ; i++) {
c = list[i].c;
- if (opcode_reads_register(c, reg)) {
- if (i > 0 && has_delay_slot(list[i - 1].c))
- break;
-
+ if (opcode_reads_register(c, reg))
return i;
- }
- if (op_flag_sync(list[i].flags) ||
- has_delay_slot(c) || opcode_writes_register(c, reg))
+ if (op_flag_sync(list[i].flags)
+ || (op_flag_no_ds(list[i].flags) && has_delay_slot(c))
+ || is_delay_slot(list, i)
+ || opcode_writes_register(c, reg))
break;
}
{
unsigned int i;
- if (op_flag_sync(list[offset].flags))
+ if (op_flag_sync(list[offset].flags) || is_delay_slot(list, offset))
return false;
for (i = offset + 1; ; i++) {
return false;
}
-static u32 lightrec_propagate_consts(const struct opcode *op,
- const struct opcode *prev,
- u32 known, u32 *v)
+static void lightrec_optimize_sll_sra(struct opcode *list, unsigned int offset,
+ struct constprop_data *v)
{
- union code c = prev->c;
-
- /* Register $zero is always, well, zero */
- known |= BIT(0);
- v[0] = 0;
-
- if (op_flag_sync(op->flags))
- return BIT(0);
-
- switch (c.i.op) {
- case OP_SPECIAL:
- switch (c.r.op) {
- case OP_SPECIAL_SLL:
- if (known & BIT(c.r.rt)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = v[c.r.rt] << c.r.imm;
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_SRL:
- if (known & BIT(c.r.rt)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = v[c.r.rt] >> c.r.imm;
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_SRA:
- if (known & BIT(c.r.rt)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = (s32)v[c.r.rt] >> c.r.imm;
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_SLLV:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = v[c.r.rt] << (v[c.r.rs] & 0x1f);
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_SRLV:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = v[c.r.rt] >> (v[c.r.rs] & 0x1f);
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_SRAV:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = (s32)v[c.r.rt]
- >> (v[c.r.rs] & 0x1f);
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_ADD:
- case OP_SPECIAL_ADDU:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = (s32)v[c.r.rt] + (s32)v[c.r.rs];
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_SUB:
- case OP_SPECIAL_SUBU:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = v[c.r.rt] - v[c.r.rs];
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_AND:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = v[c.r.rt] & v[c.r.rs];
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_OR:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = v[c.r.rt] | v[c.r.rs];
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_XOR:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = v[c.r.rt] ^ v[c.r.rs];
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_NOR:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = ~(v[c.r.rt] | v[c.r.rs]);
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_SLT:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = (s32)v[c.r.rs] < (s32)v[c.r.rt];
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_SLTU:
- if (known & BIT(c.r.rt) && known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = v[c.r.rs] < v[c.r.rt];
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_SPECIAL_MULT:
- case OP_SPECIAL_MULTU:
- case OP_SPECIAL_DIV:
- case OP_SPECIAL_DIVU:
- if (OPT_FLAG_MULT_DIV && c.r.rd)
- known &= ~BIT(c.r.rd);
- if (OPT_FLAG_MULT_DIV && c.r.imm)
- known &= ~BIT(c.r.imm);
- break;
- case OP_SPECIAL_MFLO:
- case OP_SPECIAL_MFHI:
- known &= ~BIT(c.r.rd);
- break;
- default:
- break;
- }
- break;
- case OP_META_MULT2:
- case OP_META_MULTU2:
- if (OPT_FLAG_MULT_DIV && (known & BIT(c.r.rs))) {
- if (c.r.rd) {
- known |= BIT(c.r.rd);
-
- if (c.r.op < 32)
- v[c.r.rd] = v[c.r.rs] << c.r.op;
- else
- v[c.r.rd] = 0;
- }
-
- if (c.r.imm) {
- known |= BIT(c.r.imm);
-
- if (c.r.op >= 32)
- v[c.r.imm] = v[c.r.rs] << (c.r.op - 32);
- else if (c.i.op == OP_META_MULT2)
- v[c.r.imm] = (s32) v[c.r.rs] >> (32 - c.r.op);
- else
- v[c.r.imm] = v[c.r.rs] >> (32 - c.r.op);
- }
- } else {
- if (OPT_FLAG_MULT_DIV && c.r.rd)
- known &= ~BIT(c.r.rd);
- if (OPT_FLAG_MULT_DIV && c.r.imm)
- known &= ~BIT(c.r.imm);
- }
- break;
- case OP_REGIMM:
- break;
- case OP_ADDI:
- case OP_ADDIU:
- if (known & BIT(c.i.rs)) {
- known |= BIT(c.i.rt);
- v[c.i.rt] = v[c.i.rs] + (s32)(s16)c.i.imm;
- } else {
- known &= ~BIT(c.i.rt);
- }
- break;
- case OP_SLTI:
- if (known & BIT(c.i.rs)) {
- known |= BIT(c.i.rt);
- v[c.i.rt] = (s32)v[c.i.rs] < (s32)(s16)c.i.imm;
- } else {
- known &= ~BIT(c.i.rt);
- }
- break;
- case OP_SLTIU:
- if (known & BIT(c.i.rs)) {
- known |= BIT(c.i.rt);
- v[c.i.rt] = v[c.i.rs] < (u32)(s32)(s16)c.i.imm;
- } else {
- known &= ~BIT(c.i.rt);
- }
- break;
- case OP_ANDI:
- if (known & BIT(c.i.rs)) {
- known |= BIT(c.i.rt);
- v[c.i.rt] = v[c.i.rs] & c.i.imm;
- } else {
- known &= ~BIT(c.i.rt);
- }
- break;
- case OP_ORI:
- if (known & BIT(c.i.rs)) {
- known |= BIT(c.i.rt);
- v[c.i.rt] = v[c.i.rs] | c.i.imm;
- } else {
- known &= ~BIT(c.i.rt);
- }
- break;
- case OP_XORI:
- if (known & BIT(c.i.rs)) {
- known |= BIT(c.i.rt);
- v[c.i.rt] = v[c.i.rs] ^ c.i.imm;
- } else {
- known &= ~BIT(c.i.rt);
- }
- break;
- case OP_LUI:
- known |= BIT(c.i.rt);
- v[c.i.rt] = c.i.imm << 16;
- break;
- case OP_CP0:
- switch (c.r.rs) {
- case OP_CP0_MFC0:
- case OP_CP0_CFC0:
- known &= ~BIT(c.r.rt);
- break;
- }
- break;
- case OP_CP2:
- if (c.r.op == OP_CP2_BASIC) {
- switch (c.r.rs) {
- case OP_CP2_BASIC_MFC2:
- case OP_CP2_BASIC_CFC2:
- known &= ~BIT(c.r.rt);
- break;
- }
- }
- break;
- case OP_LB:
- case OP_LH:
- case OP_LWL:
- case OP_LW:
- case OP_LBU:
- case OP_LHU:
- case OP_LWR:
- case OP_LWC2:
- known &= ~BIT(c.i.rt);
- break;
- case OP_META_MOV:
- if (known & BIT(c.r.rs)) {
- known |= BIT(c.r.rd);
- v[c.r.rd] = v[c.r.rs];
- } else {
- known &= ~BIT(c.r.rd);
- }
- break;
- case OP_META_EXTC:
- if (known & BIT(c.i.rs)) {
- known |= BIT(c.i.rt);
- v[c.i.rt] = (s32)(s8)v[c.i.rs];
- } else {
- known &= ~BIT(c.i.rt);
- }
- break;
- case OP_META_EXTS:
- if (known & BIT(c.i.rs)) {
- known |= BIT(c.i.rt);
- v[c.i.rt] = (s32)(s16)v[c.i.rs];
- } else {
- known &= ~BIT(c.i.rt);
- }
- break;
- default:
- break;
- }
-
- return known;
-}
-
-static void lightrec_optimize_sll_sra(struct opcode *list, unsigned int offset)
-{
- struct opcode *prev, *prev2 = NULL, *curr = &list[offset];
+ struct opcode *ldop = NULL, *curr = &list[offset], *next;
struct opcode *to_change, *to_nop;
int idx, idx2;
if (curr->r.imm != 24 && curr->r.imm != 16)
return;
- idx = find_prev_writer(list, offset, curr->r.rt);
+ if (is_delay_slot(list, offset))
+ return;
+
+ idx = find_next_reader(list, offset + 1, curr->r.rd);
if (idx < 0)
return;
- prev = &list[idx];
+ next = &list[idx];
- if (prev->i.op != OP_SPECIAL || prev->r.op != OP_SPECIAL_SLL ||
- prev->r.imm != curr->r.imm || prev->r.rd != curr->r.rt)
+ if (next->i.op != OP_SPECIAL || next->r.op != OP_SPECIAL_SRA ||
+ next->r.imm != curr->r.imm || next->r.rt != curr->r.rd)
return;
- if (prev->r.rd != prev->r.rt && curr->r.rd != curr->r.rt) {
+ if (curr->r.rd != curr->r.rt && next->r.rd != next->r.rt) {
/* sll rY, rX, 16
* ...
- * srl rZ, rY, 16 */
+ * sra rZ, rY, 16 */
- if (!reg_is_dead(list, offset, curr->r.rt) ||
- reg_is_read_or_written(list, idx, offset, curr->r.rd))
+ if (!reg_is_dead(list, idx, curr->r.rd) ||
+ reg_is_read_or_written(list, offset, idx, next->r.rd))
return;
/* If rY is dead after the SRL, and rZ is not used after the SLL,
* we can change rY to rZ */
pr_debug("Detected SLL/SRA with middle temp register\n");
- prev->r.rd = curr->r.rd;
- curr->r.rt = prev->r.rd;
+ curr->r.rd = next->r.rd;
+ next->r.rt = curr->r.rd;
}
- /* We got a SLL/SRA combo. If imm #16, that's a cast to u16.
- * If imm #24 that's a cast to u8.
+ /* We got a SLL/SRA combo. If imm #16, that's a cast to s16.
+ * If imm #24 that's a cast to s8.
*
* First of all, make sure that the target register of the SLL is not
- * read before the SRA. */
+ * read after the SRA. */
- if (prev->r.rd == prev->r.rt) {
+ if (curr->r.rd == curr->r.rt) {
/* sll rX, rX, 16
* ...
- * srl rY, rX, 16 */
- to_change = curr;
- to_nop = prev;
+ * sra rY, rX, 16 */
+ to_change = next;
+ to_nop = curr;
/* rX is used after the SRA - we cannot convert it. */
- if (prev->r.rd != curr->r.rd && !reg_is_dead(list, offset, prev->r.rd))
+ if (curr->r.rd != next->r.rd && !reg_is_dead(list, idx, curr->r.rd))
return;
} else {
/* sll rY, rX, 16
* ...
- * srl rY, rY, 16 */
- to_change = prev;
- to_nop = curr;
+ * sra rY, rY, 16 */
+ to_change = curr;
+ to_nop = next;
}
- idx2 = find_prev_writer(list, idx, prev->r.rt);
+ idx2 = find_prev_writer(list, offset, curr->r.rt);
if (idx2 >= 0) {
/* Note that PSX games sometimes do casts after
* a LHU or LBU; in this case we can change the
* load opcode to a LH or LB, and the cast can
* be changed to a MOV or a simple NOP. */
- prev2 = &list[idx2];
+ ldop = &list[idx2];
- if (curr->r.rd != prev2->i.rt &&
- !reg_is_dead(list, offset, prev2->i.rt))
- prev2 = NULL;
- else if (curr->r.imm == 16 && prev2->i.op == OP_LHU)
- prev2->i.op = OP_LH;
- else if (curr->r.imm == 24 && prev2->i.op == OP_LBU)
- prev2->i.op = OP_LB;
+ if (next->r.rd != ldop->i.rt &&
+ !reg_is_dead(list, idx, ldop->i.rt))
+ ldop = NULL;
+ else if (curr->r.imm == 16 && ldop->i.op == OP_LHU)
+ ldop->i.op = OP_LH;
+ else if (curr->r.imm == 24 && ldop->i.op == OP_LBU)
+ ldop->i.op = OP_LB;
else
- prev2 = NULL;
+ ldop = NULL;
- if (prev2) {
- if (curr->r.rd == prev2->i.rt) {
+ if (ldop) {
+ if (next->r.rd == ldop->i.rt) {
to_change->opcode = 0;
- } else if (reg_is_dead(list, offset, prev2->i.rt) &&
- !reg_is_read_or_written(list, idx2 + 1, offset, curr->r.rd)) {
+ } else if (reg_is_dead(list, idx, ldop->i.rt) &&
+ !reg_is_read_or_written(list, idx2 + 1, idx, next->r.rd)) {
/* The target register of the SRA is dead after the
* LBU/LHU; we can change the target register of the
* LBU/LHU to the one of the SRA. */
- prev2->i.rt = curr->r.rd;
+ v[ldop->i.rt].known = 0;
+ v[ldop->i.rt].sign = 0;
+ ldop->i.rt = next->r.rd;
to_change->opcode = 0;
} else {
to_change->i.op = OP_META_MOV;
- to_change->r.rd = curr->r.rd;
- to_change->r.rs = prev2->i.rt;
+ to_change->r.rd = next->r.rd;
+ to_change->r.rs = ldop->i.rt;
}
if (to_nop->r.imm == 24)
pr_debug("Convert LBU+SLL+SRA to LB\n");
else
pr_debug("Convert LHU+SLL+SRA to LH\n");
+
+ v[ldop->i.rt].known = 0;
+ v[ldop->i.rt].sign = 0xffffff80 << 24 - curr->r.imm;
}
}
- if (!prev2) {
+ if (!ldop) {
pr_debug("Convert SLL/SRA #%u to EXT%c\n",
- prev->r.imm,
- prev->r.imm == 24 ? 'C' : 'S');
+ curr->r.imm, curr->r.imm == 24 ? 'C' : 'S');
- if (to_change == prev) {
- to_change->i.rs = prev->r.rt;
- to_change->i.rt = curr->r.rd;
+ if (to_change == curr) {
+ to_change->i.rs = curr->r.rt;
+ to_change->i.rt = next->r.rd;
} else {
- to_change->i.rt = curr->r.rd;
- to_change->i.rs = prev->r.rt;
+ to_change->i.rt = next->r.rd;
+ to_change->i.rs = curr->r.rt;
}
if (to_nop->r.imm == 24)
to_nop->opcode = 0;
}
-static void lightrec_remove_useless_lui(struct block *block, unsigned int offset,
- u32 known, u32 *values)
+static void
+lightrec_remove_useless_lui(struct block *block, unsigned int offset,
+ const struct constprop_data *v)
{
struct opcode *list = block->opcode_list,
*op = &block->opcode_list[offset];
int reader;
- if (!op_flag_sync(op->flags) && (known & BIT(op->i.rt)) &&
- values[op->i.rt] == op->i.imm << 16) {
+ if (!op_flag_sync(op->flags) && is_known(v, op->i.rt) &&
+ v[op->i.rt].value == op->i.imm << 16) {
pr_debug("Converting duplicated LUI to NOP\n");
op->opcode = 0x0;
return;
}
- if (op->i.imm != 0 || op->i.rt == 0)
+ if (op->i.imm != 0 || op->i.rt == 0 || offset == block->nb_ops - 1)
return;
reader = find_next_reader(list, offset + 1, op->i.rt);
return popcount32(value) == 1;
}
+static void lightrec_patch_known_zero(struct opcode *op,
+ const struct constprop_data *v)
+{
+ switch (op->i.op) {
+ case OP_SPECIAL:
+ switch (op->r.op) {
+ case OP_SPECIAL_JR:
+ case OP_SPECIAL_JALR:
+ case OP_SPECIAL_MTHI:
+ case OP_SPECIAL_MTLO:
+ if (is_known_zero(v, op->r.rs))
+ op->r.rs = 0;
+ break;
+ default:
+ if (is_known_zero(v, op->r.rs))
+ op->r.rs = 0;
+ fallthrough;
+ case OP_SPECIAL_SLL:
+ case OP_SPECIAL_SRL:
+ case OP_SPECIAL_SRA:
+ if (is_known_zero(v, op->r.rt))
+ op->r.rt = 0;
+ break;
+ case OP_SPECIAL_SYSCALL:
+ case OP_SPECIAL_BREAK:
+ case OP_SPECIAL_MFHI:
+ case OP_SPECIAL_MFLO:
+ break;
+ }
+ break;
+ case OP_CP0:
+ switch (op->r.rs) {
+ case OP_CP0_MTC0:
+ case OP_CP0_CTC0:
+ if (is_known_zero(v, op->r.rt))
+ op->r.rt = 0;
+ break;
+ default:
+ break;
+ }
+ break;
+ case OP_CP2:
+ if (op->r.op == OP_CP2_BASIC) {
+ switch (op->r.rs) {
+ case OP_CP2_BASIC_MTC2:
+ case OP_CP2_BASIC_CTC2:
+ if (is_known_zero(v, op->r.rt))
+ op->r.rt = 0;
+ break;
+ default:
+ break;
+ }
+ }
+ break;
+ case OP_BEQ:
+ case OP_BNE:
+ if (is_known_zero(v, op->i.rt))
+ op->i.rt = 0;
+ fallthrough;
+ case OP_REGIMM:
+ case OP_BLEZ:
+ case OP_BGTZ:
+ case OP_ADDI:
+ case OP_ADDIU:
+ case OP_SLTI:
+ case OP_SLTIU:
+ case OP_ANDI:
+ case OP_ORI:
+ case OP_XORI:
+ case OP_META_MOV:
+ case OP_META_EXTC:
+ case OP_META_EXTS:
+ case OP_META_MULT2:
+ case OP_META_MULTU2:
+ if (is_known_zero(v, op->i.rs))
+ op->i.rs = 0;
+ break;
+ case OP_SB:
+ case OP_SH:
+ case OP_SWL:
+ case OP_SW:
+ case OP_SWR:
+ if (is_known_zero(v, op->i.rt))
+ op->i.rt = 0;
+ fallthrough;
+ case OP_LB:
+ case OP_LH:
+ case OP_LWL:
+ case OP_LW:
+ case OP_LBU:
+ case OP_LHU:
+ case OP_LWR:
+ case OP_LWC2:
+ case OP_SWC2:
+ if (is_known(v, op->i.rs)
+ && kunseg(v[op->i.rs].value) == 0)
+ op->i.rs = 0;
+ break;
+ default:
+ break;
+ }
+}
+
+static void lightrec_reset_syncs(struct block *block)
+{
+ struct opcode *op, *list = block->opcode_list;
+ unsigned int i;
+ s32 offset;
+
+ for (i = 0; i < block->nb_ops; i++)
+ list[i].flags &= ~LIGHTREC_SYNC;
+
+ for (i = 0; i < block->nb_ops; i++) {
+ op = &list[i];
+
+ if (op_flag_local_branch(op->flags) && has_delay_slot(op->c)) {
+ offset = i + 1 + (s16)op->i.imm;
+ list[offset].flags |= LIGHTREC_SYNC;
+ }
+ }
+}
+
static int lightrec_transform_ops(struct lightrec_state *state, struct block *block)
{
- struct opcode *list = block->opcode_list;
- struct opcode *prev, *op = NULL;
- u32 known = BIT(0);
- u32 values[32] = { 0 };
+ struct opcode *op, *list = block->opcode_list;
+ struct constprop_data v[32] = LIGHTREC_CONSTPROP_INITIALIZER;
unsigned int i;
+ bool local;
u8 tmp;
for (i = 0; i < block->nb_ops; i++) {
- prev = op;
op = &list[i];
- if (prev)
- known = lightrec_propagate_consts(op, prev, known, values);
+ lightrec_consts_propagate(list, i, v);
+
+ lightrec_patch_known_zero(op, v);
/* Transform all opcodes detected as useless to real NOPs
* (0x0: SLL r0, r0, #0) */
switch (op->i.op) {
case OP_BEQ:
- if (op->i.rs == op->i.rt) {
+ if (op->i.rs == op->i.rt ||
+ (is_known(v, op->i.rs) && is_known(v, op->i.rt) &&
+ v[op->i.rs].value == v[op->i.rt].value)) {
+ if (op->i.rs != op->i.rt)
+ pr_debug("Found always-taken BEQ\n");
+
op->i.rs = 0;
op->i.rt = 0;
+ } else if (v[op->i.rs].known & v[op->i.rt].known &
+ (v[op->i.rs].value ^ v[op->i.rt].value)) {
+ pr_debug("Found never-taken BEQ\n");
+
+ local = op_flag_local_branch(op->flags);
+ op->opcode = 0;
+ op->flags = 0;
+
+ if (local)
+ lightrec_reset_syncs(block);
} else if (op->i.rs == 0) {
op->i.rs = op->i.rt;
op->i.rt = 0;
break;
case OP_BNE:
- if (op->i.rs == 0) {
+ if (v[op->i.rs].known & v[op->i.rt].known &
+ (v[op->i.rs].value ^ v[op->i.rt].value)) {
+ pr_debug("Found always-taken BNE\n");
+
+ op->i.op = OP_BEQ;
+ op->i.rs = 0;
+ op->i.rt = 0;
+ } else if (is_known(v, op->i.rs) && is_known(v, op->i.rt) &&
+ v[op->i.rs].value == v[op->i.rt].value) {
+ pr_debug("Found never-taken BNE\n");
+
+ local = op_flag_local_branch(op->flags);
+ op->opcode = 0;
+ op->flags = 0;
+
+ if (local)
+ lightrec_reset_syncs(block);
+ } else if (op->i.rs == 0) {
op->i.rs = op->i.rt;
op->i.rt = 0;
}
break;
+ case OP_BLEZ:
+ if (v[op->i.rs].known & BIT(31) &&
+ v[op->i.rs].value & BIT(31)) {
+ pr_debug("Found always-taken BLEZ\n");
+
+ op->i.op = OP_BEQ;
+ op->i.rs = 0;
+ op->i.rt = 0;
+ }
+ break;
+
+ case OP_BGTZ:
+ if (v[op->i.rs].known & BIT(31) &&
+ v[op->i.rs].value & BIT(31)) {
+ pr_debug("Found never-taken BGTZ\n");
+
+ local = op_flag_local_branch(op->flags);
+ op->opcode = 0;
+ op->flags = 0;
+
+ if (local)
+ lightrec_reset_syncs(block);
+ }
+ break;
+
case OP_LUI:
- if (!prev || !has_delay_slot(prev->c))
+ if (i == 0 || !has_delay_slot(list[i - 1].c))
lightrec_modify_lui(block, i);
- lightrec_remove_useless_lui(block, i, known, values);
+ lightrec_remove_useless_lui(block, i, v);
break;
/* Transform ORI/ADDI/ADDIU with imm #0 or ORR/ADD/ADDU/SUB/SUBU
op->r.rd = op->i.rt;
}
break;
+ case OP_ANDI:
+ if (bits_are_known_zero(v, op->i.rs, ~op->i.imm)) {
+ pr_debug("Found useless ANDI 0x%x\n", op->i.imm);
+
+ if (op->i.rs == op->i.rt) {
+ op->opcode = 0;
+ } else {
+ op->i.op = OP_META_MOV;
+ op->r.rd = op->i.rt;
+ }
+ }
+ break;
+ case OP_REGIMM:
+ switch (op->r.rt) {
+ case OP_REGIMM_BLTZ:
+ case OP_REGIMM_BGEZ:
+ if (!(v[op->r.rs].known & BIT(31)))
+ break;
+
+ if (!!(v[op->r.rs].value & BIT(31))
+ ^ (op->r.rt == OP_REGIMM_BGEZ)) {
+ pr_debug("Found always-taken BLTZ/BGEZ\n");
+ op->i.op = OP_BEQ;
+ op->i.rs = 0;
+ op->i.rt = 0;
+ } else {
+ pr_debug("Found never-taken BLTZ/BGEZ\n");
+
+ local = op_flag_local_branch(op->flags);
+ op->opcode = 0;
+ op->flags = 0;
+
+ if (local)
+ lightrec_reset_syncs(block);
+ }
+ break;
+ case OP_REGIMM_BLTZAL:
+ case OP_REGIMM_BGEZAL:
+ /* TODO: Detect always-taken and replace with JAL */
+ break;
+ }
+ break;
case OP_SPECIAL:
switch (op->r.op) {
+ case OP_SPECIAL_SRAV:
+ if ((v[op->r.rs].known & 0x1f) != 0x1f)
+ break;
+
+ pr_debug("Convert SRAV to SRA\n");
+ op->r.imm = v[op->r.rs].value & 0x1f;
+ op->r.op = OP_SPECIAL_SRA;
+
+ fallthrough;
case OP_SPECIAL_SRA:
if (op->r.imm == 0) {
pr_debug("Convert SRA #0 to MOV\n");
op->r.rs = op->r.rt;
break;
}
-
- lightrec_optimize_sll_sra(block->opcode_list, i);
break;
+
+ case OP_SPECIAL_SLLV:
+ if ((v[op->r.rs].known & 0x1f) != 0x1f)
+ break;
+
+ pr_debug("Convert SLLV to SLL\n");
+ op->r.imm = v[op->r.rs].value & 0x1f;
+ op->r.op = OP_SPECIAL_SLL;
+
+ fallthrough;
case OP_SPECIAL_SLL:
+ if (op->r.imm == 0) {
+ pr_debug("Convert SLL #0 to MOV\n");
+ op->i.op = OP_META_MOV;
+ op->r.rs = op->r.rt;
+ }
+
+ lightrec_optimize_sll_sra(block->opcode_list, i, v);
+ break;
+
+ case OP_SPECIAL_SRLV:
+ if ((v[op->r.rs].known & 0x1f) != 0x1f)
+ break;
+
+ pr_debug("Convert SRLV to SRL\n");
+ op->r.imm = v[op->r.rs].value & 0x1f;
+ op->r.op = OP_SPECIAL_SRL;
+
+ fallthrough;
case OP_SPECIAL_SRL:
if (op->r.imm == 0) {
- pr_debug("Convert SLL/SRL #0 to MOV\n");
+ pr_debug("Convert SRL #0 to MOV\n");
op->i.op = OP_META_MOV;
op->r.rs = op->r.rt;
}
break;
+
case OP_SPECIAL_MULT:
case OP_SPECIAL_MULTU:
- if ((known & BIT(op->r.rs)) &&
- is_power_of_two(values[op->r.rs])) {
+ if (is_known(v, op->r.rs) &&
+ is_power_of_two(v[op->r.rs].value)) {
tmp = op->c.i.rs;
op->c.i.rs = op->c.i.rt;
op->c.i.rt = tmp;
- } else if (!(known & BIT(op->r.rt)) ||
- !is_power_of_two(values[op->r.rt])) {
+ } else if (!is_known(v, op->r.rt) ||
+ !is_power_of_two(v[op->r.rt].value)) {
break;
}
pr_debug("Multiply by power-of-two: %u\n",
- values[op->r.rt]);
+ v[op->r.rt].value);
if (op->r.op == OP_SPECIAL_MULT)
op->i.op = OP_META_MULT2;
else
op->i.op = OP_META_MULTU2;
- op->r.op = ctz32(values[op->r.rt]);
+ op->r.op = ctz32(v[op->r.rt].value);
break;
case OP_SPECIAL_OR:
case OP_SPECIAL_ADD:
op.opcode == 0 || next_op.opcode == 0)
continue;
- if (i && has_delay_slot(block->opcode_list[i - 1].c) &&
- !op_flag_no_ds(block->opcode_list[i - 1].flags))
+ if (is_delay_slot(block->opcode_list, i))
continue;
if (op_flag_sync(next->flags))
continue;
}
- pr_debug("Adding sync at offset 0x%x\n", offset << 2);
-
- block->opcode_list[offset].flags |= LIGHTREC_SYNC;
list->flags |= LIGHTREC_LOCAL_BRANCH;
}
+ lightrec_reset_syncs(block);
+
return 0;
}
}
}
+bool is_delay_slot(const struct opcode *list, unsigned int offset)
+{
+ return offset > 0
+ && !op_flag_no_ds(list[offset - 1].flags)
+ && has_delay_slot(list[offset - 1].c);
+}
+
bool should_emulate(const struct opcode *list)
{
return op_flag_emulate_branch(list->flags) && has_delay_slot(list->c);
static int lightrec_flag_io(struct lightrec_state *state, struct block *block)
{
- struct opcode *prev = NULL, *list = NULL;
+ struct opcode *list;
enum psx_map psx_map;
- u32 known = BIT(0);
- u32 values[32] = { 0 };
+ struct constprop_data v[32] = LIGHTREC_CONSTPROP_INITIALIZER;
unsigned int i;
u32 val, kunseg_val;
bool no_mask;
for (i = 0; i < block->nb_ops; i++) {
- prev = list;
list = &block->opcode_list[i];
- if (prev)
- known = lightrec_propagate_consts(list, prev, known, values);
+ lightrec_consts_propagate(block->opcode_list, i, v);
switch (list->i.op) {
case OP_SB:
/* Detect writes whose destination address is inside the
* current block, using constant propagation. When these
* occur, we mark the blocks as not compilable. */
- if ((known & BIT(list->i.rs)) &&
- kunseg(values[list->i.rs]) >= kunseg(block->pc) &&
- kunseg(values[list->i.rs]) < (kunseg(block->pc) +
- block->nb_ops * 4)) {
+ if (is_known(v, list->i.rs) &&
+ kunseg(v[list->i.rs].value) >= kunseg(block->pc) &&
+ kunseg(v[list->i.rs].value) < (kunseg(block->pc) +
+ block->nb_ops * 4)) {
pr_debug("Self-modifying block detected\n");
block_set_flags(block, BLOCK_NEVER_COMPILE);
list->flags |= LIGHTREC_SMC;
case OP_LWL:
case OP_LWR:
case OP_LWC2:
- if (OPT_FLAG_IO && (known & BIT(list->i.rs))) {
- val = values[list->i.rs] + (s16) list->i.imm;
- kunseg_val = kunseg(val);
- psx_map = lightrec_get_map_idx(state, kunseg_val);
+ if (OPT_FLAG_IO &&
+ (v[list->i.rs].known | v[list->i.rs].sign)) {
+ psx_map = lightrec_get_constprop_map(state, v,
+ list->i.rs,
+ (s16) list->i.imm);
+
+ if (psx_map != PSX_MAP_UNKNOWN && !is_known(v, list->i.rs))
+ pr_debug("Detected map thanks to bit-level const propagation!\n");
list->flags &= ~LIGHTREC_IO_MASK;
- no_mask = val == kunseg_val;
+
+ val = v[list->i.rs].value + (s16) list->i.imm;
+ kunseg_val = kunseg(val);
+
+ no_mask = (v[list->i.rs].known & ~v[list->i.rs].value
+ & 0xe0000000) == 0xe0000000;
switch (psx_map) {
case PSX_MAP_KERNEL_USER_RAM:
if (no_mask)
list->flags |= LIGHTREC_NO_MASK;
- break;
+ } else {
+ pr_debug("Flagging opcode %u as I/O access\n",
+ i);
+ list->flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
}
- fallthrough;
+ break;
default:
- pr_debug("Flagging opcode %u as I/O access\n",
- i);
- list->flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
break;
}
}
static int lightrec_flag_mults_divs(struct lightrec_state *state, struct block *block)
{
- struct opcode *prev, *list = NULL;
+ struct opcode *list = NULL;
+ struct constprop_data v[32] = LIGHTREC_CONSTPROP_INITIALIZER;
u8 reg_hi, reg_lo;
unsigned int i;
- u32 known = BIT(0);
- u32 values[32] = { 0 };
for (i = 0; i < block->nb_ops - 1; i++) {
- prev = list;
list = &block->opcode_list[i];
- if (prev)
- known = lightrec_propagate_consts(list, prev, known, values);
+ lightrec_consts_propagate(block->opcode_list, i, v);
switch (list->i.op) {
case OP_SPECIAL:
/* If we are dividing by a non-zero constant, don't
* emit the div-by-zero check. */
if (lightrec_always_skip_div_check() ||
- ((known & BIT(list->c.r.rt)) && values[list->c.r.rt]))
+ (v[list->r.rt].known & v[list->r.rt].value)) {
list->flags |= LIGHTREC_NO_DIV_CHECK;
+ }
fallthrough;
case OP_SPECIAL_MULT:
case OP_SPECIAL_MULTU:
}
/* Don't support opcodes in delay slots */
- if ((i && has_delay_slot(block->opcode_list[i - 1].c)) ||
+ if (is_delay_slot(block->opcode_list, i) ||
op_flag_no_ds(list->flags)) {
continue;
}