static void fix_wval(MVMThreadContext *tc, MVMSpeshGraph *inliner,
MVMSpeshGraph *inlinee, MVMSpeshIns *to_fix) {
/* Resolve object, then just put it into a spesh slot. (Could do some
* smarter things like trying to see if the SC is referenced by both
* compilation units, too.) */
MVMCompUnit *cu = inlinee->sf->body.cu;
MVMint16 dep = to_fix->operands[1].lit_i16;
MVMint64 idx = to_fix->info->opcode == MVM_OP_wval
? to_fix->operands[2].lit_i16
: to_fix->operands[2].lit_i64;
if (dep >= 0 && dep < cu->body.num_scs) {
MVMSerializationContext *sc = MVM_sc_get_sc(tc, cu, dep);
if (sc) {
MVMObject *obj = MVM_sc_get_object(tc, sc, idx);
MVMint16 ss = MVM_spesh_add_spesh_slot(tc, inliner, (MVMCollectable *)obj);
to_fix->info = MVM_op_get_op(MVM_OP_sp_getspeshslot);
to_fix->operands[1].lit_i16 = ss;
}
else {
MVM_oops(tc,
"Spesh inline: SC not yet resolved; lookup failed");
}
}
else {
MVM_oops(tc,
"Spesh inline: invalid SC index found");
}
}
/* Tweak the successor of a BB, also updating the target BBs pred. */
static void tweak_succ(MVMThreadContext *tc, MVMSpeshGraph *g, MVMSpeshBB *bb, MVMSpeshBB *new_succ) {
if (bb->num_succ == 0) {
bb->succ = MVM_spesh_alloc(tc, g, sizeof(MVMSpeshBB *));
bb->num_succ = 1;
}
if (bb->num_succ == 1)
bb->succ[0] = new_succ;
else
MVM_oops(tc, "Spesh inline: unexpected num_succ");
if (new_succ->num_pred == 0) {
new_succ->pred = MVM_spesh_alloc(tc, g, sizeof(MVMSpeshBB *));
new_succ->num_pred = 1;
new_succ->pred[0] = bb;
}
else {
MVMint32 found = 0;
MVMint32 i;
for (i = 0; i < new_succ->num_pred; i++)
if (new_succ->pred[i]->idx + 1 == new_succ->idx) {
new_succ->pred[i] = bb;
found = 1;
break;
}
if (!found)
MVM_oops(tc,
"Spesh inline: could not find appropriate pred to update\n");
}
}
/* Does a lookup of information held about a synthetic. The synth parameter
* must be a synthetic codepoint (that is, negative). The memory returned is
* not to be freed by the caller; it also is only valid until the next GC
* safe point. */
MVMNFGSynthetic * MVM_nfg_get_synthetic_info(MVMThreadContext *tc, MVMGrapheme32 synth) {
MVMNFGState *nfg = tc->instance->nfg;
MVMint32 synth_idx = -synth - 1;
if (synth >= 0)
MVM_oops(tc, "MVM_nfg_get_synthetic_info illegally called on a non-synthetic codepoint.\nRequested codepoint %i.", synth);
if (synth_idx >= nfg->num_synthetics)
MVM_oops(tc, "MVM_nfg_get_synthetic_info call requested a synthetic codepoint that does not exist.\nRequested synthetic %i when only %i have been created.", -synth, nfg->num_synthetics);
return &(nfg->synthetics[synth_idx]);
}
/* Finds the deopt index of the return. */
static MVMint32 return_deopt_idx(MVMThreadContext *tc, MVMSpeshIns *invoke_ins) {
MVMSpeshAnn *ann = invoke_ins->annotations;
while (ann) {
if (ann->type == MVM_SPESH_ANN_DEOPT_ALL_INS)
return ann->data.deopt_idx;
ann = ann->next;
}
MVM_oops(tc, "Spesh inline: return_deopt_idx failed");
}
static MVMint32 find_deopt_target(MVMThreadContext *tc, MVMFrame *f, MVMint32 deopt_offset) {
MVMint32 i;
for (i = 0; i < f->spesh_cand->num_deopts * 2; i += 2) {
if (f->spesh_cand->deopts[i + 1] == deopt_offset) {
return f->spesh_cand->deopts[i];
}
}
MVM_oops(tc, "find_deopt_target failed for %s (%s)",
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.name),
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.cuuid));
}
/* Locates deopt index matching OSR point. */
static MVMint32 get_osr_deopt_index(MVMThreadContext *tc, MVMSpeshCandidate *cand) {
/* Calculate offset. */
MVMint32 offset = (*(tc->interp_cur_op) - *(tc->interp_bytecode_start));
/* Locate it in the deopt table. */
MVMint32 i;
for (i = 0; i < cand->num_deopts; i++)
if (cand->deopts[2 * i] == offset)
return i;
/* If we couldn't locate it, something is really very wrong. */
MVM_oops(tc, "Spesh: get_osr_deopt_index failed");
}
static void rewrite_returns(MVMThreadContext *tc, MVMSpeshGraph *inliner,
MVMSpeshGraph *inlinee, MVMSpeshBB *invoke_bb,
MVMSpeshIns *invoke_ins) {
/* Locate return instructions. */
MVMSpeshBB *bb = inlinee->entry;
while (bb) {
MVMSpeshIns *ins = bb->first_ins;
while (ins) {
MVMuint16 opcode = ins->info->opcode;
switch (opcode) {
case MVM_OP_return:
if (invoke_ins->info->opcode == MVM_OP_invoke_v) {
MVM_spesh_manipulate_insert_goto(tc, inliner, bb, ins,
invoke_bb->succ[0]);
tweak_succ(tc, inliner, bb, invoke_bb->succ[0]);
}
else {
MVM_oops(tc,
"Spesh inline: return_v/invoke_[!v] mismatch");
}
break;
case MVM_OP_return_i:
MVM_spesh_manipulate_insert_goto(tc, inliner, bb, ins,
invoke_bb->succ[0]);
tweak_succ(tc, inliner, bb, invoke_bb->succ[0]);
rewrite_int_return(tc, inliner, bb, ins, invoke_bb, invoke_ins);
break;
case MVM_OP_return_n:
MVM_spesh_manipulate_insert_goto(tc, inliner, bb, ins,
invoke_bb->succ[0]);
tweak_succ(tc, inliner, bb, invoke_bb->succ[0]);
rewrite_num_return(tc, inliner, bb, ins, invoke_bb, invoke_ins);
break;
case MVM_OP_return_s:
MVM_spesh_manipulate_insert_goto(tc, inliner, bb, ins,
invoke_bb->succ[0]);
tweak_succ(tc, inliner, bb, invoke_bb->succ[0]);
rewrite_str_return(tc, inliner, bb, ins, invoke_bb, invoke_ins);
break;
case MVM_OP_return_o:
MVM_spesh_manipulate_insert_goto(tc, inliner, bb, ins,
invoke_bb->succ[0]);
tweak_succ(tc, inliner, bb, invoke_bb->succ[0]);
rewrite_obj_return(tc, inliner, bb, ins, invoke_bb, invoke_ins);
break;
}
ins = ins->next;
}
bb = bb->linear_next;
}
}
/* De-optimizes the current frame by directly specifying the addresses */
void MVM_spesh_deopt_one_direct(MVMThreadContext *tc, MVMint32 deopt_offset,
MVMint32 deopt_target) {
MVMFrame *f = tc->cur_frame;
if (tc->instance->profiling)
MVM_profiler_log_deopt_one(tc);
if (f->effective_bytecode != f->static_info->body.bytecode) {
deopt_frame(tc, tc->cur_frame, deopt_offset, deopt_target);
} else {
MVM_oops(tc, "deopt_one_direct failed for %s (%s)",
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.name),
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.cuuid));
}
}
static void rewrite_obj_return(MVMThreadContext *tc, MVMSpeshGraph *g,
MVMSpeshBB *return_bb, MVMSpeshIns *return_ins,
MVMSpeshBB *invoke_bb, MVMSpeshIns *invoke_ins) {
switch (invoke_ins->info->opcode) {
case MVM_OP_invoke_v:
MVM_spesh_manipulate_delete_ins(tc, g, return_bb, return_ins);
break;
case MVM_OP_invoke_o:
return_to_set(tc, g, return_ins, invoke_ins->operands[0]);
break;
default:
MVM_oops(tc,
"Spesh inline: unhandled case of return_o");
}
}
/* De-optimizes the currently executing frame, provided it is specialized and
* at a valid de-optimization point. Typically used when a guard fails. */
void MVM_spesh_deopt_one(MVMThreadContext *tc) {
MVMFrame *f = tc->cur_frame;
if (tc->instance->profiling)
MVM_profiler_log_deopt_one(tc);
/*fprintf(stderr, "deopt_one requested in frame '%s' (cuid '%s')\n",
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.name),
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.cuuid));*/
if (f->effective_bytecode != f->static_info->body.bytecode) {
MVMint32 deopt_offset = *(tc->interp_cur_op) - f->effective_bytecode;
MVMint32 deopt_target = find_deopt_target(tc, f, deopt_offset);
deopt_frame(tc, tc->cur_frame, deopt_offset, deopt_target);
}
else {
MVM_oops(tc, "deopt_one failed for %s (%s)",
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.name),
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.cuuid));
}
}
/* Ensures that a given compilation unit has access to the specified extop. */
static void demand_extop(MVMThreadContext *tc, MVMCompUnit *target_cu, MVMCompUnit *source_cu,
const MVMOpInfo *info) {
MVMExtOpRecord *extops;
MVMuint16 i, num_extops;
MVM_reentrantmutex_lock(tc, (MVMReentrantMutex *)target_cu->body.update_mutex);
/* See if the target compunit already has the extop. */
extops = target_cu->body.extops;
num_extops = target_cu->body.num_extops;
for (i = 0; i < num_extops; i++)
if (extops[i].info == info) {
MVM_reentrantmutex_unlock(tc, (MVMReentrantMutex *)target_cu->body.update_mutex);
return;
}
/* If not, need to add it. Locate it in the source CU. */
extops = source_cu->body.extops;
num_extops = source_cu->body.num_extops;
for (i = 0; i < num_extops; i++) {
if (extops[i].info == info) {
MVMuint32 size = (target_cu->body.num_extops + 1) * sizeof(MVMExtOpRecord);
target_cu->body.extops = target_cu->body.extops
? MVM_realloc(target_cu->body.extops, size)
: MVM_malloc(size);
memcpy(&target_cu->body.extops[target_cu->body.num_extops],
&extops[i], sizeof(MVMExtOpRecord));
target_cu->body.num_extops++;
MVM_reentrantmutex_unlock(tc, (MVMReentrantMutex *)target_cu->body.update_mutex);
return;
}
}
/* Didn't find it; should be impossible. */
MVM_reentrantmutex_unlock(tc, (MVMReentrantMutex *)target_cu->body.update_mutex);
MVM_oops(tc, "Spesh: inline failed to find source CU extop entry");
}
/* Merges the inlinee's spesh graph into the inliner. */
static void merge_graph(MVMThreadContext *tc, MVMSpeshGraph *inliner,
MVMSpeshGraph *inlinee, MVMCode *inlinee_code,
MVMSpeshIns *invoke_ins) {
MVMSpeshFacts **merged_facts;
MVMuint16 *merged_fact_counts;
MVMint32 i, total_inlines, orig_deopt_addrs;
/* If the inliner and inlinee are from different compilation units, we
* potentially have to fix up extra things. */
MVMint32 same_comp_unit = inliner->sf->body.cu == inlinee->sf->body.cu;
/* Renumber the locals, lexicals, and basic blocks of the inlinee; also
* re-write any indexes in annotations that need it. */
MVMSpeshBB *bb = inlinee->entry;
while (bb) {
MVMSpeshIns *ins = bb->first_ins;
while (ins) {
MVMuint16 opcode = ins->info->opcode;
MVMSpeshAnn *ann = ins->annotations;
while (ann) {
switch (ann->type) {
case MVM_SPESH_ANN_FH_START:
case MVM_SPESH_ANN_FH_END:
case MVM_SPESH_ANN_FH_GOTO:
ann->data.frame_handler_index += inliner->num_handlers;
break;
case MVM_SPESH_ANN_DEOPT_INLINE:
ann->data.deopt_idx += inliner->num_deopt_addrs;
break;
case MVM_SPESH_ANN_INLINE_START:
case MVM_SPESH_ANN_INLINE_END:
ann->data.inline_idx += inliner->num_inlines;
break;
}
ann = ann->next;
}
if (opcode == MVM_SSA_PHI) {
for (i = 0; i < ins->info->num_operands; i++)
ins->operands[i].reg.orig += inliner->num_locals;
}
else {
for (i = 0; i < ins->info->num_operands; i++) {
MVMuint8 flags = ins->info->operands[i];
switch (flags & MVM_operand_rw_mask) {
case MVM_operand_read_reg:
case MVM_operand_write_reg:
ins->operands[i].reg.orig += inliner->num_locals;
break;
case MVM_operand_read_lex:
case MVM_operand_write_lex:
ins->operands[i].lex.idx += inliner->num_lexicals;
break;
default: {
MVMuint32 type = flags & MVM_operand_type_mask;
if (type == MVM_operand_spesh_slot) {
ins->operands[i].lit_i16 += inliner->num_spesh_slots;
}
else if (type == MVM_operand_callsite) {
if (!same_comp_unit)
fix_callsite(tc, inliner, inlinee, &(ins->operands[i]));
}
else if (type == MVM_operand_coderef) {
if (!same_comp_unit)
fix_coderef(tc, inliner, inlinee, &(ins->operands[i]));
}
else if (type == MVM_operand_str) {
if (!same_comp_unit)
fix_str(tc, inliner, inlinee, &(ins->operands[i]));
}
break;
}
}
}
}
ins = ins->next;
}
bb->idx += inliner->num_bbs - 1; /* -1 as we won't include entry */
bb->inlined = 1;
bb = bb->linear_next;
}
/* Incorporate the basic blocks by concatening them onto the end of the
* linear_next chain of the inliner; skip the inlinee's fake entry BB. */
bb = inliner->entry;
while (bb) {
if (!bb->linear_next) {
/* Found the end; insert and we're done. */
bb->linear_next = inlinee->entry->linear_next;
bb = NULL;
}
else {
bb = bb->linear_next;
}
}
/* Merge facts. */
merged_facts = MVM_spesh_alloc(tc, inliner,
(inliner->num_locals + inlinee->num_locals) * sizeof(MVMSpeshFacts *));
memcpy(merged_facts, inliner->facts,
inliner->num_locals * sizeof(MVMSpeshFacts *));
memcpy(merged_facts + inliner->num_locals, inlinee->facts,
inlinee->num_locals * sizeof(MVMSpeshFacts *));
inliner->facts = merged_facts;
merged_fact_counts = MVM_spesh_alloc(tc, inliner,
(inliner->num_locals + inlinee->num_locals) * sizeof(MVMuint16));
memcpy(merged_fact_counts, inliner->fact_counts,
inliner->num_locals * sizeof(MVMuint16));
memcpy(merged_fact_counts + inliner->num_locals, inlinee->fact_counts,
inlinee->num_locals * sizeof(MVMuint16));
inliner->fact_counts = merged_fact_counts;
/* Copy over spesh slots. */
for (i = 0; i < inlinee->num_spesh_slots; i++)
MVM_spesh_add_spesh_slot(tc, inliner, inlinee->spesh_slots[i]);
/* If they are from separate compilation units, make another pass through
* to fix up on wvals. Note we can't do this in the first pass as we must
* not modify the spesh slots once we've got started with the rewrites.
* Now we've resolved all that, we're good to map wvals elsewhere into
* some extra spesh slots. */
if (!same_comp_unit) {
bb = inlinee->entry;
while (bb) {
MVMSpeshIns *ins = bb->first_ins;
while (ins) {
MVMuint16 opcode = ins->info->opcode;
if (opcode == MVM_OP_wval || opcode == MVM_OP_wval_wide)
fix_wval(tc, inliner, inlinee, ins);
ins = ins->next;
}
bb = bb->linear_next;
}
}
/* Merge de-opt tables, if needed. */
orig_deopt_addrs = inliner->num_deopt_addrs;
if (inlinee->num_deopt_addrs) {
assert(inlinee->deopt_addrs != inliner->deopt_addrs);
inliner->alloc_deopt_addrs += inlinee->alloc_deopt_addrs;
if (inliner->deopt_addrs)
inliner->deopt_addrs = MVM_realloc(inliner->deopt_addrs,
inliner->alloc_deopt_addrs * sizeof(MVMint32) * 2);
else
inliner->deopt_addrs = MVM_malloc(inliner->alloc_deopt_addrs * sizeof(MVMint32) * 2);
memcpy(inliner->deopt_addrs + inliner->num_deopt_addrs * 2,
inlinee->deopt_addrs, inlinee->alloc_deopt_addrs * sizeof(MVMint32) * 2);
inliner->num_deopt_addrs += inlinee->num_deopt_addrs;
}
/* Merge inlines table, and add us an entry too. */
total_inlines = inliner->num_inlines + inlinee->num_inlines + 1;
inliner->inlines = inliner->num_inlines
? MVM_realloc(inliner->inlines, total_inlines * sizeof(MVMSpeshInline))
: MVM_malloc(total_inlines * sizeof(MVMSpeshInline));
memcpy(inliner->inlines + inliner->num_inlines, inlinee->inlines,
inlinee->num_inlines * sizeof(MVMSpeshInline));
for (i = inliner->num_inlines; i < total_inlines - 1; i++) {
inliner->inlines[i].locals_start += inliner->num_locals;
inliner->inlines[i].lexicals_start += inliner->num_lexicals;
inliner->inlines[i].return_deopt_idx += orig_deopt_addrs;
}
inliner->inlines[total_inlines - 1].code = inlinee_code;
inliner->inlines[total_inlines - 1].g = inlinee;
inliner->inlines[total_inlines - 1].locals_start = inliner->num_locals;
inliner->inlines[total_inlines - 1].lexicals_start = inliner->num_lexicals;
switch (invoke_ins->info->opcode) {
case MVM_OP_invoke_v:
inliner->inlines[total_inlines - 1].res_type = MVM_RETURN_VOID;
break;
case MVM_OP_invoke_o:
inliner->inlines[total_inlines - 1].res_reg = invoke_ins->operands[0].reg.orig;
inliner->inlines[total_inlines - 1].res_type = MVM_RETURN_OBJ;
break;
case MVM_OP_invoke_i:
inliner->inlines[total_inlines - 1].res_reg = invoke_ins->operands[0].reg.orig;
inliner->inlines[total_inlines - 1].res_type = MVM_RETURN_INT;
break;
case MVM_OP_invoke_n:
inliner->inlines[total_inlines - 1].res_reg = invoke_ins->operands[0].reg.orig;
inliner->inlines[total_inlines - 1].res_type = MVM_RETURN_NUM;
break;
case MVM_OP_invoke_s:
inliner->inlines[total_inlines - 1].res_reg = invoke_ins->operands[0].reg.orig;
inliner->inlines[total_inlines - 1].res_type = MVM_RETURN_STR;
break;
default:
MVM_oops(tc, "Spesh inline: unknown invoke instruction");
}
inliner->inlines[total_inlines - 1].return_deopt_idx = return_deopt_idx(tc, invoke_ins);
inliner->num_inlines = total_inlines;
/* Create/update per-specialization local and lexical type maps. */
if (!inliner->local_types) {
MVMint32 local_types_size = inliner->num_locals * sizeof(MVMuint16);
inliner->local_types = MVM_malloc(local_types_size);
memcpy(inliner->local_types, inliner->sf->body.local_types, local_types_size);
}
inliner->local_types = MVM_realloc(inliner->local_types,
(inliner->num_locals + inlinee->num_locals) * sizeof(MVMuint16));
memcpy(inliner->local_types + inliner->num_locals,
inlinee->local_types ? inlinee->local_types : inlinee->sf->body.local_types,
inlinee->num_locals * sizeof(MVMuint16));
if (!inliner->lexical_types) {
MVMint32 lexical_types_size = inliner->num_lexicals * sizeof(MVMuint16);
inliner->lexical_types = MVM_malloc(lexical_types_size);
memcpy(inliner->lexical_types, inliner->sf->body.lexical_types, lexical_types_size);
}
inliner->lexical_types = MVM_realloc(inliner->lexical_types,
(inliner->num_lexicals + inlinee->num_lexicals) * sizeof(MVMuint16));
memcpy(inliner->lexical_types + inliner->num_lexicals,
inlinee->lexical_types ? inlinee->lexical_types : inlinee->sf->body.lexical_types,
inlinee->num_lexicals * sizeof(MVMuint16));
/* Merge handlers. */
if (inlinee->num_handlers) {
MVMuint32 total_handlers = inliner->num_handlers + inlinee->num_handlers;
if (inliner->handlers == inliner->sf->body.handlers) {
/* Original handlers table; need a copy. */
MVMFrameHandler *new_handlers = MVM_malloc(total_handlers * sizeof(MVMFrameHandler));
memcpy(new_handlers, inliner->handlers,
inliner->num_handlers * sizeof(MVMFrameHandler));
inliner->handlers = new_handlers;
}
else {
/* Probably already did some inlines into this frame; resize. */
inliner->handlers = MVM_realloc(inliner->handlers,
total_handlers * sizeof(MVMFrameHandler));
}
memcpy(inliner->handlers + inliner->num_handlers, inlinee->handlers,
inlinee->num_handlers * sizeof(MVMFrameHandler));
for (i = inliner->num_handlers; i < total_handlers; i++) {
inliner->handlers[i].block_reg += inliner->num_locals;
inliner->handlers[i].label_reg += inliner->num_locals;
}
}
/* Update total locals, lexicals, basic blocks, and handlers of the
* inliner. */
inliner->num_bbs += inlinee->num_bbs - 1;
inliner->num_locals += inlinee->num_locals;
inliner->num_lexicals += inlinee->num_lexicals;
inliner->num_handlers += inlinee->num_handlers;
}
/* Writes instructions within a basic block boundary. */
static void write_instructions(MVMThreadContext *tc, MVMSpeshGraph *g, SpeshWriterState *ws, MVMSpeshBB *bb) {
MVMSpeshIns *ins = bb->first_ins;
while (ins) {
MVMint32 i;
/* Process any annotations. */
MVMSpeshAnn *ann = ins->annotations;
MVMSpeshAnn *deopt_one_ann = NULL;
MVMSpeshAnn *deopt_all_ann = NULL;
MVMSpeshAnn *deopt_inline_ann = NULL;
while (ann) {
switch (ann->type) {
case MVM_SPESH_ANN_FH_START:
ws->handlers[ann->data.frame_handler_index].start_offset =
ws->bytecode_pos;
break;
case MVM_SPESH_ANN_FH_END:
ws->handlers[ann->data.frame_handler_index].end_offset =
ws->bytecode_pos;
break;
case MVM_SPESH_ANN_FH_GOTO:
ws->handlers[ann->data.frame_handler_index].goto_offset =
ws->bytecode_pos;
break;
case MVM_SPESH_ANN_DEOPT_ONE_INS:
deopt_one_ann = ann;
break;
case MVM_SPESH_ANN_DEOPT_ALL_INS:
deopt_all_ann = ann;
break;
case MVM_SPESH_ANN_INLINE_START:
g->inlines[ann->data.inline_idx].start = ws->bytecode_pos;
break;
case MVM_SPESH_ANN_INLINE_END:
g->inlines[ann->data.inline_idx].end = ws->bytecode_pos;
break;
case MVM_SPESH_ANN_DEOPT_INLINE:
deopt_inline_ann = ann;
break;
case MVM_SPESH_ANN_DEOPT_OSR:
g->deopt_addrs[2 * ann->data.deopt_idx + 1] = ws->bytecode_pos;
break;
}
ann = ann->next;
}
if (ins->info->opcode != MVM_SSA_PHI) {
/* Real instruction, not a phi. Emit opcode. */
if (ins->info->opcode == (MVMuint16)-1) {
/* Ext op; resolve. */
MVMExtOpRecord *extops = g->sf->body.cu->body.extops;
MVMuint16 num_extops = g->sf->body.cu->body.num_extops;
MVMint32 found = 0;
for (i = 0; i < num_extops; i++) {
if (extops[i].info == ins->info) {
write_int16(ws, MVM_OP_EXT_BASE + i);
found = 1;
break;
}
}
if (!found)
MVM_oops(tc, "Spesh: failed to resolve extop in code-gen");
}
else {
/* Core op. */
write_int16(ws, ins->info->opcode);
}
/* Write out operands. */
for (i = 0; i < ins->info->num_operands; i++) {
MVMuint8 flags = ins->info->operands[i];
MVMuint8 rw = flags & MVM_operand_rw_mask;
switch (rw) {
case MVM_operand_read_reg:
case MVM_operand_write_reg:
write_int16(ws, ins->operands[i].reg.orig);
break;
case MVM_operand_read_lex:
case MVM_operand_write_lex:
write_int16(ws, ins->operands[i].lex.idx);
write_int16(ws, ins->operands[i].lex.outers);
break;
case MVM_operand_literal: {
MVMuint8 type = flags & MVM_operand_type_mask;
switch (type) {
case MVM_operand_int8:
write_int8(ws, ins->operands[i].lit_i8);
break;
case MVM_operand_int16:
write_int16(ws, ins->operands[i].lit_i16);
break;
case MVM_operand_int32:
write_int32(ws, ins->operands[i].lit_i32);
break;
case MVM_operand_int64:
write_int64(ws, ins->operands[i].lit_i64);
break;
case MVM_operand_num32:
write_num32(ws, ins->operands[i].lit_n32);
break;
case MVM_operand_num64:
write_num64(ws, ins->operands[i].lit_n64);
break;
case MVM_operand_callsite:
write_int16(ws, ins->operands[i].callsite_idx);
break;
case MVM_operand_coderef:
write_int16(ws, ins->operands[i].coderef_idx);
break;
case MVM_operand_str:
write_int32(ws, ins->operands[i].lit_str_idx);
break;
case MVM_operand_ins: {
MVMint32 offset = ws->bb_offsets[ins->operands[i].ins_bb->idx];
if (offset >= 0) {
/* Already know where it is, so just write it. */
write_int32(ws, offset);
}
else {
/* Need to fix it up. */
if (ws->num_fixups == ws->alloc_fixups) {
ws->alloc_fixups *= 2;
ws->fixup_locations = MVM_realloc(ws->fixup_locations,
ws->alloc_fixups * sizeof(MVMint32));
ws->fixup_bbs = MVM_realloc(ws->fixup_bbs,
ws->alloc_fixups * sizeof(MVMSpeshBB *));
}
ws->fixup_locations[ws->num_fixups] = ws->bytecode_pos;
ws->fixup_bbs[ws->num_fixups] = ins->operands[i].ins_bb;
write_int32(ws, 0);
ws->num_fixups++;
}
break;
}
case MVM_operand_spesh_slot:
write_int16(ws, ins->operands[i].lit_i16);
break;
default:
MVM_oops(tc,
"Spesh: unknown operand type %d in codegen (op %s)",
(int)type, ins->info->name);
}
}
break;
default:
MVM_oops(tc, "Spesh: unknown operand type in codegen");
}
}
}
/* If there was a deopt point annotation, update table. */
if (deopt_one_ann)
g->deopt_addrs[2 * deopt_one_ann->data.deopt_idx + 1] = ws->bytecode_pos;
if (deopt_all_ann)
g->deopt_addrs[2 * deopt_all_ann->data.deopt_idx + 1] = ws->bytecode_pos;
if (deopt_inline_ann)
g->deopt_addrs[2 * deopt_inline_ann->data.deopt_idx + 1] = ws->bytecode_pos;
ins = ins->next;
}
}
/* Finalizes OSR. */
void MVM_spesh_osr_finalize(MVMThreadContext *tc) {
/* Find deopt index using existing deopt table, for entering the updated
* code later. */
MVMSpeshCandidate *specialized = tc->cur_frame->spesh_cand;
MVMint32 osr_index = get_osr_deopt_finalize_index(tc, specialized);
MVMJitCode *jc;
/* Finish up the specialization. */
MVM_spesh_candidate_specialize(tc, tc->cur_frame->static_info, specialized);
/* Resize work area if needed. */
if (specialized->num_locals > tc->cur_frame->static_info->body.num_locals) {
/* Resize work area. */
MVMRegister *new_work = MVM_fixed_size_alloc_zeroed(tc, tc->instance->fsa,
specialized->work_size);
memcpy(new_work, tc->cur_frame->work,
tc->cur_frame->static_info->body.num_locals * sizeof(MVMRegister));
MVM_fixed_size_free(tc, tc->instance->fsa, tc->cur_frame->allocd_work,
tc->cur_frame->work);
tc->cur_frame->work = new_work;
tc->cur_frame->allocd_work = specialized->work_size;
tc->cur_frame->args = tc->cur_frame->work + specialized->num_locals;
}
/* Resize environment if needed. */
if (specialized->num_lexicals > tc->cur_frame->static_info->body.num_lexicals) {
MVMRegister *new_env = MVM_fixed_size_alloc_zeroed(tc, tc->instance->fsa,
specialized->env_size);
if (tc->cur_frame->allocd_env) {
memcpy(new_env, tc->cur_frame->env,
tc->cur_frame->static_info->body.num_lexicals * sizeof(MVMRegister));
MVM_fixed_size_free(tc, tc->instance->fsa, tc->cur_frame->allocd_env,
tc->cur_frame->env);
}
tc->cur_frame->env = new_env;
tc->cur_frame->allocd_env = specialized->env_size;
}
/* Sync frame with updates. */
tc->cur_frame->effective_bytecode = specialized->bytecode;
tc->cur_frame->effective_handlers = specialized->handlers;
tc->cur_frame->effective_spesh_slots = specialized->spesh_slots;
tc->cur_frame->spesh_log_slots = NULL;
tc->cur_frame->spesh_log_idx = -1;
/* Sync interpreter with updates. */
jc = specialized->jitcode;
if (jc && jc->num_deopts) {
MVMint32 i;
*(tc->interp_bytecode_start) = specialized->jitcode->bytecode;
*(tc->interp_cur_op) = specialized->jitcode->bytecode;
for (i = 0; i < jc->num_deopts; i++) {
if (jc->deopts[i].idx == osr_index) {
tc->cur_frame->jit_entry_label = jc->labels[jc->deopts[i].label];
break;
}
}
if (i == jc->num_deopts)
MVM_oops(tc, "JIT: Could not find OSR label");
if (tc->instance->profiling)
MVM_profiler_log_osr(tc, 1);
} else {
*(tc->interp_bytecode_start) = specialized->bytecode;
*(tc->interp_cur_op) = specialized->bytecode +
specialized->deopts[2 * osr_index + 1];
if (tc->instance->profiling)
MVM_profiler_log_osr(tc, 0);
}
*(tc->interp_reg_base) = tc->cur_frame->work;
/* Tweak frame invocation count so future invocations will use the code
* produced by OSR. */
tc->cur_frame->static_info->body.invocations +=
tc->cur_frame->static_info->body.spesh_threshold;
}
/* Does the jump into the optimized code. */
void perform_osr(MVMThreadContext *tc, MVMSpeshCandidate *specialized) {
MVMJitCode *jit_code;
MVMint32 num_locals;
/* Work out the OSR deopt index, to locate the entry point. */
MVMint32 osr_index = get_osr_deopt_index(tc, specialized);
#if MVM_LOG_OSR
fprintf(stderr, "Performing OSR of frame '%s' (cuid: %s) at index %d\n",
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.name),
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.cuuid),
osr_index);
#endif
jit_code = specialized->jitcode;
num_locals = jit_code && jit_code->local_types ?
jit_code->num_locals : specialized->num_locals;
/* Resize work area if needed. */
if (specialized->work_size > tc->cur_frame->allocd_work) {
/* Resize work area. */
MVMRegister *new_work = MVM_fixed_size_alloc_zeroed(tc, tc->instance->fsa,
specialized->work_size);
MVMRegister *new_args = new_work + num_locals;
memcpy(new_work, tc->cur_frame->work,
tc->cur_frame->static_info->body.num_locals * sizeof(MVMRegister));
memcpy(new_args, tc->cur_frame->args,
tc->cur_frame->static_info->body.cu->body.max_callsite_size * sizeof(MVMRegister));
MVM_fixed_size_free(tc, tc->instance->fsa, tc->cur_frame->allocd_work,
tc->cur_frame->work);
tc->cur_frame->work = new_work;
tc->cur_frame->allocd_work = specialized->work_size;
tc->cur_frame->args = new_args;
#if MVM_LOG_OSR
fprintf(stderr, "OSR resized work area of frame '%s' (cuid: %s)\n",
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.name),
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.cuuid));
#endif
}
else if (specialized->work_size > tc->cur_frame->static_info->body.work_size) {
size_t keep_bytes = tc->cur_frame->static_info->body.num_locals * sizeof(MVMRegister);
size_t to_null = specialized->work_size - keep_bytes;
memset((char *)tc->cur_frame->work + keep_bytes, 0, to_null);
}
/* Resize environment if needed. */
if (specialized->num_lexicals > tc->cur_frame->static_info->body.num_lexicals) {
MVMRegister *new_env = MVM_fixed_size_alloc_zeroed(tc, tc->instance->fsa,
specialized->env_size);
if (tc->cur_frame->allocd_env) {
memcpy(new_env, tc->cur_frame->env,
tc->cur_frame->static_info->body.num_lexicals * sizeof(MVMRegister));
MVM_fixed_size_free(tc, tc->instance->fsa, tc->cur_frame->allocd_env,
tc->cur_frame->env);
}
tc->cur_frame->env = new_env;
tc->cur_frame->allocd_env = specialized->env_size;
#if MVM_LOG_OSR
fprintf(stderr, "OSR resized environment of frame '%s' (cuid: %s)\n",
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.name),
MVM_string_utf8_encode_C_string(tc, tc->cur_frame->static_info->body.cuuid));
#endif
}
else if (specialized->env_size > tc->cur_frame->static_info->body.env_size) {
size_t keep_bytes = tc->cur_frame->static_info->body.num_lexicals * sizeof(MVMRegister);
size_t to_null = specialized->env_size - keep_bytes;
memset((char *)tc->cur_frame->env + keep_bytes, 0, to_null);
}
/* Set up frame to point to spesh candidate/slots. */
tc->cur_frame->effective_spesh_slots = specialized->spesh_slots;
tc->cur_frame->spesh_cand = specialized;
/* Move into the optimized (and maybe JIT-compiled) code. */
if (jit_code && jit_code->num_deopts) {
MVMint32 i;
*(tc->interp_bytecode_start) = jit_code->bytecode;
*(tc->interp_cur_op) = jit_code->bytecode;
for (i = 0; i < jit_code->num_deopts; i++) {
if (jit_code->deopts[i].idx == osr_index) {
tc->cur_frame->jit_entry_label = jit_code->labels[jit_code->deopts[i].label];
break;
}
}
if (i == jit_code->num_deopts)
MVM_oops(tc, "JIT: Could not find OSR label");
if (tc->instance->profiling)
MVM_profiler_log_osr(tc, 1);
} else {
*(tc->interp_bytecode_start) = specialized->bytecode;
*(tc->interp_cur_op) = specialized->bytecode +
specialized->deopts[2 * osr_index + 1];
if (tc->instance->profiling)
MVM_profiler_log_osr(tc, 0);
}
*(tc->interp_reg_base) = tc->cur_frame->work;
}
static void pop_inline(MVMThreadContext *tc, InlineIndexStack *stack) {
stack->cur_depth--;
if (stack->cur_depth < -1)
MVM_oops(tc, "Too many levels of inlining popped");
}
static void push_inline(MVMThreadContext *tc, InlineIndexStack *stack, MVMint32 idx) {
if (stack->cur_depth == 63)
MVM_oops(tc, "Too many levels of inlining to dump");
stack->cur_depth++;
stack->inline_idx[stack->cur_depth] = idx;
}
void MVM_jit_expr_load_operands(MVMThreadContext *tc, MVMJitExprTree *tree, MVMSpeshIns *ins,
struct ValueDefinition *values, MVMint32 *operands) {
MVMint32 i;
for (i = 0; i < ins->info->num_operands; i++) {
MVMSpeshOperand opr = ins->operands[i];
MVMint8 opr_kind = ins->info->operands[i];
switch(opr_kind & MVM_operand_rw_mask) {
case MVM_operand_read_reg:
if (values[opr.reg.orig].node >= 0) {
operands[i] = values[opr.reg.orig].node;
} else {
MVMint32 addr = MVM_jit_expr_add_regaddr(tc, tree, opr.reg.orig);
operands[i] = MVM_jit_expr_add_load(tc, tree, addr);
values[opr.reg.orig].node = operands[i];
values[opr.reg.orig].addr = addr;
values[opr.reg.orig].root = -1; /* load is not part of a root */
}
break;
case MVM_operand_write_reg:
/* get address of register to write */
operands[i] = MVM_jit_expr_add_regaddr(tc, tree, opr.reg.orig);
break;
case MVM_operand_literal:
operands[i] = MVM_jit_expr_add_const(tc, tree, opr, ins->info->operands[i]);
break;
case MVM_operand_read_lex:
{
MVMint32 addr = MVM_jit_expr_add_lexaddr(tc, tree, opr.lex.outers, opr.lex.idx);
operands[i] = MVM_jit_expr_add_load(tc, tree, addr);
break;
}
case MVM_operand_write_lex:
operands[i] = MVM_jit_expr_add_lexaddr(tc, tree, opr.lex.outers, opr.lex.idx);
break;
default:
continue;
}
if (operands[i] >= tree->nodes_num || operands[i] < 0) {
MVM_oops(tc, "JIT: something is wrong with operand loading");
}
}
/* A HACK.
*
* dec_i and inc_i have a single operand that acts both as input and output.
* This is marked only as an output operand, though. Thus, we load the
* address here, and define the value later. However, if we have multiple of
* these in sequence, each will load the old value from memory, disregarding
* the value that an earlier operator has defined, i.e. losing the update.
* That's a bug, and this tries to fix it, by forcing a 'split' between the
* input and the output operand.
*/
if (ins_has_single_input_output_operand(ins)) {
MVMuint16 reg = ins->operands[0].reg.orig;
if (values[reg].node >= 0) {
operands[1] = values[reg].node;
} else {
/* operands[0] has the address */
operands[1] = MVM_jit_expr_add_load(tc, tree, operands[0]);
/* no need to insert it in the table since it will be directly
* overwritten */
}
}
}
/* Re-writes argument passing and parameter taking instructions to simple
* register set operations. */
static void rewrite_args(MVMThreadContext *tc, MVMSpeshGraph *inliner,
MVMSpeshGraph *inlinee, MVMSpeshBB *invoke_bb,
MVMSpeshCallInfo *call_info) {
/* Look for param-taking instructions. Track what arg instructions we
* use in the process. */
MVMSpeshBB *bb = inlinee->entry;
while (bb) {
MVMSpeshIns *ins = bb->first_ins;
while (ins) {
MVMuint16 opcode = ins->info->opcode;
MVMSpeshIns *next = ins->next;
switch (opcode) {
case MVM_OP_sp_getarg_o:
case MVM_OP_sp_getarg_i:
case MVM_OP_sp_getarg_n:
case MVM_OP_sp_getarg_s: {
MVMuint16 idx = ins->operands[1].lit_i16;
MVMSpeshIns *arg_ins = call_info->arg_ins[idx];
switch (arg_ins->info->opcode) {
case MVM_OP_arg_i:
case MVM_OP_arg_n:
case MVM_OP_arg_s:
case MVM_OP_arg_o:
/* Arg passer just becomes a set instruction; delete the
* parameter-taking instruction. */
arg_ins->info = MVM_op_get_op(MVM_OP_set);
arg_ins->operands[0] = ins->operands[0];
MVM_spesh_manipulate_delete_ins(tc, inliner, bb, ins);
MVM_spesh_get_facts(tc, inliner, arg_ins->operands[0])->usages++;
break;
case MVM_OP_argconst_i:
arg_ins->info = MVM_op_get_op(MVM_OP_const_i64);
arg_ins->operands[0] = ins->operands[0];
MVM_spesh_manipulate_delete_ins(tc, inliner, bb, ins);
MVM_spesh_get_facts(tc, inliner, arg_ins->operands[0])->usages++;
break;
case MVM_OP_argconst_n:
arg_ins->info = MVM_op_get_op(MVM_OP_const_n64);
arg_ins->operands[0] = ins->operands[0];
MVM_spesh_manipulate_delete_ins(tc, inliner, bb, ins);
MVM_spesh_get_facts(tc, inliner, arg_ins->operands[0])->usages++;
break;
case MVM_OP_argconst_s:
arg_ins->info = MVM_op_get_op(MVM_OP_const_s);
arg_ins->operands[0] = ins->operands[0];
MVM_spesh_manipulate_delete_ins(tc, inliner, bb, ins);
MVM_spesh_get_facts(tc, inliner, arg_ins->operands[0])->usages++;
break;
default:
MVM_oops(tc,
"Spesh inline: unhandled arg instruction %d",
arg_ins->info->opcode);
}
break;
}
}
ins = next;
}
bb = bb->linear_next;
}
/* Delete the prepargs instruction. */
MVM_spesh_manipulate_delete_ins(tc, inliner, invoke_bb, call_info->prepargs_ins);
}
static void fix_coderef(MVMThreadContext *tc, MVMSpeshGraph *inliner,
MVMSpeshGraph *inlinee, MVMSpeshOperand *to_fix) {
MVM_oops(tc, "Spesh inline: fix_coderef NYI");
}
/* Generate bytecode from a spesh graph. */
MVMSpeshCode * MVM_spesh_codegen(MVMThreadContext *tc, MVMSpeshGraph *g) {
MVMSpeshCode *res;
MVMSpeshBB *bb;
MVMint32 i, hanlen;
/* Initialize writer state. */
SpeshWriterState *ws = MVM_malloc(sizeof(SpeshWriterState));
ws->bytecode_pos = 0;
ws->bytecode_alloc = 1024;
ws->bytecode = MVM_malloc(ws->bytecode_alloc);
ws->bb_offsets = MVM_malloc(g->num_bbs * sizeof(MVMint32));
ws->num_fixups = 0;
ws->alloc_fixups = 64;
ws->fixup_locations = MVM_malloc(ws->alloc_fixups * sizeof(MVMint32));
ws->fixup_bbs = MVM_malloc(ws->alloc_fixups * sizeof(MVMSpeshBB *));
for (i = 0; i < g->num_bbs; i++)
ws->bb_offsets[i] = -1;
/* Create copy of handlers, and -1 all offsets so we can catch missing
* updates. */
hanlen = g->num_handlers * sizeof(MVMFrameHandler);
if (hanlen) {
ws->handlers = MVM_malloc(hanlen);
memcpy(ws->handlers, g->handlers, hanlen);
for (i = 0; i < g->num_handlers; i++) {
ws->handlers[i].start_offset = -1;
ws->handlers[i].end_offset = -1;
ws->handlers[i].goto_offset = -1;
}
}
else {
ws->handlers = NULL;
}
/* -1 all the deopt targets, so we'll easily catch those that don't get
* mapped if we try to use them. Same for inlines. */
for (i = 0; i < g->num_deopt_addrs; i++)
g->deopt_addrs[i * 2 + 1] = -1;
for (i = 0; i < g->num_inlines; i++) {
g->inlines[i].start = -1;
g->inlines[i].end = -1;
}
/* Write out each of the basic blocks, in linear order. Skip the first,
* dummy, block. */
bb = g->entry->linear_next;
while (bb) {
ws->bb_offsets[bb->idx] = ws->bytecode_pos;
write_instructions(tc, g, ws, bb);
bb = bb->linear_next;
}
/* Fixup labels we were too early for. */
for (i = 0; i < ws->num_fixups; i++)
*((MVMuint32 *)(ws->bytecode + ws->fixup_locations[i])) =
ws->bb_offsets[ws->fixup_bbs[i]->idx];
/* Ensure all handlers got fixed up. */
for (i = 0; i < g->num_handlers; i++) {
if (ws->handlers[i].start_offset == -1 ||
ws->handlers[i].end_offset == -1 ||
ws->handlers[i].goto_offset == -1)
MVM_oops(tc, "Spesh: failed to fix up handlers (%d, %d, %d)",
(int)ws->handlers[i].start_offset,
(int)ws->handlers[i].end_offset,
(int)ws->handlers[i].goto_offset);
}
/* Ensure all inlines got fixed up. */
for (i = 0; i < g->num_inlines; i++)
if (g->inlines[i].start == -1 || g->inlines[i].end == -1)
MVM_oops(tc, "Spesh: failed to fix up inline %d", i);
/* Produce result data structure. */
res = MVM_malloc(sizeof(MVMSpeshCode));
res->bytecode = ws->bytecode;
res->bytecode_size = ws->bytecode_pos;
res->handlers = ws->handlers;
/* Cleanup. */
MVM_free(ws->bb_offsets);
MVM_free(ws->fixup_locations);
MVM_free(ws->fixup_bbs);
MVM_free(ws);
return res;
}
/* Merges the inlinee's spesh graph into the inliner. */
static void merge_graph(MVMThreadContext *tc, MVMSpeshGraph *inliner,
MVMSpeshGraph *inlinee, MVMCode *inlinee_code,
MVMSpeshIns *invoke_ins) {
MVMSpeshFacts **merged_facts;
MVMuint16 *merged_fact_counts;
MVMint32 i, total_inlines, orig_deopt_addrs;
MVMSpeshBB *inlinee_first_bb = NULL, *inlinee_last_bb = NULL;
MVMint32 active_handlers_at_invoke = 0;
/* If the inliner and inlinee are from different compilation units, we
* potentially have to fix up extra things. */
MVMint32 same_comp_unit = inliner->sf->body.cu == inlinee->sf->body.cu;
/* Renumber the locals, lexicals, and basic blocks of the inlinee; also
* re-write any indexes in annotations that need it. */
MVMSpeshBB *bb = inlinee->entry;
while (bb) {
MVMSpeshIns *ins = bb->first_ins;
while (ins) {
MVMuint16 opcode = ins->info->opcode;
MVMSpeshAnn *ann = ins->annotations;
while (ann) {
switch (ann->type) {
case MVM_SPESH_ANN_FH_START:
case MVM_SPESH_ANN_FH_END:
case MVM_SPESH_ANN_FH_GOTO:
ann->data.frame_handler_index += inliner->num_handlers;
break;
case MVM_SPESH_ANN_DEOPT_INLINE:
ann->data.deopt_idx += inliner->num_deopt_addrs;
break;
case MVM_SPESH_ANN_INLINE_START:
case MVM_SPESH_ANN_INLINE_END:
ann->data.inline_idx += inliner->num_inlines;
break;
}
ann = ann->next;
}
if (opcode == MVM_SSA_PHI) {
for (i = 0; i < ins->info->num_operands; i++)
ins->operands[i].reg.orig += inliner->num_locals;
}
else {
for (i = 0; i < ins->info->num_operands; i++) {
MVMuint8 flags = ins->info->operands[i];
switch (flags & MVM_operand_rw_mask) {
case MVM_operand_read_reg:
case MVM_operand_write_reg:
ins->operands[i].reg.orig += inliner->num_locals;
break;
case MVM_operand_read_lex:
case MVM_operand_write_lex:
ins->operands[i].lex.idx += inliner->num_lexicals;
break;
default: {
MVMuint32 type = flags & MVM_operand_type_mask;
if (type == MVM_operand_spesh_slot) {
ins->operands[i].lit_i16 += inliner->num_spesh_slots;
}
else if (type == MVM_operand_callsite) {
if (!same_comp_unit)
fix_callsite(tc, inliner, inlinee, &(ins->operands[i]));
}
else if (type == MVM_operand_coderef) {
if (!same_comp_unit)
fix_coderef(tc, inliner, inlinee, &(ins->operands[i]));
}
else if (type == MVM_operand_str) {
if (!same_comp_unit)
fix_str(tc, inliner, inlinee, &(ins->operands[i]));
}
break;
}
}
}
}
ins = ins->next;
}
bb->idx += inliner->num_bbs - 1; /* -1 as we won't include entry */
bb->inlined = 1;
if (!bb->linear_next)
inlinee_last_bb = bb;
bb = bb->linear_next;
}
/* Incorporate the basic blocks by concatening them onto the end of the
* linear_next chain of the inliner; skip the inlinee's fake entry BB. */
bb = inliner->entry;
while (bb) {
if (!bb->linear_next) {
/* Found the end; insert and we're done. */
bb->linear_next = inlinee_first_bb = inlinee->entry->linear_next;
bb = NULL;
}
else {
bb = bb->linear_next;
}
}
/* Merge facts. */
merged_facts = MVM_spesh_alloc(tc, inliner,
(inliner->num_locals + inlinee->num_locals) * sizeof(MVMSpeshFacts *));
memcpy(merged_facts, inliner->facts,
inliner->num_locals * sizeof(MVMSpeshFacts *));
memcpy(merged_facts + inliner->num_locals, inlinee->facts,
inlinee->num_locals * sizeof(MVMSpeshFacts *));
inliner->facts = merged_facts;
merged_fact_counts = MVM_spesh_alloc(tc, inliner,
(inliner->num_locals + inlinee->num_locals) * sizeof(MVMuint16));
memcpy(merged_fact_counts, inliner->fact_counts,
inliner->num_locals * sizeof(MVMuint16));
memcpy(merged_fact_counts + inliner->num_locals, inlinee->fact_counts,
inlinee->num_locals * sizeof(MVMuint16));
inliner->fact_counts = merged_fact_counts;
/* Copy over spesh slots. */
for (i = 0; i < inlinee->num_spesh_slots; i++)
MVM_spesh_add_spesh_slot(tc, inliner, inlinee->spesh_slots[i]);
/* If they are from separate compilation units, make another pass through
* to fix up on wvals. Note we can't do this in the first pass as we must
* not modify the spesh slots once we've got started with the rewrites.
* Now we've resolved all that, we're good to map wvals elsewhere into
* some extra spesh slots. */
if (!same_comp_unit) {
bb = inlinee->entry;
while (bb) {
MVMSpeshIns *ins = bb->first_ins;
while (ins) {
MVMuint16 opcode = ins->info->opcode;
if (opcode == MVM_OP_wval || opcode == MVM_OP_wval_wide)
fix_wval(tc, inliner, inlinee, ins);
ins = ins->next;
}
bb = bb->linear_next;
}
}
/* Merge de-opt tables, if needed. */
orig_deopt_addrs = inliner->num_deopt_addrs;
if (inlinee->num_deopt_addrs) {
assert(inlinee->deopt_addrs != inliner->deopt_addrs);
inliner->alloc_deopt_addrs += inlinee->alloc_deopt_addrs;
if (inliner->deopt_addrs)
inliner->deopt_addrs = MVM_realloc(inliner->deopt_addrs,
inliner->alloc_deopt_addrs * sizeof(MVMint32) * 2);
else
inliner->deopt_addrs = MVM_malloc(inliner->alloc_deopt_addrs * sizeof(MVMint32) * 2);
memcpy(inliner->deopt_addrs + inliner->num_deopt_addrs * 2,
inlinee->deopt_addrs, inlinee->alloc_deopt_addrs * sizeof(MVMint32) * 2);
inliner->num_deopt_addrs += inlinee->num_deopt_addrs;
}
/* Merge inlines table, and add us an entry too. */
total_inlines = inliner->num_inlines + inlinee->num_inlines + 1;
inliner->inlines = inliner->num_inlines
? MVM_realloc(inliner->inlines, total_inlines * sizeof(MVMSpeshInline))
: MVM_malloc(total_inlines * sizeof(MVMSpeshInline));
memcpy(inliner->inlines + inliner->num_inlines, inlinee->inlines,
inlinee->num_inlines * sizeof(MVMSpeshInline));
for (i = inliner->num_inlines; i < total_inlines - 1; i++) {
inliner->inlines[i].locals_start += inliner->num_locals;
inliner->inlines[i].lexicals_start += inliner->num_lexicals;
inliner->inlines[i].return_deopt_idx += orig_deopt_addrs;
}
inliner->inlines[total_inlines - 1].code = inlinee_code;
inliner->inlines[total_inlines - 1].g = inlinee;
inliner->inlines[total_inlines - 1].locals_start = inliner->num_locals;
inliner->inlines[total_inlines - 1].lexicals_start = inliner->num_lexicals;
switch (invoke_ins->info->opcode) {
case MVM_OP_invoke_v:
inliner->inlines[total_inlines - 1].res_type = MVM_RETURN_VOID;
break;
case MVM_OP_invoke_o:
inliner->inlines[total_inlines - 1].res_reg = invoke_ins->operands[0].reg.orig;
inliner->inlines[total_inlines - 1].res_type = MVM_RETURN_OBJ;
break;
case MVM_OP_invoke_i:
inliner->inlines[total_inlines - 1].res_reg = invoke_ins->operands[0].reg.orig;
inliner->inlines[total_inlines - 1].res_type = MVM_RETURN_INT;
break;
case MVM_OP_invoke_n:
inliner->inlines[total_inlines - 1].res_reg = invoke_ins->operands[0].reg.orig;
inliner->inlines[total_inlines - 1].res_type = MVM_RETURN_NUM;
break;
case MVM_OP_invoke_s:
inliner->inlines[total_inlines - 1].res_reg = invoke_ins->operands[0].reg.orig;
inliner->inlines[total_inlines - 1].res_type = MVM_RETURN_STR;
break;
default:
MVM_oops(tc, "Spesh inline: unknown invoke instruction");
}
inliner->inlines[total_inlines - 1].return_deopt_idx = return_deopt_idx(tc, invoke_ins);
inliner->num_inlines = total_inlines;
/* Create/update per-specialization local and lexical type maps. */
if (!inliner->local_types) {
MVMint32 local_types_size = inliner->num_locals * sizeof(MVMuint16);
inliner->local_types = MVM_malloc(local_types_size);
memcpy(inliner->local_types, inliner->sf->body.local_types, local_types_size);
}
inliner->local_types = MVM_realloc(inliner->local_types,
(inliner->num_locals + inlinee->num_locals) * sizeof(MVMuint16));
memcpy(inliner->local_types + inliner->num_locals,
inlinee->local_types ? inlinee->local_types : inlinee->sf->body.local_types,
inlinee->num_locals * sizeof(MVMuint16));
if (!inliner->lexical_types) {
MVMint32 lexical_types_size = inliner->num_lexicals * sizeof(MVMuint16);
inliner->lexical_types = MVM_malloc(lexical_types_size);
memcpy(inliner->lexical_types, inliner->sf->body.lexical_types, lexical_types_size);
}
inliner->lexical_types = MVM_realloc(inliner->lexical_types,
(inliner->num_lexicals + inlinee->num_lexicals) * sizeof(MVMuint16));
memcpy(inliner->lexical_types + inliner->num_lexicals,
inlinee->lexical_types ? inlinee->lexical_types : inlinee->sf->body.lexical_types,
inlinee->num_lexicals * sizeof(MVMuint16));
/* Merge handlers from inlinee. */
if (inlinee->num_handlers) {
MVMuint32 total_handlers = inliner->num_handlers + inlinee->num_handlers;
resize_handlers_table(tc, inliner, total_handlers);
memcpy(inliner->handlers + inliner->num_handlers, inlinee->handlers,
inlinee->num_handlers * sizeof(MVMFrameHandler));
for (i = inliner->num_handlers; i < total_handlers; i++) {
inliner->handlers[i].block_reg += inliner->num_locals;
inliner->handlers[i].label_reg += inliner->num_locals;
}
}
/* If the inliner has handlers in effect at the point of the call that we
* are inlining, then we duplicate those and place them surrounding the
* inlinee, but with the goto still pointing to the original location.
* This means that we can still do a linear scan when searching for an
* exception handler, and don't have to try the (costly and fiddly) matter
* of trying to traverse the post-inlined call chain. */
if (inliner->sf->body.num_handlers) {
/* Walk inliner looking for handlers in effect at the point we hit the
* invoke instruction we're currently inlining; also record all of the
* instructions where the handler "goto" annotation lives. */
MVMuint32 orig_handlers = inliner->sf->body.num_handlers;
MVMuint8 *active = MVM_spesh_alloc(tc, inliner, orig_handlers);
MVMSpeshIns **handler_goto_ins = MVM_spesh_alloc(tc, inliner,
orig_handlers * sizeof(MVMSpeshIns *));
MVMint32 found_invoke = 0;
bb = inliner->entry;
while (bb && !bb->inlined) {
MVMSpeshIns *ins = bb->first_ins;
while (ins) {
MVMSpeshAnn *ann = ins->annotations;
while (ann) {
if (ann->type == MVM_SPESH_ANN_FH_GOTO) {
if (ann->data.frame_handler_index < orig_handlers)
handler_goto_ins[ann->data.frame_handler_index] = ins;
}
else if (!found_invoke) {
/* Only update these to the point we found the invoke
* being inlined, so it serves as a snapshot of what
* is active. */
if (ann->type == MVM_SPESH_ANN_FH_START)
active[ann->data.frame_handler_index] = 1;
else if (ann->type == MVM_SPESH_ANN_FH_END)
active[ann->data.frame_handler_index] = 0;
}
ann = ann->next;
}
if (ins == invoke_ins) {
/* Found it; see if we have any handlers active. If so, we
* will continue walking to collect goto annotations. */
found_invoke = 1;
for (i = 0; i < orig_handlers; i++)
active_handlers_at_invoke += active[i];
if (!active_handlers_at_invoke)
break;
}
ins = ins->next;
}
if (found_invoke && !active_handlers_at_invoke)
break;
bb = bb->linear_next;
}
/* If we found handlers active at the point of invoke, duplicate them
* in the handlers table and add annotations. */
if (active_handlers_at_invoke) {
MVMuint32 insert_pos = inliner->num_handlers + inlinee->num_handlers;
resize_handlers_table(tc, inliner, insert_pos + active_handlers_at_invoke);
for (i = orig_handlers - 1; i >= 0; i--) {
if (active[i]) {
/* Add handler start annotation to first inlinee instruction. */
MVMSpeshAnn *new_ann = MVM_spesh_alloc(tc, inliner, sizeof(MVMSpeshAnn));
new_ann->type = MVM_SPESH_ANN_FH_START;
new_ann->data.frame_handler_index = insert_pos;
new_ann->next = inlinee_first_bb->first_ins->annotations;
inlinee_first_bb->first_ins->annotations = new_ann;
/* Add handler end annotation to last inlinee instruction. */
new_ann = MVM_spesh_alloc(tc, inliner, sizeof(MVMSpeshAnn));
new_ann->type = MVM_SPESH_ANN_FH_END;
new_ann->data.frame_handler_index = insert_pos;
new_ann->next = inlinee_last_bb->last_ins->annotations;
inlinee_last_bb->last_ins->annotations = new_ann;
/* Add handler goto annotation to original target in inliner. */
new_ann = MVM_spesh_alloc(tc, inliner, sizeof(MVMSpeshAnn));
new_ann->type = MVM_SPESH_ANN_FH_GOTO;
new_ann->data.frame_handler_index = insert_pos;
new_ann->next = handler_goto_ins[i]->annotations;
handler_goto_ins[i]->annotations = new_ann;
/* Copy handler entry to new slot. */
memcpy(inliner->handlers + insert_pos, inliner->handlers + i,
sizeof(MVMFrameHandler));
insert_pos++;
}
}
}
}
/* Update total locals, lexicals, basic blocks, and handlers of the
* inliner. */
inliner->num_bbs += inlinee->num_bbs - 1;
inliner->num_locals += inlinee->num_locals;
inliner->num_lexicals += inlinee->num_lexicals;
inliner->num_handlers += inlinee->num_handlers + active_handlers_at_invoke;
}