void dumpPrettyIR(llvm::Function* f) {
// f->getParent()->dump();
// return;
std::unique_ptr<llvm::Module> tmp_module(llvm::CloneModule(f->getParent()));
// std::unique_ptr<llvm::Module> tmp_module(new llvm::Module("tmp", g.context));
llvm::Function* new_f = tmp_module->begin();
llvm::ValueToValueMapTy VMap;
llvm::RemapFlags flags = llvm::RF_None;
llvm::ValueMapTypeRemapper* type_remapper = NULL;
PrettifyingMaterializer materializer(tmp_module.get(), VMap, flags, type_remapper);
for (llvm::Function::iterator I = new_f->begin(), E = new_f->end(); I != E; ++I) {
VMap[I] = I;
}
for (llvm::inst_iterator it = inst_begin(new_f), end = inst_end(new_f); it != end; ++it) {
llvm::RemapInstruction(&*it, VMap, flags, type_remapper, &materializer);
if (llvm::IntrinsicInst* ii = llvm::dyn_cast<llvm::IntrinsicInst>(&*it)) {
if (ii->getIntrinsicID() == llvm::Intrinsic::experimental_patchpoint_i64
|| ii->getIntrinsicID() == llvm::Intrinsic::experimental_patchpoint_void
|| ii->getIntrinsicID() == llvm::Intrinsic::experimental_patchpoint_double) {
remapPatchpoint(ii);
}
} else if (llvm::InvokeInst* ii = llvm::dyn_cast<llvm::InvokeInst>(&*it)) {
if (ii->getCalledFunction() && ii->getCalledFunction()->isIntrinsic()) {
remapPatchpoint(ii);
}
}
}
tmp_module->begin()->dump();
// tmp_module->dump();
}
void dumpPrettyIR(llvm::Function* f) {
std::unique_ptr<llvm::Module> tmp_module(llvm::CloneModule(f->getParent()));
// std::unique_ptr<llvm::Module> tmp_module(new llvm::Module("tmp", g.context));
llvm::Function* new_f = tmp_module->begin();
llvm::ValueToValueMapTy VMap;
PrettifyingMaterializer materializer(tmp_module.get());
for (llvm::Function::iterator I = new_f->begin(), E = new_f->end(); I != E; ++I) {
VMap[I] = I;
}
for (llvm::inst_iterator it = inst_begin(new_f), end = inst_end(new_f); it != end; ++it) {
llvm::RemapInstruction(&*it, VMap, llvm::RF_None, NULL, &materializer);
}
tmp_module->begin()->dump();
// tmp_module->dump();
}
bool _runOnFunction(llvm::Function& f) {
Timer _t2("(sum)");
Timer _t("initializing");
initialize();
_t.split("overhead");
// f.dump();
llvm::Module* cur_module = f.getParent();
#if LLVMREV < 217548
llvm::PassManager fake_pm;
#else
llvm::legacy::PassManager fake_pm;
#endif
llvm::InlineCostAnalysis* cost_analysis = new llvm::InlineCostAnalysis();
fake_pm.add(cost_analysis);
// llvm::errs() << "doing fake run\n";
fake_pm.run(*fake_module);
// llvm::errs() << "done with fake run\n";
bool did_any_inlining = false;
// TODO I haven't gotten the callgraph-updating part of the inliner to work,
// so it's not easy to tell what callsites have been inlined into (ie added to)
// the function.
// One simple-but-not-great way to handle it is to just iterate over the entire function
// multiple times and re-inline things until we don't want to inline any more;
// NPASSES controls the maximum number of times to attempt that.
// Right now we actually don't need that, since we only inline fully-optimized
// functions (from the stdlib), and those will already have had inlining
// applied recursively.
const int NPASSES = 1;
for (int passnum = 0; passnum < NPASSES; passnum++) {
_t.split("collecting calls");
std::vector<llvm::CallSite> calls;
for (llvm::inst_iterator I = llvm::inst_begin(f), E = llvm::inst_end(f); I != E; ++I) {
llvm::CallInst* call = llvm::dyn_cast<llvm::CallInst>(&(*I));
// From Inliner.cpp:
if (!call || llvm::isa<llvm::IntrinsicInst>(call))
continue;
// I->dump();
llvm::CallSite CS(call);
llvm::Value* v = CS.getCalledValue();
llvm::ConstantExpr* ce = llvm::dyn_cast<llvm::ConstantExpr>(v);
if (!ce)
continue;
assert(ce->isCast());
llvm::ConstantInt* l_addr = llvm::cast<llvm::ConstantInt>(ce->getOperand(0));
int64_t addr = l_addr->getSExtValue();
if (addr == (int64_t)printf)
continue;
llvm::Function* f = g.func_addr_registry.getLLVMFuncAtAddress((void*)addr);
if (f == NULL) {
if (VERBOSITY()) {
printf("Giving up on inlining %s:\n",
g.func_addr_registry.getFuncNameAtAddress((void*)addr, true).c_str());
call->dump();
}
continue;
}
// We load the bitcode lazily, so check if we haven't yet fully loaded the function:
if (f->isMaterializable()) {
#if LLVMREV < 220600
f->Materialize();
#else
f->materialize();
#endif
}
// It could still be a declaration, though I think the code won't generate this case any more:
if (f->isDeclaration())
continue;
// Keep this section as a release_assert since the code-to-be-inlined, as well as the inlining
// decisions, can be different in release mode:
int op_idx = -1;
for (llvm::Argument& arg : f->args()) {
++op_idx;
llvm::Type* op_type = call->getOperand(op_idx)->getType();
if (arg.getType() != op_type) {
llvm::errs() << f->getName() << " has arg " << op_idx << " mismatched!\n";
llvm::errs() << "Given ";
op_type->dump();
llvm::errs() << " but underlying function expected ";
arg.getType()->dump();
llvm::errs() << '\n';
}
RELEASE_ASSERT(arg.getType() == call->getOperand(op_idx)->getType(), "");
}
assert(!f->isDeclaration());
CS.setCalledFunction(f);
calls.push_back(CS);
}
// assert(0 && "TODO");
// printf("%ld\n", calls.size());
bool did_inline = false;
_t.split("doing inlining");
while (calls.size()) {
llvm::CallSite cs = calls.back();
calls.pop_back();
// if (VERBOSITY("irgen.inlining") >= 1) {
// llvm::errs() << "Evaluating callsite ";
// cs->dump();
//}
llvm::InlineCost IC = cost_analysis->getInlineCost(cs, threshold);
bool do_inline = false;
if (IC.isAlways()) {
if (VERBOSITY("irgen.inlining") >= 2)
llvm::errs() << "always inline\n";
do_inline = true;
} else if (IC.isNever()) {
if (VERBOSITY("irgen.inlining") >= 2)
llvm::errs() << "never inline\n";
do_inline = false;
} else {
if (VERBOSITY("irgen.inlining") >= 2)
llvm::errs() << "Inline cost: " << IC.getCost() << '\n';
do_inline = (bool)IC;
}
if (VERBOSITY("irgen.inlining") >= 1) {
if (!do_inline)
llvm::outs() << "not ";
llvm::outs() << "inlining ";
cs->dump();
}
if (do_inline) {
static StatCounter num_inlines("num_inlines");
num_inlines.log();
// llvm::CallGraph cg(*f.getParent());
////cg.addToCallGraph(cs->getCalledFunction());
// llvm::InlineFunctionInfo InlineInfo(&cg);
llvm::InlineFunctionInfo InlineInfo;
bool inlined = llvm::InlineFunction(cs, InlineInfo, false);
did_inline = did_inline || inlined;
did_any_inlining = did_any_inlining || inlined;
// if (inlined)
// f.dump();
}
}
if (!did_inline) {
if (passnum >= NPASSES - 1 && VERBOSITY("irgen.inlining"))
printf("quitting after %d passes\n", passnum + 1);
break;
}
}
// TODO would be nice to break out here and not have to rematerialize the function;
// I think I have to do that even if no inlining happened from the "setCalledFunction" call above.
// I thought that'd just change the CS object, but maybe it changes the underlying instruction as well?
// if (!did_any_inlining)
// return false;
_t.split("remapping");
llvm::ValueToValueMapTy VMap;
for (llvm::Function::iterator I = f.begin(), E = f.end(); I != E; ++I) {
VMap[I] = I;
}
MyMaterializer materializer(cur_module);
for (llvm::inst_iterator I = llvm::inst_begin(f), E = llvm::inst_end(f); I != E; ++I) {
RemapInstruction(&(*I), VMap, llvm::RF_None, NULL, &materializer);
}
_t.split("cleaning up");
std::vector<llvm::GlobalValue*> to_remove;
for (llvm::Module::global_iterator I = cur_module->global_begin(), E = cur_module->global_end(); I != E; ++I) {
if (I->use_empty()) {
to_remove.push_back(I);
continue;
}
}
for (int i = 0; i < to_remove.size(); i++) {
to_remove[i]->eraseFromParent();
}
for (llvm::Module::iterator I = cur_module->begin(), E = cur_module->end(); I != E;) {
if (!I->isDeclaration()) {
++I;
continue;
}
if (I->use_empty()) {
I = cur_module->getFunctionList().erase(I);
} else {
++I;
}
}
return did_any_inlining;
}
