/** Write just the specified regions back to the file */
void
SgAsmGenericSection::unparse(std::ostream &f, const ExtentMap &map) const
{
for (ExtentMap::const_iterator i=map.begin(); i!=map.end(); ++i) {
Extent e = i->first;
assert(e.first()+e.size() <= get_size());
const unsigned char *extent_data;
if (e.first() >= p_data.size()) {
extent_data = NULL;
} else if (e.first() + e.size() > p_data.size()) {
extent_data = &p_data[e.first()];
} else {
extent_data = &p_data[e.first()];
}
if (extent_data)
write(f, e.first(), e.size(), extent_data);
}
}
void
SgAsmGenericFile::shift_extend(SgAsmGenericSection *s, rose_addr_t sa, rose_addr_t sn, AddressSpace space, Elasticity elasticity)
{
ROSE_ASSERT(s!=NULL);
ROSE_ASSERT(s->get_file()==this);
ROSE_ASSERT((space & (ADDRSP_FILE|ADDRSP_MEMORY)) != 0);
const bool debug = false;
static size_t ncalls=0;
char p[256];
if (debug) {
const char *space_s="unknown";
if (space & ADDRSP_FILE) {
space_s = "file";
} else if (space & ADDRSP_MEMORY) {
space_s = "memory";
}
sprintf(p, "SgAsmGenericFile::shift_extend[%" PRIuPTR "]: ", ncalls++);
fprintf(stderr, "%s -- START --\n", p);
fprintf(stderr, "%s S = [%d] \"%s\"\n", p, s->get_id(), s->get_name()->get_string(true).c_str());
fprintf(stderr, "%s %s Sa=0x%08" PRIx64 " (%" PRIu64 "), Sn=0x%08" PRIx64 " (%" PRIu64 ")\n",
p, space_s, sa, sa, sn, sn);
fprintf(stderr, "%s elasticity = %s\n", p, (ELASTIC_NONE==elasticity ? "none" :
ELASTIC_UNREF==elasticity ? "unref" :
ELASTIC_HOLE==elasticity ? "unref+holes" :
"unknown"));
}
/* No-op case */
if (0==sa && 0==sn) {
if (debug) {
fprintf(stderr, "%s No change necessary.\n", p);
fprintf(stderr, "%s -- END --\n", p);
}
return;
}
bool filespace = (space & ADDRSP_FILE)!=0;
bool memspace = (space & ADDRSP_MEMORY)!=0;
rose_addr_t align=1, aligned_sa, aligned_sasn;
SgAsmGenericSectionPtrList neighbors, villagers;
ExtentMap amap; /* address mappings for all extents */
Extent sp;
/* Get a list of all sections that may need to be adjusted. */
SgAsmGenericSectionPtrList all;
switch (elasticity) {
case ELASTIC_NONE:
case ELASTIC_UNREF:
all = filespace ? get_sections() : get_mapped_sections();
break;
case ELASTIC_HOLE:
all = filespace ? get_sections(false) : get_mapped_sections();
break;
}
if (debug) {
fprintf(stderr, "%s Following sections are in 'all' set:\n", p);
for (size_t i=0; i<all.size(); i++) {
Extent ep;
if (filespace) {
ep = all[i]->get_file_extent();
} else {
ROSE_ASSERT(all[i]->is_mapped());
ep = all[i]->get_mapped_preferred_extent();
}
fprintf(stderr, "%s 0x%08" PRIx64 " 0x%08" PRIx64 " 0x%08" PRIx64 " [%d] \"%s\"\n",
p, ep.relaxed_first(), ep.size(), ep.relaxed_first()+ep.size(), all[i]->get_id(),
all[i]->get_name()->get_string(true).c_str());
}
}
for (size_t pass=0; pass<2; pass++) {
if (debug) {
fprintf(stderr, "%s -- %s --\n",
p, 0==pass?"FIRST PASS":"******");
}
/* S offset and size in file or memory address space */
if (filespace) {
sp = s->get_file_extent();
} else if (!memspace || !s->is_mapped()) {
return; /*nothing to do*/
} else {
sp = s->get_mapped_preferred_extent();
}
/* Build address map */
for (size_t i=0; i<all.size(); i++) {
if (filespace) {
amap.insert(all[i]->get_file_extent());
} else {
ROSE_ASSERT(all[i]->is_mapped());
amap.insert(all[i]->get_mapped_preferred_extent());
}
}
if (debug) {
fprintf(stderr, "%s Address map:\n", p);
amap.dump_extents(stderr, (std::string(p)+" ").c_str(), "amap");
fprintf(stderr, "%s Extent of S:\n", p);
fprintf(stderr, "%s start=0x%08" PRIx64 " size=0x%08" PRIx64 " end=0x%08" PRIx64 "\n",
p, sp.relaxed_first(), sp.size(), sp.relaxed_first()+sp.size());
}
/* Neighborhood (nhs) of S is a single extent. However, if S is zero size then nhs might be empty. The neighborhood of
* S is S plus all sections that overlap with S and all sections that are right-contiguous with S. */
ExtentMap nhs_map;
for (ExtentMap::iterator amapi=amap.begin(); amapi!=amap.end(); ++amapi) {
if (amapi->first.relaxed_first() <= sp.relaxed_first()+sp.size() &&
amapi->first.relaxed_first()+amapi->first.size() > sp.relaxed_first())
nhs_map.insert(amapi->first, amapi->second);
}
if (debug) {
fprintf(stderr, "%s Neighborhood of S:\n", p);
nhs_map.dump_extents(stderr, (std::string(p)+" ").c_str(), "nhs_map");
}
Extent nhs;
if (nhs_map.size()>0) {
assert(nhs_map.nranges()==1);
nhs = nhs_map.begin()->first;
} else {
nhs = sp;
}
/* What sections are in the neighborhood (including S), and right of the neighborhood? */
neighbors.clear(); /*sections in neighborhood*/
neighbors.push_back(s);
villagers.clear(); /*sections right of neighborhood*/
if (debug)
fprintf(stderr, "%s Ignoring left (L) sections:\n", p);
for (size_t i=0; i<all.size(); i++) {
SgAsmGenericSection *a = all[i];
if (a==s) continue; /*already pushed onto neighbors*/
Extent ap;
if (filespace) {
ap = a->get_file_extent();
} else if (!a->is_mapped()) {
continue;
} else {
ap = a->get_mapped_preferred_extent();
}
switch (ExtentMap::category(ap, nhs)) {
case 'L':
if (debug)
fprintf(stderr, "%s L 0x%08" PRIx64 " 0x%08" PRIx64 " 0x%08" PRIx64 " [%d] \"%s\"\n",
p, ap.relaxed_first(), ap.size(), ap.relaxed_first()+ap.size(),
a->get_id(), a->get_name()->get_string(true).c_str());
break;
case 'R':
if (ap.relaxed_first()==nhs.relaxed_first()+nhs.size() && 0==ap.size()) {
/* Empty sections immediately right of the neighborhood of S should actually be considered part of the
* neighborhood rather than right of it. */
neighbors.push_back(a);
} else if (elasticity!=ELASTIC_NONE) {
/* If holes are elastic then treat things right of the hole as being part of the right village; otherwise
* add those sections to the neighborhood of S even though they fall outside 'nhs' (it's OK because this
* partitioning of sections is the only thing we use 'nhs' for anyway. */
villagers.push_back(a);
} else if ('L'==ExtentMap::category(ap, sp)) {
/*ignore sections left of S*/
} else {
neighbors.push_back(a);
}
break;
default:
if ('L'!=ExtentMap::category(ap, sp)) /*ignore sections left of S*/
neighbors.push_back(a);
break;
}
}
if (debug) {
fprintf(stderr, "%s Neighbors:\n", p);
for (size_t i=0; i<neighbors.size(); i++) {
SgAsmGenericSection *a = neighbors[i];
Extent ap = filespace ? a->get_file_extent() : a->get_mapped_preferred_extent();
rose_addr_t align = filespace ? a->get_file_alignment() : a->get_mapped_alignment();
char cat = ExtentMap::category(ap, sp);
fprintf(stderr, "%s %c %c0x%08" PRIx64 " 0x%08" PRIx64 " 0x%08" PRIx64,
p, cat, 0==ap.relaxed_first() % (align?align:1) ? ' ' : '!',
ap.relaxed_first(), ap.size(), ap.relaxed_first()+ap.size());
if (strchr("RICE", cat)) {
fprintf(stderr, " align=0x%08" PRIx64, align);
} else {
fputs(" ", stderr);
}
fprintf(stderr, " [%2d] \"%s\"\n", a->get_id(), a->get_name()->get_string(true).c_str());
}
if (villagers.size()>0) fprintf(stderr, "%s Villagers:\n", p);
for (size_t i=0; i<villagers.size(); i++) {
SgAsmGenericSection *a = villagers[i];
Extent ap = filespace ? a->get_file_extent() : a->get_mapped_preferred_extent();
rose_addr_t align = filespace ? a->get_file_alignment() : a->get_mapped_alignment();
fprintf(stderr, "%s %c %c0x%08" PRIx64 " 0x%08" PRIx64 " 0x%08" PRIx64,
p, ExtentMap::category(ap, sp), /*cat should always be R*/
0==ap.relaxed_first() % (align?align:1) ? ' ' : '!',
ap.relaxed_first(), ap.size(), ap.relaxed_first()+ap.size());
fputs(" ", stderr);
fprintf(stderr, " [%2d] \"%s\"\n", a->get_id(), a->get_name()->get_string(true).c_str());
}
}
/* Adjust Sa to satisfy all alignment constraints in neighborhood(S) for sections that will move (cats R, I, C, and E). */
align = 1;
for (size_t i=0; i<neighbors.size(); i++) {
SgAsmGenericSection *a = neighbors[i];
Extent ap = filespace ? a->get_file_extent() : a->get_mapped_preferred_extent();
if (strchr("RICE", ExtentMap::category(ap, sp))) {
rose_addr_t x = filespace ? a->get_file_alignment() : a->get_mapped_alignment();
#if BOOST_VERSION < 106900
align = boost::math::lcm(align, x?x:1); // deprecated in boost-1.69.0
#else
align = boost::integer::lcm(align, x?x:1); // not present before boost-1.60.0
#endif
}
}
aligned_sa = (sa/align + (sa%align?1:0))*align;
aligned_sasn = ((sa+sn)/align + ((sa+sn)%align?1:0))*align;
if (debug) {
fprintf(stderr, "%s Alignment LCM = 0x%08" PRIx64 " (%" PRIu64 ")\n", p, align, align);
fprintf(stderr, "%s Aligned Sa = 0x%08" PRIx64 " (%" PRIu64 ")\n", p, aligned_sa, aligned_sa);
fprintf(stderr, "%s Aligned Sa+Sn = 0x%08" PRIx64 " (%" PRIu64 ")\n", p, aligned_sasn, aligned_sasn);
}
/* Are there any sections to the right of neighborhood(S)? If so, find the one with the lowest start address and use
* that to define the size of the hole right of neighborhood(S). */
if (0==villagers.size()) break;
SgAsmGenericSection *after_hole = NULL;
Extent hp(0, 0);
for (size_t i=0; i<villagers.size(); i++) {
SgAsmGenericSection *a = villagers[i];
Extent ap = filespace ? a->get_file_extent() : a->get_mapped_preferred_extent();
if (!after_hole || ap.relaxed_first()<hp.relaxed_first()) {
after_hole = a;
hp = ap;
}
}
ROSE_ASSERT(after_hole);
ROSE_ASSERT(hp.relaxed_first() > nhs.relaxed_first()+nhs.size());
rose_addr_t hole_size = hp.relaxed_first() - (nhs.relaxed_first()+nhs.size());
if (debug) {
fprintf(stderr, "%s hole size = 0x%08" PRIx64 " (%" PRIu64 "); need 0x%08" PRIx64 " (%" PRIu64 "); %s\n",
p, hole_size, hole_size, aligned_sasn, aligned_sasn,
hole_size>=aligned_sasn ? "large enough" : "not large enough");
}
if (hole_size >= aligned_sasn) break;
rose_addr_t need_more = aligned_sasn - hole_size;
/* Hole is not large enough. We need to recursively move things that are right of our neighborhood, then recompute the
* all-sections address map and neighborhood(S). */
ROSE_ASSERT(0==pass); /*logic problem since the recursive call should have enlarged the hole enough*/
if (debug) {
fprintf(stderr, "%s Calling recursively to increase hole size by 0x%08" PRIx64 " (%" PRIu64 ") bytes\n",
p, need_more, need_more);
}
shift_extend(after_hole, need_more, 0, space, elasticity);
if (debug) fprintf(stderr, "%s Returned from recursive call\n", p);
}
/* Consider sections that are in the same neighborhood as S */
if (debug) fprintf(stderr, "%s -- ADJUSTING --\n", p);
bool resized_mem = false;
for (size_t i=0; i<neighbors.size(); i++) {
SgAsmGenericSection *a = neighbors[i];
Extent ap = filespace ? a->get_file_extent() : a->get_mapped_preferred_extent();
switch (ExtentMap::category(ap, sp)) {
case 'L':
break;
case 'R':
if (filespace) {
a->set_offset(a->get_offset()+aligned_sasn);
} else {
a->set_mapped_preferred_rva(a->get_mapped_preferred_rva()+aligned_sasn);
}
break;
case 'C': /*including S itself*/
case 'E':
if (filespace) {
a->set_offset(a->get_offset()+aligned_sa);
a->set_size(a->get_size()+sn);
if (memspace && !resized_mem && a->is_mapped()) {
shift_extend(a, 0, sn, ADDRSP_MEMORY, elasticity);
resized_mem = true;
}
} else {
a->set_mapped_preferred_rva(a->get_mapped_preferred_rva()+aligned_sa);
a->set_mapped_size(a->get_mapped_size()+sn);
}
break;
case 'O':
if (ap.relaxed_first()==sp.relaxed_first()) {
if (filespace) {
a->set_offset(a->get_offset()+aligned_sa);
a->set_size(a->get_size()+sn);
} else {
a->set_mapped_preferred_rva(a->get_mapped_preferred_rva()+aligned_sa);
a->set_mapped_size(a->get_mapped_size()+sn);
}
} else {
if (filespace) {
a->set_size(a->get_size()+aligned_sasn);
if (memspace && !resized_mem && a->is_mapped()) {
shift_extend(a, 0, aligned_sasn, ADDRSP_MEMORY, elasticity);
resized_mem = true;
}
} else {
a->set_mapped_size(a->get_mapped_size()+aligned_sasn);
}
}
break;
case 'I':
if (filespace) {
a->set_offset(a->get_offset()+aligned_sa);
} else {
a->set_mapped_preferred_rva(a->get_mapped_preferred_rva()+aligned_sa);
}
break;
case 'B':
if (filespace) {
a->set_size(a->get_size()+sn);
if (memspace && !resized_mem && a->is_mapped()) {
shift_extend(a, 0, sn, ADDRSP_MEMORY, elasticity);
resized_mem = true;
}
} else {
a->set_mapped_size(a->get_size()+sn);
}
break;
default:
ROSE_ASSERT(!"invalid extent category");
break;
}
if (debug) {
const char *space_name = filespace ? "file" : "mem";
rose_addr_t x = filespace ? a->get_file_alignment() : a->get_mapped_alignment();
fprintf(stderr, "%s %4s-%c %c0x%08" PRIx64 " 0x%08" PRIx64 " 0x%08" PRIx64,
p, space_name, ExtentMap::category(ap, sp),
0==ap.relaxed_first()%(x?x:1)?' ':'!',
ap.relaxed_first(), ap.size(), ap.relaxed_first()+ap.size());
Extent newap = filespace ? a->get_file_extent() : a->get_mapped_preferred_extent();
fprintf(stderr, " -> %c0x%08" PRIx64 " 0x%08" PRIx64 " 0x%08" PRIx64,
0==newap.relaxed_first()%(x?x:1)?' ':'!',
newap.relaxed_first(), newap.size(), newap.relaxed_first()+newap.size());
fprintf(stderr, " [%2d] \"%s\"\n", a->get_id(), a->get_name()->get_string(true).c_str());
}
}
if (debug) fprintf(stderr, "%s -- END --\n", p);
}
AddressIntervalSet toAddressIntervalSet(const ExtentMap &x) {
AddressIntervalSet retval;
for (ExtentMap::const_iterator iter=x.begin(); iter!=x.end(); ++iter)
retval.insert(toAddressInterval(iter->first));
return retval;
}
Partitioner::RegionStats *
Partitioner::region_statistics(const ExtentMap &addresses)
{
RegionStats *stats = new_region_stats();
assert(stats!=NULL);
size_t nbytes = addresses.size();
if (0==nbytes)
return stats;
stats->add_sample(RegionStats::RA_NBYTES, nbytes);
ExtentMap not_addresses = addresses.invert<ExtentMap>();
Disassembler::AddressSet worklist; // addresses waiting to be disassembled recursively
InstructionMap insns_found; // all the instructions we found herein
ExtentMap insns_extent; // memory used by the instructions we've found
ExtentMap pending = addresses; // addresses we haven't looked at yet
/* Undirected local control flow graph used to count connected components */
typedef boost::adjacency_list<boost::vecS, boost::vecS, boost::undirectedS> CFG;
typedef boost::graph_traits<CFG>::vertex_descriptor CFGVertex;
typedef std::map<rose_addr_t, CFGVertex> Addr2Vertex;
CFG cfg;
Addr2Vertex va2id;
/* Statistics */
size_t nstarts=0; // number of times the recursive disassembler was started
size_t nfails=0; // number of disassembler failures
size_t noverlaps=0; // instructions overlapping with a previously found instruction
size_t nincomplete=0; // number of instructions with unknown successors
size_t nfallthrough=0; // number of branches to fall-through address within our "addresses"
size_t ncalls=0; // number of function calls outside our "addresses"
size_t nnoncalls=0; // number of branches to non-functions outside our "addresses"
size_t ninternal=0; // number of non-fallthrough internal branches
while (!pending.empty()) {
rose_addr_t start_va = pending.min();
worklist.insert(start_va);
++nstarts;
while (!worklist.empty()) {
rose_addr_t va = *worklist.begin();
worklist.erase(worklist.begin());
/* Obtain (disassemble) the instruction and make sure it falls entirely within the "addresses" */
Instruction *insn = find_instruction(va);
if (!insn) {
++nfails;
pending.erase(Extent(va));
continue;
}
Extent ie(va, insn->get_size());
if (not_addresses.overlaps(ie)) {
++nfails;
pending.erase(Extent(va));
continue;
}
/* The disassembler can also return an "unknown" instruction when failing, depending on how it is invoked. */
if (insn->node->is_unknown()) {
++nfails;
pending.erase(Extent(va, insn->get_size()));
continue;
}
insns_found.insert(std::make_pair(va, insn));
rose_addr_t fall_through_va = va + insn->get_size();
/* Does this instruction overlap with any we've already found? */
if (insns_extent.overlaps(ie))
++noverlaps;
pending.erase(Extent(va, insn->get_size()));
insns_extent.insert(ie);
/* Find instruction successors by looking only at the instruction itself. This is simpler, but less rigorous
* method than finding successors a basic block at a time. For instance, we'll find both sides of a branch
* instruction even if the more rigorous method determined that one side or the other is always taken. But this is
* probably what we want here anyway for determining whether something looks like code. */
bool complete;
Disassembler::AddressSet succs = insn->get_successors(&complete);
if (!complete)
++nincomplete;
/* Add instruction as vertex to CFG */
std::pair<Addr2Vertex::iterator, bool> inserted = va2id.insert(std::make_pair(va, va2id.size()));
if (inserted.second) {
CFGVertex vertex __attribute__((unused)) = add_vertex(cfg);
assert(vertex==inserted.first->second);
}
/* Classify the various successors. */
for (Disassembler::AddressSet::const_iterator si=succs.begin(); si!=succs.end(); ++si) {
rose_addr_t succ_va = *si;
if (succ_va==fall_through_va) {
++nfallthrough;
if (pending.find(succ_va)!=pending.end())
worklist.insert(succ_va);
/* Add edge to CFG graph */
va2id.insert(std::make_pair(succ_va, va2id.size()));
add_edge(va2id[va], va2id[succ_va], cfg);
} else if (addresses.find(succ_va)==addresses.end()) {
/* A non-fallthrough branch to something outside this memory region */
if (functions.find(succ_va)!=functions.end()) {
/* A branch to a function entry point we've previously discovered. */
++ncalls;
} else {
++nnoncalls;
}
} else {
/* A non-fallthrough branch to something in our address range. */
++ninternal;
if (pending.find(succ_va)!=pending.end())
worklist.insert(succ_va);
/* Add edge to CFG graph */
va2id.insert(std::make_pair(succ_va, va2id.size()));
add_edge(va2id[va], va2id[succ_va], cfg);
}
}
}
}
/* Statistics */
stats->add_sample(RegionStats::RA_NFAILS, nfails);
stats->add_sample(RegionStats::RA_NINSNS, insns_found.size());
stats->add_sample(RegionStats::RA_NOVERLAPS, noverlaps);
stats->add_sample(RegionStats::RA_NSTARTS, nstarts);
stats->add_sample(RegionStats::RA_NCOVERAGE, insns_extent.size());
stats->add_sample(RegionStats::RA_NINCOMPLETE, nincomplete);
stats->add_sample(RegionStats::RA_NBRANCHES, ncalls+nnoncalls+ninternal);
stats->add_sample(RegionStats::RA_NCALLS, ncalls);
stats->add_sample(RegionStats::RA_NNONCALLS, nnoncalls);
stats->add_sample(RegionStats::RA_NINTERNAL, ninternal);
stats->add_sample(RegionStats::RA_NICFGEDGES, ninternal + nfallthrough);
stats->add_sample(RegionStats::RA_NIUNIQUE, count_kinds(insns_found));
stats->add_sample(RegionStats::RA_NPRIV, count_privileged(insns_found));
stats->add_sample(RegionStats::RA_NFLOAT, count_floating_point(insns_found));
double regsz, regvar;
stats->add_sample(RegionStats::RA_NREGREFS, count_registers(insns_found, ®sz, ®var));
stats->add_sample(RegionStats::RA_REGSZ, regsz);
stats->add_sample(RegionStats::RA_REGVAR, regvar);
/* Count the number of connected components in the undirected CFG */
if (!va2id.empty()) {
std::vector<int> component(num_vertices(cfg));
stats->add_sample(RegionStats::RA_NCOMPS, connected_components(cfg, &component[0]));
}
stats->compute_ratios();
return stats;
}