SWDEV-520093+SWDEV-521137+SWDEV-523996: Fix for missing AMDGPU_HSA_V6 (#20)

* Adding AMDGPU_HSA_V6

* Fix segment

* Fix conflicts

* ident fix

* Symbol lookup fix

---------

Co-authored-by: Giovanni Baraldi <gbaraldi@amd.com>

[ROCm/rocprofiler commit: 6301f46d13]
This commit is contained in:
Baraldi, Giovanni
2025-03-28 18:02:49 +01:00
committed by GitHub
parent be0f948018
commit 1a3e33774c
5 changed files with 69 additions and 145 deletions
@@ -119,7 +119,7 @@ CodeObjDecoderComponent::CodeObjDecoderComponent(
Dwarf_Off cu_offset{0}, next_offset;
size_t header_size;
std::unordered_set<uint64_t> used_addrs;
std::map<uint64_t, std::string> used_addrs;
while (!dwarf_nextcu(dbg.get(), cu_offset, &next_offset, &header_size, nullptr, nullptr,
nullptr)) {
@@ -142,32 +142,33 @@ CodeObjDecoderComponent::CodeObjDecoderComponent(
if (used_addrs.find(addr) != used_addrs.end())
{
size_t pos = m_line_number_map.lower_bound(addr);
m_line_number_map.data()[pos].str += ' ' + dwarf_line;
used_addrs.at(addr) += ' ' + dwarf_line;
continue;
}
used_addrs.insert(addr);
m_line_number_map.insert(DSourceLine{addr, 0, std::move(dwarf_line)});
used_addrs.emplace(addr, std::move(dwarf_line));
}
}
cu_offset = next_offset;
}
auto it = used_addrs.begin();
if(it != used_addrs.end())
{
while(std::next(it) != used_addrs.end())
{
uint64_t delta = std::next(it)->first - it->first;
auto segment = address_range_t{it->first, delta, 0};
m_line_number_map.emplace(segment, std::move(it->second));
it++;
}
auto segment = address_range_t{it->first, codeobj_size - it->first, 0};
m_line_number_map.emplace(segment, std::move(it->second));
}
}
// Can throw
disassembly = std::make_unique<DisassemblyInstance>(codeobj_data, codeobj_size, gpu_id);
if (m_line_number_map.size())
{
size_t total_size = 0;
for (size_t i=0; i<m_line_number_map.size()-1; i++)
{
size_t s = m_line_number_map.get(i+1).vaddr - m_line_number_map.get(i).vaddr;
m_line_number_map.data()[i].size = s;
total_size += s;
}
m_line_number_map.back().size = std::max(total_size, codeobj_size) - total_size;
}
try {
m_symbol_map = disassembly->GetKernelMap(); // Can throw
} catch(...) {}
@@ -209,10 +210,8 @@ CodeObjDecoderComponent::disassemble_instruction(uint64_t faddr, uint64_t vaddr)
const char* cpp_line = nullptr;
try {
const DSourceLine& it = m_line_number_map.find_obj(vaddr);
cpp_line = it.str.data();
} catch(...) {}
auto it = m_line_number_map.find({vaddr, 0, 0});
if(it != m_line_number_map.end()) cpp_line = it->second.data();
size_t size = disassembly->ReadInstruction(faddr, vaddr, cpp_line);
return {disassembly->last_instruction, size};
@@ -57,7 +57,7 @@ public:
int m_fd;
cached_ordered_vector<DSourceLine> m_line_number_map;
std::map<address_range_t, std::string> m_line_number_map{};
std::map<uint64_t, SymbolInfo> m_symbol_map{};
std::string m_uri;
@@ -176,7 +176,7 @@ public:
{
this->Super::addDecoder(filepath, id, loadbase, memsize, gpu_id);
auto ptr = decoders.at(id);
table.insert({ptr->begin(), static_cast<uint32_t>(ptr->size()), id, 0});
table.insert(address_range_t{ptr->begin(), static_cast<uint32_t>(ptr->size()), id});
}
virtual bool removeDecoder(uint32_t id, uint64_t loadbase)
@@ -186,8 +186,8 @@ public:
instruction_info_t get(uint64_t vaddr)
{
auto& addr_range = table.find_codeobj_in_range(vaddr);
return get(addr_range.id, vaddr - addr_range.vbegin);
auto addr_range = table.find_codeobj_in_range(vaddr);
return get(addr_range.id, vaddr - addr_range.addr);
}
instruction_info_t get(uint32_t id, uint64_t offset) { return this->Super::get(id, offset); }
@@ -303,7 +303,8 @@ std::optional<uint64_t> DisassemblyInstance::va2fo(uint64_t va)
CHECK_VA2FO(
e_ident[EI_ABIVERSION] == 2 || // ELFABIVERSION_AMDGPU_HSA_V4
e_ident[EI_ABIVERSION] == 3, "unexpected ei_abiversion"); // ELFABIVERSION_AMDGPU_HSA_V5
e_ident[EI_ABIVERSION] == 3 || // ELFABIVERSION_AMDGPU_HSA_V5
e_ident[EI_ABIVERSION] == 4, "unexpected ei_abiversion"); // ELFABIVERSION_AMDGPU_HSA_V6
Elf64_Ehdr *ehdr = (Elf64_Ehdr*)buffer.data();
CHECK_VA2FO(ehdr, "ehdr is nullptr");
@@ -356,7 +357,8 @@ std::vector<std::pair<uint64_t, uint64_t>> DisassemblyInstance::getSegments()
CHECK_VA2FO(
e_ident[EI_ABIVERSION] == 2 || // ELFABIVERSION_AMDGPU_HSA_V4
e_ident[EI_ABIVERSION] == 3, "unexpected ei_abiversion"); // ELFABIVERSION_AMDGPU_HSA_V5
e_ident[EI_ABIVERSION] == 3 || // ELFABIVERSION_AMDGPU_HSA_V5
e_ident[EI_ABIVERSION] == 4, "unexpected ei_abiversion"); // ELFABIVERSION_AMDGPU_HSA_V6
Elf64_Ehdr *ehdr = (Elf64_Ehdr*)buffer.data();
CHECK_VA2FO(ehdr, "ehdr is nullptr");
+41 -118
View File
@@ -23,132 +23,55 @@
#include <vector>
#include <iostream>
#include <random>
#include <unordered_set>
#include <set>
#include <algorithm>
template<typename Type>
class ordered_vector : public std::vector<Type>
{
using Super = std::vector<Type>;
public:
void insert(const Type& elem)
{
size_t loc = lower_bound(elem.begin());
if (this->size() && get(loc).begin() < elem.begin())
loc ++;
this->Super::insert(this->begin()+loc, elem);
}
bool remove(const Type& elem)
{
if (!this->size()) return false;
size_t loc = lower_bound(elem.begin());
if (get(loc) != elem) return false;
this->Super::erase(this->begin()+loc);
return true;
}
bool remove(uint64_t elem_begin)
{
if (!this->size()) return false;
size_t loc = lower_bound(elem_begin);
if (get(loc).begin() != elem_begin) return false;
this->Super::erase(this->begin()+loc);
return true;
}
size_t lower_bound(size_t addr) const
{
if (!this->size()) return 0;
return binary_search(addr, 0, this->size()-1);
}
size_t binary_search(size_t addr, size_t s, size_t e) const
{
if (s >= e)
return s;
else if (s+1 == e)
return (get(e).begin() <= addr) ? e : s;
size_t mid = (s+e)/2;
if (get(mid).begin() <= addr)
return binary_search(addr, mid, e);
else
return binary_search(addr, s, mid);
}
const Type& get(size_t i) const { return this->operator[](i); }
};
/**
* @brief Finds a candidate codeobj for the given vaddr
*/
template<typename Type>
class cached_ordered_vector : public ordered_vector<Type>
{
using Super = ordered_vector<Type>;
public:
cached_ordered_vector<Type>() { reset(); }
const Type& find_obj(uint64_t addr)
{
if (testCache(addr))
return get(cached_segment);
size_t lb = this->lower_bound(addr);
if (lb >= this->size() || !get(lb).inrange(addr))
throw std::string("segment addr out of range");
cached_segment = lb;
return get(cached_segment);
}
uint64_t find_addr(uint64_t addr) {
return find_obj(addr).begin();
}
bool testCache(uint64_t addr) const {
return this->cached_segment < this->size() && get(cached_segment).inrange(addr);
}
const Type& get(size_t index) const { return this->data()[index]; }
void insert(const Type& elem) { this->Super::insert(elem); }
void insert_list(std::vector<Type> arange)
{
for (auto& elem : arange) push_back(elem);
std::sort(
this->begin(),
this->end(),
[](const Type& a, const Type& b) { return a.begin() < b.begin(); }
);
};
void reset() { cached_segment = ~0; }
void clear() { reset(); this->Super::clear(); }
bool remove(uint64_t addr) { reset(); return this->Super::remove(addr); }
private:
size_t cached_segment = ~0;
};
struct address_range_t
{
uint64_t vbegin;
uint32_t size;
uint32_t id;
uint32_t offset;
uint64_t addr{0};
uint64_t size{0};
uint64_t id{0};
bool operator<(const address_range_t& other) const { return vbegin < other.vbegin; }
bool inrange(uint64_t addr) const { return addr >= vbegin && addr < vbegin+size; };
uint64_t begin() const { return vbegin; }
bool operator==(const address_range_t& other) const
{
return (addr >= other.addr && addr < other.addr + other.size) ||
(other.addr >= addr && other.addr < addr + size);
}
bool operator<(const address_range_t& other) const
{
if(*this == other) return false;
return addr < other.addr;
}
bool inrange(uint64_t _addr) const { return addr <= _addr && addr + size > _addr; };
};
/**
* @brief Finds a candidate codeobj for the given vaddr
*/
class CodeobjTableTranslator : public cached_ordered_vector<address_range_t>
*/
class CodeobjTableTranslator : public std::set<address_range_t>
{
public:
const address_range_t& find_codeobj_in_range(uint64_t addr) { return this->find_obj(addr); }
using Super = std::set<address_range_t>;
public:
address_range_t find_codeobj_in_range(uint64_t addr)
{
if(!cached_segment.inrange(addr))
{
auto it = this->find(address_range_t{addr, 0, 0});
if(it == this->end()) throw std::exception();
cached_segment = *it;
}
return cached_segment;
}
void clear_cache() { cached_segment = {}; }
bool remove(const address_range_t& range)
{
clear_cache();
return this->erase(range) != 0;
}
bool remove(uint64_t addr) { return remove(address_range_t{addr, 0, 0}); }
private:
address_range_t cached_segment{};
};
+1 -1
View File
@@ -155,7 +155,7 @@ class RegisterWatchList:
class PCTranslator:
def __init__(self, insts, code, raw_code, reverse_map, codeservice):
self.codeservice = codeservice
self.insts = insts
self.addrmap = {c[-3] : (c, self.codeservice.GetInstruction(c[-3])[3]) for c in code if c[-3] > 0}