In addition to the existing support for libbfd and calling out to
an external addr2line command, add support for using libllvm directly.
This is both faster than libbfd, and can be enabled in distro builds
(the LLVM license has an explicit provision for GPLv2 compatibility).
Thus, it is set as the primary choice if available.
As an example, running perf report on a medium-size profile with
DWARF-based backtraces took 58 seconds with LLVM, 78 seconds with
libbfd, 153 seconds with external llvm-addr2line, and I got tired
and aborted the test after waiting for 55 minutes with external
bfd addr2line (which is the default for perf as compiled by distributions
today). Evidently, for this case, the bfd addr2line process needs
18 seconds (on a 5.2 GHz Zen 3) to load the .debug ELF in question,
hits the 1-second timeout and gets killed during initialization,
getting restarted anew every time. Having an in-process addr2line
makes this much more robust.
As future extensions, libllvm can be used in many other places where
we currently use libbfd or other libraries:
- Symbol enumeration (in particular, for PE binaries).
- Demangling (including non-Itanium demangling, e.g. Microsoft
or Rust).
- Disassembling (perf annotate).
However, these are much less pressing; most people don't profile
PE binaries, and perf has non-bfd paths for ELF. The same with
demangling; the default _cxa_demangle path works fine for most
users. Disassembling is coming in a later patch in the series;
however do note that while bfd objdump can be slow on large binaries,
it is possible to use --objdump=llvm-objdump to get the speed benefits.
(It appears LLVM-based demangling is very simple, should we want
that.)
Tested with LLVM 14, 15, 16, 18 and 19. For some reason, LLVM 12 was not
correctly detected using feature_check, and thus was not tested.
Signed-off-by: Steinar H. Gunderson <[email protected]>
---
tools/build/Makefile.feature | 1 +
tools/perf/Makefile.config | 15 ++++
tools/perf/builtin-version.c | 1 +
tools/perf/util/Build | 1 +
tools/perf/util/llvm-c-helpers.cpp | 130 +++++++++++++++++++++++++++++
tools/perf/util/llvm-c-helpers.h | 47 +++++++++++
tools/perf/util/srcline.c | 57 ++++++++++++-
7 files changed, 251 insertions(+), 1 deletion(-)
create mode 100644 tools/perf/util/llvm-c-helpers.cpp
create mode 100644 tools/perf/util/llvm-c-helpers.h
diff --git a/tools/build/Makefile.feature b/tools/build/Makefile.feature
index 1e2ab148d5db..278b26216254 100644
--- a/tools/build/Makefile.feature
+++ b/tools/build/Makefile.feature
@@ -136,6 +136,7 @@ FEATURE_DISPLAY ?= \
libunwind \
libdw-dwarf-unwind \
libcapstone \
+ llvm \
zlib \
lzma \
get_cpuid \
diff --git a/tools/perf/Makefile.config b/tools/perf/Makefile.config
index 7f1e016a9253..414a37f712bd 100644
--- a/tools/perf/Makefile.config
+++ b/tools/perf/Makefile.config
@@ -969,6 +969,21 @@ ifdef BUILD_NONDISTRO
endif
endif
+ifndef NO_LLVM
+ $(call feature_check,llvm)
+ ifeq ($(feature-llvm), 1)
+ CFLAGS += -DHAVE_LLVM_SUPPORT
+ CXXFLAGS += -DHAVE_LLVM_SUPPORT
+ CXXFLAGS += $(shell $(LLVM_CONFIG) --cxxflags)
+ LIBLLVM = $(shell $(LLVM_CONFIG) --libs all) $(shell $(LLVM_CONFIG) --system-libs)
+ EXTLIBS += -L$(shell $(LLVM_CONFIG) --libdir) $(LIBLLVM)
+ $(call detected,CONFIG_LLVM)
+ else
+ $(warning No libllvm found, slower source file resolution, please install llvm-devel/llvm-dev)
+ NO_LLVM := 1
+ endif
+endif
+
ifndef NO_DEMANGLE
$(call feature_check,cxa-demangle)
ifeq ($(feature-cxa-demangle), 1)
diff --git a/tools/perf/builtin-version.c b/tools/perf/builtin-version.c
index 398aa53e9e2e..94d8e13a688f 100644
--- a/tools/perf/builtin-version.c
+++ b/tools/perf/builtin-version.c
@@ -65,6 +65,7 @@ static void library_status(void)
STATUS(HAVE_LIBBFD_SUPPORT, libbfd);
STATUS(HAVE_DEBUGINFOD_SUPPORT, debuginfod);
STATUS(HAVE_LIBELF_SUPPORT, libelf);
+ STATUS(HAVE_LLVM_SUPPORT, libllvm);
STATUS(HAVE_LIBNUMA_SUPPORT, libnuma);
STATUS(HAVE_LIBNUMA_SUPPORT, numa_num_possible_cpus);
STATUS(HAVE_LIBPERL_SUPPORT, libperl);
diff --git a/tools/perf/util/Build b/tools/perf/util/Build
index da64efd8718f..32c4e5e634ed 100644
--- a/tools/perf/util/Build
+++ b/tools/perf/util/Build
@@ -226,6 +226,7 @@ perf-$(CONFIG_CXX_DEMANGLE) += demangle-cxx.o
perf-y += demangle-ocaml.o
perf-y += demangle-java.o
perf-y += demangle-rust.o
+perf-$(CONFIG_LLVM) += llvm-c-helpers.o
ifdef CONFIG_JITDUMP
perf-$(CONFIG_LIBELF) += jitdump.o
diff --git a/tools/perf/util/llvm-c-helpers.cpp b/tools/perf/util/llvm-c-helpers.cpp
new file mode 100644
index 000000000000..cde3f1f03af0
--- /dev/null
+++ b/tools/perf/util/llvm-c-helpers.cpp
@@ -0,0 +1,130 @@
+// SPDX-License-Identifier: GPL-2.0
+
+/*
+ * Must come before the linux/compiler.h include, which defines several
+ * macros (e.g. noinline) that conflict with compiler builtins used
+ * by LLVM.
+ */
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wunused-parameter" /* Needed for LLVM <= 15 */
+#include <llvm/DebugInfo/Symbolize/Symbolize.h>
+#pragma GCC diagnostic pop
+
+#include <stdio.h>
+#include <sys/types.h>
+#include <linux/compiler.h>
+#include "symbol_conf.h"
+#include "llvm-c-helpers.h"
+
+using namespace llvm;
+using llvm::symbolize::LLVMSymbolizer;
+
+/*
+ * Allocate a static LLVMSymbolizer, which will live to the end of the program.
+ * Unlike the bfd paths, LLVMSymbolizer has its own cache, so we do not need
+ * to store anything in the dso struct.
+ */
+static LLVMSymbolizer *get_symbolizer()
+{
+ static LLVMSymbolizer *instance = nullptr;
+ if (instance == nullptr) {
+ LLVMSymbolizer::Options opts;
+ /*
+ * LLVM sometimes demangles slightly different from the rest
+ * of the code, and this mismatch can cause new_inline_sym()
+ * to get confused and mark non-inline symbol as inlined
+ * (since the name does not properly match up with base_sym).
+ * Thus, disable the demangling and let the rest of the code
+ * handle it.
+ */
+ opts.Demangle = false;
+ instance = new LLVMSymbolizer(opts);
+ }
+ return instance;
+}
+
+/* Returns 0 on error, 1 on success. */
+static int extract_file_and_line(const DILineInfo& line_info, char **file,
+ unsigned int *line)
+{
+ if (file) {
+ if (line_info.FileName == "<invalid>") {
+ /* Match the convention of libbfd. */
+ *file = nullptr;
+ } else {
+ /* The caller expects to get something it can free(). */
+ *file = strdup(line_info.FileName.c_str());
+ if (*file == nullptr)
+ return 0;
+ }
+ }
+ if (line)
+ *line = line_info.Line;
+ return 1;
+}
+
+extern "C"
+int llvm_addr2line(const char *dso_name, u64 addr,
+ char **file, unsigned int *line,
+ bool unwind_inlines,
+ llvm_a2l_frame** inline_frames)
+{
+ LLVMSymbolizer *symbolizer = get_symbolizer();
+ object::SectionedAddress sectioned_addr = {
+ addr,
+ object::SectionedAddress::UndefSection
+ };
+
+ if (unwind_inlines) {
+ Expected<DIInliningInfo> res_or_err =
+ symbolizer->symbolizeInlinedCode(dso_name,
+ sectioned_addr);
+ if (!res_or_err)
+ return 0;
+ unsigned num_frames = res_or_err->getNumberOfFrames();
+ if (num_frames == 0)
+ return 0;
+
+ if (extract_file_and_line(
+ res_or_err->getFrame(0), file, line) == 0)
+ return 0;
+
+ *inline_frames = (llvm_a2l_frame*)malloc(
+ sizeof(**inline_frames) * num_frames);
+ if (*inline_frames == nullptr)
+ return 0;
+
+ for (unsigned i = 0; i < num_frames; ++i) {
+ const DILineInfo& src = res_or_err->getFrame(i);
+ llvm_a2l_frame& dst = (*inline_frames)[i];
+ if (src.FileName == "<invalid>")
+ /* Match the convention of libbfd. */
+ dst.filename = nullptr;
+ else
+ dst.filename = strdup(src.FileName.c_str());
+ dst.funcname = strdup(src.FunctionName.c_str());
+ dst.line = src.Line;
+
+ if (dst.filename == nullptr ||
+ dst.funcname == nullptr) {
+ for (unsigned j = 0; j <= i; ++j) {
+ free((*inline_frames)[j].filename);
+ free((*inline_frames)[j].funcname);
+ }
+ free(*inline_frames);
+ return 0;
+ }
+ }
+
+ return num_frames;
+ } else {
+ if (inline_frames)
+ *inline_frames = nullptr;
+
+ Expected<DILineInfo> res_or_err =
+ symbolizer->symbolizeCode(dso_name, sectioned_addr);
+ if (!res_or_err)
+ return 0;
+ return extract_file_and_line(*res_or_err, file, line);
+ }
+}
diff --git a/tools/perf/util/llvm-c-helpers.h b/tools/perf/util/llvm-c-helpers.h
new file mode 100644
index 000000000000..1b28cdc9f9b7
--- /dev/null
+++ b/tools/perf/util/llvm-c-helpers.h
@@ -0,0 +1,47 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __PERF_LLVM_C_HELPERS
+#define __PERF_LLVM_C_HELPERS 1
+
+/*
+ * Helpers to call into LLVM C++ code from C, for the parts that do not have
+ * C APIs.
+ */
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct llvm_a2l_frame {
+ char *filename;
+ char *funcname;
+ unsigned int line;
+};
+
+/*
+ * Implement addr2line() using libLLVM. LLVM is a C++ API, and
+ * many of the linux/ headers cannot be included in a C++ compile unit,
+ * so we need to make a little bridge code here. llvm_addr2line() will
+ * convert the inline frame information from LLVM's internal structures
+ * and put them into a flat array given in inline_frames. The caller
+ * is then responsible for taking that array and convert it into perf's
+ * regular inline frame structures (which depend on e.g. struct list_head).
+ *
+ * If the address could not be resolved, or an error occurred (e.g. OOM),
+ * returns 0. Otherwise, returns the number of inline frames (which means 1
+ * if the address was not part of an inlined function). If unwind_inlines
+ * is set and the return code is nonzero, inline_frames will be set to
+ * a newly allocated array with that length. The caller is then responsible
+ * for freeing both the strings and the array itself.
+ */
+int llvm_addr2line(const char *dso_name,
+ u64 addr,
+ char **file,
+ unsigned int *line,
+ bool unwind_inlines,
+ struct llvm_a2l_frame **inline_frames);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* __PERF_LLVM_C_HELPERS */
diff --git a/tools/perf/util/srcline.c b/tools/perf/util/srcline.c
index 9d670d8c1c08..0505b4c16608 100644
--- a/tools/perf/util/srcline.c
+++ b/tools/perf/util/srcline.c
@@ -16,6 +16,9 @@
#include "util/debug.h"
#include "util/callchain.h"
#include "util/symbol_conf.h"
+#ifdef HAVE_LLVM_SUPPORT
+#include "util/llvm-c-helpers.h"
+#endif
#include "srcline.h"
#include "string2.h"
#include "symbol.h"
@@ -130,7 +133,59 @@ static struct symbol *new_inline_sym(struct dso *dso,
#define MAX_INLINE_NEST 1024
-#ifdef HAVE_LIBBFD_SUPPORT
+#ifdef HAVE_LLVM_SUPPORT
+
+static void free_llvm_inline_frames(struct llvm_a2l_frame *inline_frames,
+ int num_frames)
+{
+ if (inline_frames != NULL) {
+ for (int i = 0; i < num_frames; ++i) {
+ free(inline_frames[i].filename);
+ free(inline_frames[i].funcname);
+ }
+ free(inline_frames);
+ }
+}
+
+static int addr2line(const char *dso_name, u64 addr,
+ char **file, unsigned int *line, struct dso *dso,
+ bool unwind_inlines, struct inline_node *node,
+ struct symbol *sym)
+{
+ struct llvm_a2l_frame *inline_frames = NULL;
+ int num_frames = llvm_addr2line(dso_name, addr, file, line,
+ node && unwind_inlines, &inline_frames);
+
+ if (num_frames == 0 || !inline_frames) {
+ /* Error, or we didn't want inlines. */
+ return num_frames;
+ }
+
+ for (int i = 0; i < num_frames; ++i) {
+ struct symbol *inline_sym =
+ new_inline_sym(dso, sym, inline_frames[i].funcname);
+ char *srcline = NULL;
+
+ if (inline_frames[i].filename)
+ srcline = srcline_from_fileline(
+ inline_frames[i].filename,
+ inline_frames[i].line);
+ if (inline_list__append(inline_sym, srcline, node) != 0) {
+ free_llvm_inline_frames(inline_frames, num_frames);
+ return 0;
+ }
+ }
+ free_llvm_inline_frames(inline_frames, num_frames);
+
+ return num_frames;
+}
+
+void dso__free_a2l(struct dso *)
+{
+ /* Nothing to free. */
+}
+
+#elif defined(HAVE_LIBBFD_SUPPORT)
/*
* Implement addr2line using libbfd.
--
2.45.1
Support using LLVM as a disassembler method, allowing helperless
annotation in non-distro builds. (It is also much faster than
using libbfd or bfd objdump on binaries with a lot of debug
information.)
This is nearly identical to the output of llvm-objdump; there are
some very rare whitespace differences, some minor changes to demangling
(since we use perf's regular demangling and not LLVM's own) and
the occasional case where llvm-objdump makes a different choice
when multiple symbols share the same address. It should work across
all of LLVM's supported architectures, although I've only tested 64-bit
x86, and finding the right triple from perf's idea of machine
architecture can sometimes be a bit tricky. Ideally, we should have
some way of finding the triplet just from the file itself.
Signed-off-by: Steinar H. Gunderson <[email protected]>
---
tools/perf/util/disasm.c | 195 +++++++++++++++++++++++++++++
tools/perf/util/llvm-c-helpers.cpp | 62 +++++++++
tools/perf/util/llvm-c-helpers.h | 11 ++
3 files changed, 268 insertions(+)
diff --git a/tools/perf/util/disasm.c b/tools/perf/util/disasm.c
index c0dbb955e61a..9c07d2a8c8a8 100644
--- a/tools/perf/util/disasm.c
+++ b/tools/perf/util/disasm.c
@@ -43,6 +43,7 @@ static int call__scnprintf(struct ins *ins, char *bf, size_t size,
static void ins__sort(struct arch *arch);
static int disasm_line__parse(char *line, const char **namep, char **rawp);
+static char *expand_tabs(char *line, char **storage, size_t *storage_len);
static __attribute__((constructor)) void symbol__init_regexpr(void)
{
@@ -1378,7 +1379,9 @@ static int open_capstone_handle(struct annotate_args *args, bool is_64bit,
return 0;
}
+#endif
+#if defined(HAVE_LIBCAPSTONE_SUPPORT) || defined(HAVE_LLVM_SUPPORT)
struct find_file_offset_data {
u64 ip;
u64 offset;
@@ -1442,7 +1445,9 @@ read_symbol(const char *filename, struct map *map, struct symbol *sym,
free(buf);
return NULL;
}
+#endif
+#ifdef HAVE_LIBCAPSTONE_SUPPORT
static void print_capstone_detail(cs_insn *insn, char *buf, size_t len,
struct annotate_args *args, u64 addr)
{
@@ -1606,6 +1611,191 @@ static int symbol__disassemble_capstone(char *filename, struct symbol *sym,
}
#endif
+#ifdef HAVE_LLVM_SUPPORT
+#include <llvm-c/Disassembler.h>
+#include <llvm-c/Target.h>
+#include "util/llvm-c-helpers.h"
+
+struct symbol_lookup_storage {
+ u64 branch_addr;
+ u64 pcrel_load_addr;
+};
+
+/*
+ * Whenever LLVM wants to resolve an address into a symbol, it calls this
+ * callback. We don't ever actually _return_ anything (in particular, because
+ * it puts quotation marks around what we return), but we use this as a hint
+ * that there is a branch or PC-relative address in the expression that we
+ * should add some textual annotation for after the instruction. The caller
+ * will use this information to add the actual annotation.
+ */
+static const char *
+symbol_lookup_callback(void *disinfo, uint64_t value,
+ uint64_t *ref_type,
+ uint64_t address __maybe_unused,
+ const char **ref __maybe_unused)
+{
+ struct symbol_lookup_storage *storage =
+ (struct symbol_lookup_storage *)disinfo;
+ if (*ref_type == LLVMDisassembler_ReferenceType_In_Branch)
+ storage->branch_addr = value;
+ else if (*ref_type == LLVMDisassembler_ReferenceType_In_PCrel_Load)
+ storage->pcrel_load_addr = value;
+ *ref_type = LLVMDisassembler_ReferenceType_InOut_None;
+ return NULL;
+}
+
+static int symbol__disassemble_llvm(char *filename, struct symbol *sym,
+ struct annotate_args *args)
+{
+ struct annotation *notes = symbol__annotation(sym);
+ struct map *map = args->ms.map;
+ struct dso *dso = map__dso(map);
+ u64 start = map__rip_2objdump(map, sym->start);
+ u8 *buf;
+ u64 len;
+ u64 pc;
+ bool is_64bit;
+ char triplet[64];
+ char disasm_buf[2048];
+ size_t disasm_len;
+ struct disasm_line *dl;
+ LLVMDisasmContextRef disasm = NULL;
+ struct symbol_lookup_storage storage;
+ char *line_storage = NULL;
+ size_t line_storage_len = 0;
+
+ if (args->options->objdump_path)
+ return -1;
+
+ LLVMInitializeAllTargetInfos();
+ LLVMInitializeAllTargetMCs();
+ LLVMInitializeAllDisassemblers();
+
+ buf = read_symbol(filename, map, sym, &len, &is_64bit);
+ if (buf == NULL)
+ return -1;
+
+ if (arch__is(args->arch, "x86")) {
+ if (is_64bit)
+ scnprintf(triplet, sizeof(triplet), "x86_64-pc-linux");
+ else
+ scnprintf(triplet, sizeof(triplet), "i686-pc-linux");
+ } else {
+ scnprintf(triplet, sizeof(triplet), "%s-linux-gnu",
+ args->arch->name);
+ }
+
+ disasm = LLVMCreateDisasm(
+ triplet, &storage, 0, NULL, symbol_lookup_callback);
+ if (disasm == NULL)
+ goto err;
+
+ if (args->options->disassembler_style &&
+ !strcmp(args->options->disassembler_style, "intel"))
+ LLVMSetDisasmOptions(
+ disasm, LLVMDisassembler_Option_AsmPrinterVariant);
+
+ /*
+ * This needs to be set after AsmPrinterVariant, due to a bug in LLVM;
+ * setting AsmPrinterVariant makes a new instruction printer, making it
+ * forget about the PrintImmHex flag (which is applied before if both
+ * are given to the same call).
+ */
+ LLVMSetDisasmOptions(disasm, LLVMDisassembler_Option_PrintImmHex);
+
+ /* add the function address and name */
+ scnprintf(disasm_buf, sizeof(disasm_buf), "%#"PRIx64" <%s>:",
+ start, sym->name);
+
+ args->offset = -1;
+ args->line = disasm_buf;
+ args->line_nr = 0;
+ args->fileloc = NULL;
+ args->ms.sym = sym;
+
+ dl = disasm_line__new(args);
+ if (dl == NULL)
+ goto err;
+
+ annotation_line__add(&dl->al, ¬es->src->source);
+
+ pc = start;
+ for (u64 offset = 0; offset < len; ) {
+ unsigned int ins_len;
+
+ storage.branch_addr = 0;
+ storage.pcrel_load_addr = 0;
+
+ ins_len = LLVMDisasmInstruction(
+ disasm, buf + offset, len - offset, pc,
+ disasm_buf, sizeof(disasm_buf));
+ if (ins_len == 0)
+ goto err;
+ disasm_len = strlen(disasm_buf);
+
+ if (storage.branch_addr != 0) {
+ char *name = llvm_name_for_code(
+ dso, filename, storage.branch_addr);
+ if (name != NULL) {
+ disasm_len += scnprintf(
+ disasm_buf + disasm_len,
+ sizeof(disasm_buf) - disasm_len,
+ " <%s>", name);
+ free(name);
+ }
+ }
+ if (storage.pcrel_load_addr != 0) {
+ char *name = llvm_name_for_data(
+ dso, filename, storage.pcrel_load_addr);
+ disasm_len += scnprintf(disasm_buf + disasm_len,
+ sizeof(disasm_buf) - disasm_len,
+ " # %#"PRIx64,
+ storage.pcrel_load_addr);
+ if (name) {
+ disasm_len += scnprintf(
+ disasm_buf + disasm_len,
+ sizeof(disasm_buf) - disasm_len,
+ " <%s>", name);
+ free(name);
+ }
+ }
+
+ args->offset = offset;
+ args->line = expand_tabs(
+ disasm_buf, &line_storage, &line_storage_len);
+ args->line_nr = 0;
+ args->fileloc = NULL;
+ args->ms.sym = sym;
+
+ llvm_addr2line(filename, pc, &args->fileloc,
+ (unsigned int *)&args->line_nr, false, NULL);
+
+ dl = disasm_line__new(args);
+ if (dl == NULL)
+ goto err;
+
+ annotation_line__add(&dl->al, ¬es->src->source);
+
+ free(args->fileloc);
+ pc += ins_len;
+ offset += ins_len;
+ }
+
+ LLVMDisasmDispose(disasm);
+ free(buf);
+ free(line_storage);
+ return 0;
+
+err:
+ LLVMDisasmDispose(disasm);
+ free(buf);
+ free(line_storage);
+ return -1;
+}
+#endif
+
+
/*
* Possibly create a new version of line with tabs expanded. Returns the
* existing or new line, storage is updated if a new line is allocated. If
@@ -1730,6 +1920,11 @@ int symbol__disassemble(struct symbol *sym, struct annotate_args *args)
strcpy(symfs_filename, tmp);
}
+#ifdef HAVE_LLVM_SUPPORT
+ err = symbol__disassemble_llvm(symfs_filename, sym, args);
+ if (err == 0)
+ goto out_remove_tmp;
+#endif
#ifdef HAVE_LIBCAPSTONE_SUPPORT
err = symbol__disassemble_capstone(symfs_filename, sym, args);
if (err == 0)
diff --git a/tools/perf/util/llvm-c-helpers.cpp b/tools/perf/util/llvm-c-helpers.cpp
index cde3f1f03af0..ef10cc191344 100644
--- a/tools/perf/util/llvm-c-helpers.cpp
+++ b/tools/perf/util/llvm-c-helpers.cpp
@@ -8,6 +8,7 @@
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wunused-parameter" /* Needed for LLVM <= 15 */
#include <llvm/DebugInfo/Symbolize/Symbolize.h>
+#include <llvm/Support/TargetSelect.h>
#pragma GCC diagnostic pop
#include <stdio.h>
@@ -16,6 +17,9 @@
#include "symbol_conf.h"
#include "llvm-c-helpers.h"
+extern "C"
+char *dso__demangle_sym(struct dso *dso, int kmodule, const char *elf_name);
+
using namespace llvm;
using llvm::symbolize::LLVMSymbolizer;
@@ -128,3 +132,61 @@ int llvm_addr2line(const char *dso_name, u64 addr,
return extract_file_and_line(*res_or_err, file, line);
}
}
+
+static char *
+make_symbol_relative_string(struct dso *dso, const char *sym_name,
+ u64 addr, u64 base_addr)
+{
+ if (!strcmp(sym_name, "<invalid>"))
+ return NULL;
+
+ char *demangled = dso__demangle_sym(dso, 0, sym_name);
+ if (base_addr && base_addr != addr) {
+ char buf[256];
+ snprintf(buf, sizeof(buf), "%s+0x%lx",
+ demangled ? demangled : sym_name, addr - base_addr);
+ free(demangled);
+ return strdup(buf);
+ } else {
+ if (demangled)
+ return demangled;
+ else
+ return strdup(sym_name);
+ }
+}
+
+extern "C"
+char *llvm_name_for_code(struct dso *dso, const char *dso_name, u64 addr)
+{
+ LLVMSymbolizer *symbolizer = get_symbolizer();
+ object::SectionedAddress sectioned_addr = {
+ addr,
+ object::SectionedAddress::UndefSection
+ };
+ Expected<DILineInfo> res_or_err =
+ symbolizer->symbolizeCode(dso_name, sectioned_addr);
+ if (!res_or_err) {
+ return NULL;
+ }
+ return make_symbol_relative_string(
+ dso, res_or_err->FunctionName.c_str(),
+ addr, res_or_err->StartAddress ? *res_or_err->StartAddress : 0);
+}
+
+extern "C"
+char *llvm_name_for_data(struct dso *dso, const char *dso_name, u64 addr)
+{
+ LLVMSymbolizer *symbolizer = get_symbolizer();
+ object::SectionedAddress sectioned_addr = {
+ addr,
+ object::SectionedAddress::UndefSection
+ };
+ Expected<DIGlobal> res_or_err =
+ symbolizer->symbolizeData(dso_name, sectioned_addr);
+ if (!res_or_err) {
+ return NULL;
+ }
+ return make_symbol_relative_string(
+ dso, res_or_err->Name.c_str(),
+ addr, res_or_err->Start);
+}
diff --git a/tools/perf/util/llvm-c-helpers.h b/tools/perf/util/llvm-c-helpers.h
index 1b28cdc9f9b7..e00e3f73a906 100644
--- a/tools/perf/util/llvm-c-helpers.h
+++ b/tools/perf/util/llvm-c-helpers.h
@@ -11,6 +11,8 @@
extern "C" {
#endif
+struct dso;
+
struct llvm_a2l_frame {
char *filename;
char *funcname;
@@ -40,6 +42,15 @@ int llvm_addr2line(const char *dso_name,
bool unwind_inlines,
struct llvm_a2l_frame **inline_frames);
+/*
+ * Simple symbolizers for addresses; will convert something like
+ * 0x12345 to "func+0x123". Will return NULL if no symbol was found.
+ *
+ * The returned value must be freed by the caller, with free().
+ */
+char *llvm_name_for_code(struct dso *dso, const char *dso_name, u64 addr);
+char *llvm_name_for_data(struct dso *dso, const char *dso_name, u64 addr);
+
#ifdef __cplusplus
}
#endif
--
2.45.1
The Capstone disassembler code has a useful code snippet to read
the bytes for a given code symbol into memory. Split it out into
its own function, so that the LLVM disassembler can use it in
the next patch.
Signed-off-by: Steinar H. Gunderson <[email protected]>
---
tools/perf/util/disasm.c | 90 +++++++++++++++++++++++++---------------
1 file changed, 56 insertions(+), 34 deletions(-)
diff --git a/tools/perf/util/disasm.c b/tools/perf/util/disasm.c
index 72aec8f61b94..c0dbb955e61a 100644
--- a/tools/perf/util/disasm.c
+++ b/tools/perf/util/disasm.c
@@ -1396,6 +1396,53 @@ static int find_file_offset(u64 start, u64 len, u64 pgoff, void *arg)
return 0;
}
+static u8 *
+read_symbol(const char *filename, struct map *map, struct symbol *sym,
+ u64 *len, bool *is_64bit)
+{
+ struct dso *dso = map__dso(map);
+ struct nscookie nsc;
+ u64 start = map__rip_2objdump(map, sym->start);
+ u64 end = map__rip_2objdump(map, sym->end);
+ int fd, count;
+ u8 *buf = NULL;
+ struct find_file_offset_data data = {
+ .ip = start,
+ };
+
+ *is_64bit = false;
+
+ nsinfo__mountns_enter(dso__nsinfo(dso), &nsc);
+ fd = open(filename, O_RDONLY);
+ nsinfo__mountns_exit(&nsc);
+ if (fd < 0)
+ return NULL;
+
+ if (file__read_maps(fd, /*exe=*/true, find_file_offset, &data,
+ is_64bit) == 0)
+ goto err;
+
+ *len = end - start;
+ buf = malloc(*len);
+ if (buf == NULL)
+ goto err;
+
+ count = pread(fd, buf, *len, data.offset);
+ close(fd);
+ fd = -1;
+
+ if ((u64)count != *len)
+ goto err;
+
+ return buf;
+
+err:
+ if (fd >= 0)
+ close(fd);
+ free(buf);
+ return NULL;
+}
+
static void print_capstone_detail(cs_insn *insn, char *buf, size_t len,
struct annotate_args *args, u64 addr)
{
@@ -1458,19 +1505,13 @@ static int symbol__disassemble_capstone(char *filename, struct symbol *sym,
{
struct annotation *notes = symbol__annotation(sym);
struct map *map = args->ms.map;
- struct dso *dso = map__dso(map);
- struct nscookie nsc;
u64 start = map__rip_2objdump(map, sym->start);
- u64 end = map__rip_2objdump(map, sym->end);
- u64 len = end - start;
+ u64 len;
u64 offset;
- int i, fd, count;
+ int i, count;
bool is_64bit = false;
bool needs_cs_close = false;
u8 *buf = NULL;
- struct find_file_offset_data data = {
- .ip = start,
- };
csh handle;
cs_insn *insn;
char disasm_buf[512];
@@ -1479,31 +1520,9 @@ static int symbol__disassemble_capstone(char *filename, struct symbol *sym,
if (args->options->objdump_path)
return -1;
- nsinfo__mountns_enter(dso__nsinfo(dso), &nsc);
- fd = open(filename, O_RDONLY);
- nsinfo__mountns_exit(&nsc);
- if (fd < 0)
- return -1;
-
- if (file__read_maps(fd, /*exe=*/true, find_file_offset, &data,
- &is_64bit) == 0)
- goto err;
-
- if (open_capstone_handle(args, is_64bit, &handle) < 0)
- goto err;
-
- needs_cs_close = true;
-
- buf = malloc(len);
+ buf = read_symbol(filename, map, sym, &len, &is_64bit);
if (buf == NULL)
- goto err;
-
- count = pread(fd, buf, len, data.offset);
- close(fd);
- fd = -1;
-
- if ((u64)count != len)
- goto err;
+ return -1;
/* add the function address and name */
scnprintf(disasm_buf, sizeof(disasm_buf), "%#"PRIx64" <%s>:",
@@ -1521,6 +1540,11 @@ static int symbol__disassemble_capstone(char *filename, struct symbol *sym,
annotation_line__add(&dl->al, ¬es->src->source);
+ if (open_capstone_handle(args, is_64bit, &handle) < 0)
+ goto err;
+
+ needs_cs_close = true;
+
count = cs_disasm(handle, buf, len, start, len, &insn);
for (i = 0, offset = 0; i < count; i++) {
int printed;
@@ -1565,8 +1589,6 @@ static int symbol__disassemble_capstone(char *filename, struct symbol *sym,
return count < 0 ? count : 0;
err:
- if (fd >= 0)
- close(fd);
if (needs_cs_close) {
struct disasm_line *tmp;
--
2.45.1
On Thu, May 23, 2024 at 11:29:18AM +0200, Steinar H. Gunderson wrote:
> In addition to the existing support for libbfd and calling out to
> an external addr2line command, add support for using libllvm directly.
> This is both faster than libbfd, and can be enabled in distro builds
> (the LLVM license has an explicit provision for GPLv2 compatibility).
> Thus, it is set as the primary choice if available.
>
> As an example, running perf report on a medium-size profile with
> DWARF-based backtraces took 58 seconds with LLVM, 78 seconds with
> libbfd, 153 seconds with external llvm-addr2line, and I got tired
> and aborted the test after waiting for 55 minutes with external
> bfd addr2line (which is the default for perf as compiled by distributions
Before this patch, using BFD's addr2line:
root@x1:~# echo 1 > /proc/sys/vm/drop_caches
root@x1:~# perf record --call-graph dwarf find / > /dev/null
[ perf record: Woken up 428 times to write data ]
[ perf record: Captured and wrote 107.598 MB perf.data (13355 samples) ]
root@x1:~#
root@x1:~# dnf debuginfo-install findutils
root@x1:~# perf report --no-children --stdio -s srcfile,srcline
# To display the perf.data header info, please use --header/--header-only options.
#
# Total Lost Samples: 0
#
# Samples: 75 of event 'cpu_atom/cycles/P'
# Event count (approx.): 22073647
#
# Overhead Source File Source:Line
# ........ ................. ................................
#
4.59% disk-io.c disk-io.c:108
|
---btrfs_buffer_uptodate
read_block_for_search
|
--4.13%--btrfs_search_slot
btrfs_real_readdir
iterate_dir
__x64_sys_getdents64
do_syscall_64
entry_SYSCALL_64
<SNIP>
root@x1:~# echo 1 > /proc/sys/vm/drop_caches
root@x1:~# taskset -c 0 perf stat -e cpu_core/cycles/P perf report --no-children --stdio -s srcfile,srcline > /dev/null
Performance counter stats for 'perf report --no-children --stdio -s srcfile,srcline':
225,837,039,942 cpu_core/cycles/P
113.192464730 seconds time elapsed
102.789965000 seconds user
6.478551000 seconds sys
root@x1:~#
Now with your patches:
Brownie points++:
Makefile.config:982: No libllvm found, slower source file resolution, please install llvm-devel/llvm-dev
... llvm: [ OFF ]
But maybe use "libllvm"?
⬢[acme@toolbox perf-tools]$ sudo dnf install llvm-devel
... libcapstone: [ on ]
... llvm: [ on ]
root@x1:~# echo 1 > /proc/sys/vm/drop_caches
root@x1:~# taskset -c 0 perf stat -e cpu_core/cycles/P perf report --no-children --stdio -s srcfile,srcline > /dev/null
Performance counter stats for 'perf report --no-children --stdio -s srcfile,srcline':
2,462,040,637 cpu_core/cycles/P
1.696059762 seconds time elapsed
1.013055000 seconds user
0.423232000 seconds sys
root@x1:~#
root@x1:~# head -25 report.srcfile.srcline.libllvm
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 75 of event 'cpu_atom/cycles/P'
# Event count (approx.): 22073647
#
# Overhead Source File Source:Line
# ........ ................. ................................
#
4.59% disk-io.c disk-io.c:108
|
---btrfs_buffer_uptodate
read_block_for_search
|
--4.13%--btrfs_search_slot
btrfs_real_readdir
iterate_dir
__x64_sys_getdents64
do_syscall_64
entry_SYSCALL_64
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
root@x1:~#
Matches, now everything:
root@x1:~# diff -u report.srcfile.srcline.bfd report.srcfile.srcline.libllvm
--- report.srcfile.srcline.bfd 2024-05-23 14:31:26.717881878 -0300
+++ report.srcfile.srcline.libllvm 2024-05-23 14:22:37.795843724 -0300
@@ -23,6 +23,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
So mostly the above, and:
root@x1:~# perf probe -x /usr/bin/find -L find | grep fts_read
44 while ( (errno=0, ent=fts_read (p)) != NULL )
/* fts_read returned NULL; distinguish between "finished" and "error". */
root@x1:~#
root@x1:~# readelf -wi /usr/lib/debug/usr/bin/find-4.9.0-5.fc39.x86_64.debug | grep "DW_AT_name.*rpl_fts_read" -B2
<1><2c69c>: Abbrev Number: 11 (DW_TAG_subprogram)
<2c69d> DW_AT_external : 1
<2c69d> DW_AT_name : (indirect string, offset: 0xda3): rpl_fts_read
--
<1><364b8>: Abbrev Number: 25 (DW_TAG_subprogram)
^^^^^^^
^^^^^^^
^^^^^^^
<364b9> DW_AT_external : 1
<364b9> DW_AT_name : (indirect string, offset: 0xda3): rpl_fts_read
root@x1:~#
<1><2c952>: Abbrev Number: 81 (DW_TAG_subprogram)
^^^^^^
^^^^^^
^^^^^^
<2c953> DW_AT_name : (indirect string, offset: 0x118f): find
<2c957> DW_AT_decl_file : 22
<2c957> DW_AT_decl_line : 482
<2c959> DW_AT_decl_column : 1
<2c959> DW_AT_prototyped : 1
<2c959> DW_AT_type : <alt 0x573b>
<2c95d> DW_AT_sibling : <0x2c996>
Then the function referring to the above that has inlined functions:
<1><def8>: Abbrev Number: 2 (DW_TAG_subprogram)
<def9> DW_AT_abstract_origin: <0x2c952> <=============================
<defd> DW_AT_ranges : 0xff0
<df01> DW_AT_frame_base : 1 byte block: 9c (DW_OP_call_frame_cfa)
<df03> DW_AT_call_all_calls: 1
<df03> DW_AT_sibling : <0xe84a>
<2><df06>: Abbrev Number: 62 (DW_TAG_formal_parameter)
<df07> DW_AT_abstract_origin: <0x2c95f>
<df0b> DW_AT_location : 0x8345 (location list)
<df0f> DW_AT_GNU_locviews: 0x833f
<2><df13>: Abbrev Number: 52 (DW_TAG_variable)
<df14> DW_AT_abstract_origin: <0x2c96b>
<df18> DW_AT_location : 3 byte block: 91 a0 7f (DW_OP_fbreg: -96)
<SNIP>
<3><df77>: Abbrev Number: 51 (DW_TAG_inlined_subroutine)
<df78> DW_AT_abstract_origin: <0x364b8> <=========================
<df7c> DW_AT_entry_pc : 0xeddf
<df84> DW_AT_GNU_entry_view: 0
<df86> DW_AT_ranges : 0x1042
<df8a> DW_AT_call_file : 6
<df8b> DW_AT_call_line : 526
<df8d> DW_AT_call_column : 29
<df8e> DW_AT_sibling : <0xe60d>
<4><df91>: Abbrev Number: 62 (DW_TAG_formal_parameter)
<df92> DW_AT_abstract_origin: <0x364c3>
<df96> DW_AT_location : 0x848f (location list)
<df9a> DW_AT_GNU_locviews: 0x8473
<4><df9e>: Abbrev Number: 14 (DW_TAG_lexical_block)
<df9f> DW_AT_ranges : 0x1042
<SNIP>
So the libllvm is even producing a better result, showing those inlines
not seen in the BFD based output.
So going from (BFD):
225,837,039,942 cpu_core/cycles/P
113.192464730 seconds time elapsed
102.789965000 seconds user
6.478551000 seconds sys
To (using libllvm):
2,462,040,637 cpu_core/cycles/P
1.696059762 seconds time elapsed
1.013055000 seconds user
0.423232000 seconds sys
100x speedup, looks like a win! 8-)
Thanks a lot, the comments I made on the patch are mostly coding style,
please consider them, but I wouldn't get in the way for this patch to
get merged because of that albeit would be nice to try to fit in more
nicely with the existing source code base.
Assuming the code remains the same after an eventual coding style
polish:
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
- Arnaldo
The rest, more comments in the patch, after the diff output.
@@ -35,6 +36,7 @@
---__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -56,6 +58,7 @@
| __GI___getdents64
| __GI___readdir64
| fts_build.constprop.0
+ | rpl_fts_read (inlined)
| find
| process_all_startpoints
| main
@@ -75,6 +78,8 @@
0x55f742a0e9a3
0x55f742a0e9a3
0x55f742a02735
+ 0x55f742a02735
+ 0x55f742a02735
0x55f742a02f6d
0x55f742a037c4
0x55f7429fc0f7
@@ -90,6 +95,7 @@
| __fdopendir
| opendirat
| fts_build.constprop.0
+ | rpl_fts_read (inlined)
| find
| process_all_startpoints
| main
@@ -98,6 +104,8 @@
| _start
|
--2.09%--__close
+ left_dir (inlined)
+ left_dir (inlined)
consider_visiting
find
process_all_startpoints
@@ -117,6 +125,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -152,6 +161,7 @@
openat_safer
opendirat
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -170,11 +180,13 @@
do_syscall_64
entry_SYSCALL_64
__mmap64 (inlined)
- __mmap64 (inlined)
+ mmap64
+ _dl_map_segment (inlined)
+ _dl_map_segments (inlined)
_dl_map_object_from_fd
_dl_map_object
openaux
- __GI__dl_catch_exception (inlined)
+ _dl_catch_exception
_dl_map_object_deps
dl_main
_dl_sysdep_start
@@ -202,7 +214,8 @@
2.21% do-rel.h do-rel.h:133
|
- ---_dl_relocate_object
+ ---elf_dynamic_do_Rela (inlined)
+ _dl_relocate_object
dl_main
_dl_sysdep_start
_dl_start_final (inlined)
@@ -216,6 +229,7 @@
entry_SYSCALL_64
__GI___fstatat64
fts_stat.isra.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -231,6 +245,8 @@
do_syscall_64
entry_SYSCALL_64
__close
+ left_dir (inlined)
+ left_dir (inlined)
consider_visiting
find
process_all_startpoints
@@ -250,6 +266,7 @@
entry_SYSCALL_64
__GI___fstatat64
fts_stat.isra.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -272,6 +289,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -303,6 +321,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -318,6 +337,7 @@
pred_print
apply_predicate
apply_predicate
+ visit (inlined)
consider_visiting
find
process_all_startpoints
@@ -331,6 +351,7 @@
---_int_free_create_chunk
_int_free_merge_chunk
cfree@GLIBC_2.2.5
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -340,7 +361,7 @@
2.09% fts-cycle.c fts-cycle.c:126
|
- ---leave_dir (inlined)
+ ---leave_dir.lto_priv.0
0x7ffd0a6562af
2.09% __printf_buffer_to_file_switch+0
@@ -351,6 +372,7 @@
pred_print
apply_predicate
apply_predicate
+ visit (inlined)
consider_visiting
find
process_all_startpoints
@@ -371,6 +393,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -387,6 +410,8 @@
do_syscall_64
entry_SYSCALL_64
__close
+ left_dir (inlined)
+ left_dir (inlined)
consider_visiting
find
process_all_startpoints
@@ -403,6 +428,7 @@
__GI___fstatat64
fts_stat.isra.0
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -421,6 +447,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -432,7 +459,8 @@
|
---_int_free
cfree@GLIBC_2.2.5
- leave_dir (inlined)
+ leave_dir.lto_priv.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -456,6 +484,7 @@
openat_safer
opendirat
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -474,6 +503,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -492,6 +522,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -505,6 +536,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -516,6 +548,7 @@
|
---pred_and
apply_predicate
+ visit (inlined)
consider_visiting
find
process_all_startpoints
@@ -536,6 +569,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -553,6 +587,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -572,6 +607,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -591,6 +627,7 @@
openat_safer
opendirat
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -610,6 +647,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -620,6 +658,7 @@
1.98% cfree@GLIBC_2.2.5+139
|
---cfree@GLIBC_2.2.5
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -635,6 +674,8 @@
do_syscall_64
entry_SYSCALL_64
__close
+ left_dir (inlined)
+ left_dir (inlined)
consider_visiting
find
process_all_startpoints
@@ -655,6 +696,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -677,6 +719,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -695,6 +738,7 @@
__GI___fstatat64
fts_stat.isra.0
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -706,8 +750,9 @@
|
---hash_find_entry
hash_remove (inlined)
- leave_dir (inlined)
+ leave_dir.lto_priv.0
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -740,6 +785,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -753,8 +799,9 @@
safe_hasher (inlined)
hash_find_entry
hash_remove (inlined)
- leave_dir (inlined)
+ leave_dir.lto_priv.0
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -799,6 +846,7 @@
| 0x7fc0e8e9ce16
| 0x55f742a1769f
| 0x55f742a03319
+ | 0x55f742a03319
| 0x55f742a037c4
| 0x55f7429fc0f7
| 0x7fc0e8deb149
@@ -815,6 +863,7 @@
0x7fc0e8e9ce16
0x55f742a1769f
0x55f742a03319
+ 0x55f742a03319
0x55f742a037c4
0x55f7429fc0f7
0x7fc0e8deb149
@@ -837,6 +886,7 @@
| | 0x7fc0e8e9ce16
| | 0x55f742a1769f
| | 0x55f742a03319
+ | | 0x55f742a03319
| | 0x55f742a037c4
| | 0x55f7429fc0f7
| | 0x7fc0e8deb149
@@ -852,6 +902,7 @@
| 0x7fc0e8e9ce16
| 0x55f742a1769f
| 0x55f742a03319
+ | 0x55f742a03319
| 0x55f742a037c4
| 0x55f7429fc0f7
| 0x7fc0e8deb149
@@ -870,6 +921,7 @@
entry_SYSCALL_64
__GI___fstatat64
fts_stat.isra.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -898,6 +950,7 @@
| __GI___getdents64
| __GI___readdir64
| fts_build.constprop.0
+ | rpl_fts_read (inlined)
| find
| process_all_startpoints
| main
@@ -913,6 +966,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -935,6 +989,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -947,6 +1002,7 @@
|
---__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -971,6 +1027,7 @@
0x55f742a17439
|
--0.90%--0x55f742a03545
+ 0x55f742a03545
0x55f742a037c4
0x55f7429fc0f7
0x7fc0e8deb149
@@ -988,6 +1045,7 @@
__GI___getdents64
__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -1019,6 +1077,7 @@
0x7fc0e8e9ce16
0x55f742a1769f
0x55f742a03319
+ 0x55f742a03319
0x55f742a037c4
0x55f7429fc0f7
0x7fc0e8deb149
@@ -1042,6 +1101,7 @@
0x7fc0e8e9ce16
0x55f742a1769f
0x55f742a03319
+ 0x55f742a03319
0x55f742a037c4
0x55f7429fc0f7
0x7fc0e8deb149
@@ -1052,6 +1112,7 @@
|
---__GI___readdir64
fts_build.constprop.0
+ rpl_fts_read (inlined)
find
process_all_startpoints
main
@@ -3292,5 +3353,5 @@
#
-# (Tip: For hierarchical output, try: perf report --hierarchy)
+# (Tip: For memory address profiling, try: perf mem record / perf mem report)
#
> today). Evidently, for this case, the bfd addr2line process needs
> 18 seconds (on a 5.2 GHz Zen 3) to load the .debug ELF in question,
> hits the 1-second timeout and gets killed during initialization,
> getting restarted anew every time. Having an in-process addr2line
> makes this much more robust.
>
> As future extensions, libllvm can be used in many other places where
> we currently use libbfd or other libraries:
>
> - Symbol enumeration (in particular, for PE binaries).
> - Demangling (including non-Itanium demangling, e.g. Microsoft
> or Rust).
> - Disassembling (perf annotate).
>
> However, these are much less pressing; most people don't profile
> PE binaries, and perf has non-bfd paths for ELF. The same with
> demangling; the default _cxa_demangle path works fine for most
> users. Disassembling is coming in a later patch in the series;
> however do note that while bfd objdump can be slow on large binaries,
> it is possible to use --objdump=llvm-objdump to get the speed benefits.
> (It appears LLVM-based demangling is very simple, should we want
> that.)
>
> Tested with LLVM 14, 15, 16, 18 and 19. For some reason, LLVM 12 was not
> correctly detected using feature_check, and thus was not tested.
>
> Signed-off-by: Steinar H. Gunderson <[email protected]>
> ---
> tools/build/Makefile.feature | 1 +
> tools/perf/Makefile.config | 15 ++++
> tools/perf/builtin-version.c | 1 +
> tools/perf/util/Build | 1 +
> tools/perf/util/llvm-c-helpers.cpp | 130 +++++++++++++++++++++++++++++
> tools/perf/util/llvm-c-helpers.h | 47 +++++++++++
> tools/perf/util/srcline.c | 57 ++++++++++++-
> 7 files changed, 251 insertions(+), 1 deletion(-)
> create mode 100644 tools/perf/util/llvm-c-helpers.cpp
> create mode 100644 tools/perf/util/llvm-c-helpers.h
>
> diff --git a/tools/build/Makefile.feature b/tools/build/Makefile.feature
> index 1e2ab148d5db..278b26216254 100644
> --- a/tools/build/Makefile.feature
> +++ b/tools/build/Makefile.feature
> @@ -136,6 +136,7 @@ FEATURE_DISPLAY ?= \
> libunwind \
> libdw-dwarf-unwind \
> libcapstone \
> + llvm \
libllvm please
> zlib \
> lzma \
> get_cpuid \
> diff --git a/tools/perf/Makefile.config b/tools/perf/Makefile.config
> index 7f1e016a9253..414a37f712bd 100644
> --- a/tools/perf/Makefile.config
> +++ b/tools/perf/Makefile.config
> @@ -969,6 +969,21 @@ ifdef BUILD_NONDISTRO
> endif
> endif
>
> +ifndef NO_LLVM
NO_LIBLLVM
> + $(call feature_check,llvm)
> + ifeq ($(feature-llvm), 1)
> + CFLAGS += -DHAVE_LLVM_SUPPORT
> + CXXFLAGS += -DHAVE_LLVM_SUPPORT
> + CXXFLAGS += $(shell $(LLVM_CONFIG) --cxxflags)
> + LIBLLVM = $(shell $(LLVM_CONFIG) --libs all) $(shell $(LLVM_CONFIG) --system-libs)
> + EXTLIBS += -L$(shell $(LLVM_CONFIG) --libdir) $(LIBLLVM)
> + $(call detected,CONFIG_LLVM)
> + else
> + $(warning No libllvm found, slower source file resolution, please install llvm-devel/llvm-dev)
> + NO_LLVM := 1
> + endif
> +endif
> +
> ifndef NO_DEMANGLE
> $(call feature_check,cxa-demangle)
> ifeq ($(feature-cxa-demangle), 1)
> diff --git a/tools/perf/builtin-version.c b/tools/perf/builtin-version.c
> index 398aa53e9e2e..94d8e13a688f 100644
> --- a/tools/perf/builtin-version.c
> +++ b/tools/perf/builtin-version.c
> @@ -65,6 +65,7 @@ static void library_status(void)
> STATUS(HAVE_LIBBFD_SUPPORT, libbfd);
> STATUS(HAVE_DEBUGINFOD_SUPPORT, debuginfod);
> STATUS(HAVE_LIBELF_SUPPORT, libelf);
> + STATUS(HAVE_LLVM_SUPPORT, libllvm);
HAVE_LIBLLVM_SUPPORT to be consistent
> STATUS(HAVE_LIBNUMA_SUPPORT, libnuma);
> STATUS(HAVE_LIBNUMA_SUPPORT, numa_num_possible_cpus);
> STATUS(HAVE_LIBPERL_SUPPORT, libperl);
> diff --git a/tools/perf/util/Build b/tools/perf/util/Build
> index da64efd8718f..32c4e5e634ed 100644
> --- a/tools/perf/util/Build
> +++ b/tools/perf/util/Build
> @@ -226,6 +226,7 @@ perf-$(CONFIG_CXX_DEMANGLE) += demangle-cxx.o
> perf-y += demangle-ocaml.o
> perf-y += demangle-java.o
> perf-y += demangle-rust.o
> +perf-$(CONFIG_LLVM) += llvm-c-helpers.o
>
> ifdef CONFIG_JITDUMP
> perf-$(CONFIG_LIBELF) += jitdump.o
> diff --git a/tools/perf/util/llvm-c-helpers.cpp b/tools/perf/util/llvm-c-helpers.cpp
> new file mode 100644
> index 000000000000..cde3f1f03af0
> --- /dev/null
> +++ b/tools/perf/util/llvm-c-helpers.cpp
> @@ -0,0 +1,130 @@
> +// SPDX-License-Identifier: GPL-2.0
> +
> +/*
> + * Must come before the linux/compiler.h include, which defines several
> + * macros (e.g. noinline) that conflict with compiler builtins used
> + * by LLVM.
> + */
> +#pragma GCC diagnostic push
> +#pragma GCC diagnostic ignored "-Wunused-parameter" /* Needed for LLVM <= 15 */
> +#include <llvm/DebugInfo/Symbolize/Symbolize.h>
> +#pragma GCC diagnostic pop
> +
> +#include <stdio.h>
> +#include <sys/types.h>
> +#include <linux/compiler.h>
> +#include "symbol_conf.h"
> +#include "llvm-c-helpers.h"
> +
> +using namespace llvm;
> +using llvm::symbolize::LLVMSymbolizer;
> +
> +/*
> + * Allocate a static LLVMSymbolizer, which will live to the end of the program.
> + * Unlike the bfd paths, LLVMSymbolizer has its own cache, so we do not need
> + * to store anything in the dso struct.
> + */
> +static LLVMSymbolizer *get_symbolizer()
> +{
> + static LLVMSymbolizer *instance = nullptr;
> + if (instance == nullptr) {
> + LLVMSymbolizer::Options opts;
> + /*
> + * LLVM sometimes demangles slightly different from the rest
> + * of the code, and this mismatch can cause new_inline_sym()
> + * to get confused and mark non-inline symbol as inlined
> + * (since the name does not properly match up with base_sym).
> + * Thus, disable the demangling and let the rest of the code
> + * handle it.
> + */
> + opts.Demangle = false;
> + instance = new LLVMSymbolizer(opts);
> + }
> + return instance;
> +}
> +
> +/* Returns 0 on error, 1 on success. */
> +static int extract_file_and_line(const DILineInfo& line_info, char **file,
> + unsigned int *line)
> +{
> + if (file) {
> + if (line_info.FileName == "<invalid>") {
> + /* Match the convention of libbfd. */
> + *file = nullptr;
> + } else {
> + /* The caller expects to get something it can free(). */
> + *file = strdup(line_info.FileName.c_str());
> + if (*file == nullptr)
> + return 0;
> + }
> + }
> + if (line)
> + *line = line_info.Line;
> + return 1;
> +}
> +
> +extern "C"
> +int llvm_addr2line(const char *dso_name, u64 addr,
> + char **file, unsigned int *line,
> + bool unwind_inlines,
> + llvm_a2l_frame** inline_frames)
> +{
> + LLVMSymbolizer *symbolizer = get_symbolizer();
> + object::SectionedAddress sectioned_addr = {
> + addr,
> + object::SectionedAddress::UndefSection
> + };
> +
> + if (unwind_inlines) {
> + Expected<DIInliningInfo> res_or_err =
> + symbolizer->symbolizeInlinedCode(dso_name,
> + sectioned_addr);
> + if (!res_or_err)
> + return 0;
> + unsigned num_frames = res_or_err->getNumberOfFrames();
> + if (num_frames == 0)
> + return 0;
> +
> + if (extract_file_and_line(
> + res_or_err->getFrame(0), file, line) == 0)
Please don't break this line or at least align 'file' with res_or_err.
> + return 0;
> +
> + *inline_frames = (llvm_a2l_frame*)malloc(
> + sizeof(**inline_frames) * num_frames);
Do we really need to cast here (my C++ is super rusty), and why not
calloc()?
> + if (*inline_frames == nullptr)
> + return 0;
> +
> + for (unsigned i = 0; i < num_frames; ++i) {
> + const DILineInfo& src = res_or_err->getFrame(i);
Add blank line here
> + llvm_a2l_frame& dst = (*inline_frames)[i];
> + if (src.FileName == "<invalid>")
> + /* Match the convention of libbfd. */
> + dst.filename = nullptr;
> + else
> + dst.filename = strdup(src.FileName.c_str());
> + dst.funcname = strdup(src.FunctionName.c_str());
If strdup fails, users will cope?
> + dst.line = src.Line;
> +
> + if (dst.filename == nullptr ||
> + dst.funcname == nullptr) {
> + for (unsigned j = 0; j <= i; ++j) {
> + free((*inline_frames)[j].filename);
> + free((*inline_frames)[j].funcname);
Normally we use zfree()
> + }
> + free(*inline_frames);
> + return 0;
> + }
> + }
> +
> + return num_frames;
> + } else {
> + if (inline_frames)
> + *inline_frames = nullptr;
> +
> + Expected<DILineInfo> res_or_err =
> + symbolizer->symbolizeCode(dso_name, sectioned_addr);
same line?
> + if (!res_or_err)
> + return 0;
> + return extract_file_and_line(*res_or_err, file, line);
> + }
> +}
> diff --git a/tools/perf/util/llvm-c-helpers.h b/tools/perf/util/llvm-c-helpers.h
> new file mode 100644
> index 000000000000..1b28cdc9f9b7
> --- /dev/null
> +++ b/tools/perf/util/llvm-c-helpers.h
> @@ -0,0 +1,47 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef __PERF_LLVM_C_HELPERS
> +#define __PERF_LLVM_C_HELPERS 1
> +
> +/*
> + * Helpers to call into LLVM C++ code from C, for the parts that do not have
> + * C APIs.
> + */
> +
> +#ifdef __cplusplus
> +extern "C" {
> +#endif
> +
> +struct llvm_a2l_frame {
> + char *filename;
> + char *funcname;
> + unsigned int line;
> +};
> +
> +/*
> + * Implement addr2line() using libLLVM. LLVM is a C++ API, and
> + * many of the linux/ headers cannot be included in a C++ compile unit,
> + * so we need to make a little bridge code here. llvm_addr2line() will
> + * convert the inline frame information from LLVM's internal structures
> + * and put them into a flat array given in inline_frames. The caller
> + * is then responsible for taking that array and convert it into perf's
> + * regular inline frame structures (which depend on e.g. struct list_head).
> + *
> + * If the address could not be resolved, or an error occurred (e.g. OOM),
> + * returns 0. Otherwise, returns the number of inline frames (which means 1
> + * if the address was not part of an inlined function). If unwind_inlines
> + * is set and the return code is nonzero, inline_frames will be set to
> + * a newly allocated array with that length. The caller is then responsible
> + * for freeing both the strings and the array itself.
> + */
> +int llvm_addr2line(const char *dso_name,
> + u64 addr,
> + char **file,
> + unsigned int *line,
> + bool unwind_inlines,
> + struct llvm_a2l_frame **inline_frames);
> +
> +#ifdef __cplusplus
> +}
> +#endif
> +
> +#endif /* __PERF_LLVM_C_HELPERS */
> diff --git a/tools/perf/util/srcline.c b/tools/perf/util/srcline.c
> index 9d670d8c1c08..0505b4c16608 100644
> --- a/tools/perf/util/srcline.c
> +++ b/tools/perf/util/srcline.c
> @@ -16,6 +16,9 @@
> #include "util/debug.h"
> #include "util/callchain.h"
> #include "util/symbol_conf.h"
> +#ifdef HAVE_LLVM_SUPPORT
> +#include "util/llvm-c-helpers.h"
> +#endif
> #include "srcline.h"
> #include "string2.h"
> #include "symbol.h"
> @@ -130,7 +133,59 @@ static struct symbol *new_inline_sym(struct dso *dso,
>
> #define MAX_INLINE_NEST 1024
>
> -#ifdef HAVE_LIBBFD_SUPPORT
> +#ifdef HAVE_LLVM_SUPPORT
> +
> +static void free_llvm_inline_frames(struct llvm_a2l_frame *inline_frames,
> + int num_frames)
> +{
> + if (inline_frames != NULL) {
> + for (int i = 0; i < num_frames; ++i) {
> + free(inline_frames[i].filename);
> + free(inline_frames[i].funcname);
zfree()?
> + }
> + free(inline_frames);
> + }
> +}
> +
> +static int addr2line(const char *dso_name, u64 addr,
> + char **file, unsigned int *line, struct dso *dso,
> + bool unwind_inlines, struct inline_node *node,
> + struct symbol *sym)
> +{
> + struct llvm_a2l_frame *inline_frames = NULL;
> + int num_frames = llvm_addr2line(dso_name, addr, file, line,
> + node && unwind_inlines, &inline_frames);
> +
> + if (num_frames == 0 || !inline_frames) {
> + /* Error, or we didn't want inlines. */
> + return num_frames;
> + }
> +
> + for (int i = 0; i < num_frames; ++i) {
> + struct symbol *inline_sym =
> + new_inline_sym(dso, sym, inline_frames[i].funcname);
> + char *srcline = NULL;
> +
> + if (inline_frames[i].filename)
> + srcline = srcline_from_fileline(
move first arg to same line of function call
> + inline_frames[i].filename,
> + inline_frames[i].line);
> + if (inline_list__append(inline_sym, srcline, node) != 0) {
> + free_llvm_inline_frames(inline_frames, num_frames);
> + return 0;
> + }
> + }
> + free_llvm_inline_frames(inline_frames, num_frames);
> +
> + return num_frames;
> +}
> +
> +void dso__free_a2l(struct dso *)
> +{
> + /* Nothing to free. */
> +}
> +
> +#elif defined(HAVE_LIBBFD_SUPPORT)
>
> /*
> * Implement addr2line using libbfd.
> --
> 2.45.1
On Thu, May 23, 2024 at 03:36:22PM -0300, Arnaldo Carvalho de Melo wrote:
> Makefile.config:982: No libllvm found, slower source file resolution, please install llvm-devel/llvm-dev
>
> ... llvm: [ OFF ]
>
> But maybe use "libllvm"?
I think I was trying to be consistent with the previous patch using LLVM
(something with Clang and eBPF, I think?), which used llvm and not
libllvm for the name here :-) And I don't think upstream actually uses
the libllvm name much (e.g. as you can see, it's llvm-devel, not
libllvm-devel). But I can change it if you think it's a better name;
I don't mind much either way.
> So mostly the above, and:
>
> root@x1:~# perf probe -x /usr/bin/find -L find | grep fts_read
> 44 while ( (errno=0, ent=fts_read (p)) != NULL )
> /* fts_read returned NULL; distinguish between "finished" and "error". */
> root@x1:~#
This part went a bit over my head, I'll just assume it's good. :-)
> So the libllvm is even producing a better result, showing those inlines
> not seen in the BFD based output.
I think maybe you need to give --inlines to BFD objdump to get it to
show inlines. (It can be useful in perf annotate sometimes.)
But annoyingly enough, LLVM objdump does not understand the flag.
> 100x speedup, looks like a win! 8-)
Great, thanks for testing. :-)
> Thanks a lot, the comments I made on the patch are mostly coding style,
> please consider them, but I wouldn't get in the way for this patch to
> get merged because of that albeit would be nice to try to fit in more
> nicely with the existing source code base.
Sure, I'll make a v6 with the requested changes. Just some
questions/comments below.
>> + *inline_frames = (llvm_a2l_frame*)malloc(
>> + sizeof(**inline_frames) * num_frames);
> Do we really need to cast here (my C++ is super rusty), and why not
> calloc()?
Yes, C++ does not have implicit cast-from-void*. The C++ way would be
new[] or use std::vector, but, that would make it impossible for the
caller to free(). I can use calloc, it won't change much either way.
> > + dst.funcname = strdup(src.FunctionName.c_str());
> If strdup fails, users will cope?
Yes, same as dst.filename. new_inline_sym() (which is where we give this
data to) has an explicit check for if (!funcname).
> > + Expected<DILineInfo> res_or_err =
> > + symbolizer->symbolizeCode(dso_name, sectioned_addr);
> same line?
That takes it way over 80 characters, is that OK?
/* Steinar */