[llvm-branch-commits] [clang][CallGraphSection] Add call graph section option and docs (PR #117037)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Nov 20 11:54:24 PST 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-clang
Author: Prabhuk (Prabhuk)
<details>
<summary>Changes</summary>
This documents the support for computing, storing, and restoring call
graphs with LLVM. This adds the options and the design documentation
for computing and storing the call graphs. To enable, use
`-fcall-graph-section`.
Inferring indirect call targets from a binary is challenging without
source-level information. Hence, the reconstruction of a fine-grained
call graph from the binary is unfeasible for indirect/virtual calls.
To address this, designed solution is to collect the necessary information
to construct the call graph while the source information is present, and
store it in a non-code section of the binary.
Original RFC: https://lists.llvm.org/pipermail/llvm-dev/2021-June/151044.html
Updated RFC: https://lists.llvm.org/pipermail/llvm-dev/2021-July/151739.html
---
Full diff: https://github.com/llvm/llvm-project/pull/117037.diff
6 Files Affected:
- (added) clang/docs/CallGraphSection.rst (+251)
- (modified) clang/include/clang/Basic/CodeGenOptions.def (+2)
- (modified) clang/include/clang/Driver/Options.td (+4)
- (modified) clang/lib/CodeGen/BackendUtil.cpp (+1)
- (modified) clang/lib/Driver/ToolChains/Clang.cpp (+4)
- (added) clang/test/Driver/call-graph-section.c (+5)
``````````diff
diff --git a/clang/docs/CallGraphSection.rst b/clang/docs/CallGraphSection.rst
new file mode 100644
index 00000000000000..d5db507f9ca9a6
--- /dev/null
+++ b/clang/docs/CallGraphSection.rst
@@ -0,0 +1,251 @@
+==================
+Call Graph Section
+==================
+
+Introduction
+============
+
+With ``-fcall-graph-section``, the compiler will create a call graph section
+in the object file. It will include type identifiers for indirect calls and
+targets. This information can be used to map indirect calls to their receivers
+with matching types. A complete and high-precision call graph can be
+reconstructed by complementing this information with disassembly
+(see ``llvm-objdump --call-graph-info``).
+
+Semantics
+=========
+
+A coarse-grained, type-agnostic call graph may allow indirect calls to target
+any function in the program. This approach ensures completeness since no
+indirect call edge is missing. However, it is generally poor in precision
+due to having unneeded edges.
+
+A call graph section provides type identifiers for indirect calls and targets.
+This information can be used to restrict the receivers of an indirect target to
+indirect calls with matching type. Consequently, the precision for indirect
+call edges are improved while maintaining the completeness.
+
+The ``llvm-objdump`` utility provides a ``--call-graph-info`` option to extract
+full call graph information by parsing the content of the call graph section
+and disassembling the program for complementary information, e.g., direct
+calls.
+
+Section layout
+==============
+
+A call graph section consists of zero or more call graph entries.
+Each entry contains information on a function and its indirect calls.
+
+An entry of a call graph section has the following layout in the binary:
+
++---------------------+-----------------------------------------------------------------------+
+| Element | Content |
++=====================+=======================================================================+
+| FormatVersionNumber | Format version number. |
++---------------------+-----------------------------------------------------------------------+
+| FunctionEntryPc | Function entry address. |
++---------------------+-----------------------------------+-----------------------------------+
+| | A flag whether the function is an | - 0: not an indirect target |
+| FunctionKind | indirect target, and if so, | - 1: indirect target, unknown id |
+| | whether its type id is known. | - 2: indirect target, known id |
++---------------------+-----------------------------------+-----------------------------------+
+| FunctionTypeId | Type id for the indirect target. Present only when FunctionKind is 2. |
++---------------------+-----------------------------------------------------------------------+
+| CallSiteCount | Number of type id to indirect call site mappings that follow. |
++---------------------+-----------------------------------------------------------------------+
+| CallSiteList | List of type id and indirect call site pc pairs. |
++---------------------+-----------------------------------------------------------------------+
+
+Each element in an entry (including each element of the contained lists and
+pairs) occupies 64-bit space.
+
+The format version number is repeated per entry to support concatenation of
+call graph sections with different format versions by the linker.
+
+As of now, the only supported format version is described above and has version
+number 0.
+
+Type identifiers
+================
+
+The type for an indirect call or target is the function signature.
+The mapping from a type to an identifier is an ABI detail.
+In the current experimental implementation, an identifier of type T is
+computed as follows:
+
+ - Obtain the generalized mangled name for “typeinfo name for T”.
+ - Compute MD5 hash of the name as a string.
+ - Reinterpret the first 8 bytes of the hash as a little-endian 64-bit integer.
+
+To avoid mismatched pointer types, generalizations are applied.
+Pointers in return and argument types are treated as equivalent as long as the
+qualifiers for the type they point to match.
+For example, ``char*``, ``char**``, and ``int*`` are considered equivalent
+types. However, ``char*`` and ``const char*`` are considered separate types.
+
+Missing type identifiers
+========================
+
+For functions, two cases need to be considered. First, if the compiler cannot
+deduce a type id for an indirect target, it will be listed as an indirect target
+without a type id. Second, if an object without a call graph section gets
+linked, the final call graph section will lack information on functions from
+the object. For completeness, these functions need to be taken as receiver to
+any indirect call regardless of their type id.
+``llvm-objdump --call-graph-info`` lists these functions as indirect targets
+with `UNKNOWN` type id.
+
+For indirect calls, current implementation guarantees a type id for each
+compiled call. However, if an object without a call graph section gets linked,
+no type id will be present for its indirect calls. For completeness, these calls
+need to be taken to target any indirect target regardless of their type id. For
+indirect calls, ``llvm-objdump --call-graph-info`` prints 1) a complete list of
+indirect calls, 2) type id to indirect call mappings. The difference of these
+lists allow to deduce the indirect calls with missing type ids.
+
+TODO: measure and report the ratio of missed type ids
+
+Performance
+===========
+
+A call graph section does not affect the executable code and does not occupy
+memory during process execution. Therefore, there is no performance overhead.
+
+The scheme has not yet been optimized for binary size.
+
+TODO: measure and report the increase in the binary size
+
+Example
+=======
+
+For example, consider the following C++ code:
+
+.. code-block:: cpp
+
+ namespace {
+ // Not an indirect target
+ void foo() {}
+ }
+
+ // Indirect target 1
+ void bar() {}
+
+ // Indirect target 2
+ int baz(char a, float *b) {
+ return 0;
+ }
+
+ // Indirect target 3
+ int main() {
+ char a;
+ float b;
+ void (*fp_bar)() = bar;
+ int (*fp_baz1)(char, float*) = baz;
+ int (*fp_baz2)(char, float*) = baz;
+
+ // Indirect call site 1
+ fp_bar();
+
+ // Indirect call site 2
+ fp_baz1(a, &b);
+
+ // Indirect call site 3: shares the type id with indirect call site 2
+ fp_baz2(a, &b);
+
+ // Direct call sites
+ foo();
+ bar();
+ baz(a, &b);
+
+ return 0;
+ }
+
+Following will compile it with a call graph section created in the binary:
+
+.. code-block:: bash
+
+ $ clang -fcall-graph-section example.cpp
+
+During the construction of the call graph section, the type identifiers are
+computed as follows:
+
++---------------+-----------------------+----------------------------+----------------------------+
+| Function name | Generalized signature | Mangled name (itanium ABI) | Numeric type id (md5 hash) |
++===============+=======================+============================+============================+
+| bar | void () | _ZTSFvvE.generalized | f85c699bb8ef20a2 |
++---------------+-----------------------+----------------------------+----------------------------+
+| baz | int (char, void*) | _ZTSFicPvE.generalized | e3804d2a7f2b03fe |
++---------------+-----------------------+----------------------------+----------------------------+
+| main | int () | _ZTSFivE.generalized | a9494def81a01dc |
++---------------+-----------------------+----------------------------+----------------------------+
+
+The call graph section will have the following content:
+
++---------------+-----------------+--------------+----------------+---------------+--------------------------------------+
+| FormatVersion | FunctionEntryPc | FunctionKind | FunctionTypeId | CallSiteCount | CallSiteList |
++===============+=================+==============+================+===============+======================================+
+| 0 | EntryPc(foo) | 0 | (empty) | 0 | (empty) |
++---------------+-----------------+--------------+----------------+---------------+--------------------------------------+
+| 0 | EntryPc(bar) | 2 | TypeId(bar) | 0 | (empty) |
++---------------+-----------------+--------------+----------------+---------------+--------------------------------------+
+| 0 | EntryPc(baz) | 2 | TypeId(baz) | 0 | (empty) |
++---------------+-----------------+--------------+----------------+---------------+--------------------------------------+
+| 0 | EntryPc(main) | 2 | TypeId(main) | 3 | * TypeId(bar), CallSitePc(fp_bar()) |
+| | | | | | * TypeId(baz), CallSitePc(fp_baz1()) |
+| | | | | | * TypeId(baz), CallSitePc(fp_baz2()) |
++---------------+-----------------+--------------+----------------+---------------+--------------------------------------+
+
+
+The ``llvm-objdump`` utility can parse the call graph section and disassemble
+the program to provide complete call graph information. This includes any
+additional call sites from the binary:
+
+.. code-block:: bash
+
+ $ llvm-objdump --call-graph-info a.out
+
+ # Comments are not a part of the llvm-objdump's output but inserted for clarifications.
+
+ a.out: file format elf64-x86-64
+ # These warnings are due to the functions and the indirect calls coming from linked objects.
+ llvm-objdump: warning: 'a.out': callgraph section does not have type ids for 3 indirect calls
+ llvm-objdump: warning: 'a.out': callgraph section does not have information for 10 functions
+
+ # Unknown targets are the 10 functions the warnings mention.
+ INDIRECT TARGET TYPES (TYPEID [FUNC_ADDR,])
+ UNKNOWN 401000 401100 401234 401050 401090 4010d0 4011d0 401020 401060 401230
+ a9494def81a01dc 401150 # main()
+ f85c699bb8ef20a2 401120 # bar()
+ e3804d2a7f2b03fe 401130 # baz()
+
+ # Notice that the call sites share the same type id as target functions
+ INDIRECT CALL TYPES (TYPEID [CALL_SITE_ADDR,])
+ f85c699bb8ef20a2 401181 # Indirect call site 1 (fp_bar())
+ e3804d2a7f2b03fe 401191 4011a1 # Indirect call site 2 and 3 (fp_baz1() and fp_baz2())
+
+ INDIRECT CALL SITES (CALLER_ADDR [CALL_SITE_ADDR,])
+ 401000 401012 # _init
+ 401150 401181 401191 4011a1 # main calls fp_bar(), fp_baz1(), fp_baz2()
+ 4011d0 401215 # __libc_csu_init
+ 401020 40104a # _start
+
+ DIRECT CALL SITES (CALLER_ADDR [(CALL_SITE_ADDR, TARGET_ADDR),])
+ 4010d0 4010e2 401060 # __do_global_dtors_aux
+ 401150 4011a6 401110 4011ab 401120 4011ba 401130 # main calls foo(), bar(), baz()
+ 4011d0 4011fd 401000 # __libc_csu_init
+
+ FUNCTIONS (FUNC_ENTRY_ADDR, SYM_NAME)
+ 401000 _init
+ 401100 frame_dummy
+ 401234 _fini
+ 401050 _dl_relocate_static_pie
+ 401090 register_tm_clones
+ 4010d0 __do_global_dtors_aux
+ 401110 _ZN12_GLOBAL__N_13fooEv # (anonymous namespace)::foo()
+ 401150 main # main
+ 4011d0 __libc_csu_init
+ 401020 _start
+ 401060 deregister_tm_clones
+ 401120 _Z3barv # bar()
+ 401130 _Z3bazcPf # baz(char, float*)
+ 401230 __libc_csu_fini
diff --git a/clang/include/clang/Basic/CodeGenOptions.def b/clang/include/clang/Basic/CodeGenOptions.def
index 4cf22c4ee08ce0..a3ee5acef4217d 100644
--- a/clang/include/clang/Basic/CodeGenOptions.def
+++ b/clang/include/clang/Basic/CodeGenOptions.def
@@ -78,6 +78,8 @@ CODEGENOPT(EnableNoundefAttrs, 1, 0) ///< Enable emitting `noundef` attributes o
CODEGENOPT(DebugPassManager, 1, 0) ///< Prints debug information for the new
///< pass manager.
CODEGENOPT(DisableRedZone , 1, 0) ///< Set when -mno-red-zone is enabled.
+CODEGENOPT(CallGraphSection, 1, 0) ///< Emit a call graph section into the
+ ///< object file.
CODEGENOPT(EmitCallSiteInfo, 1, 0) ///< Emit call site info only in the case of
///< '-g' + 'O>0' level.
CODEGENOPT(IndirectTlsSegRefs, 1, 0) ///< Set when -mno-tls-direct-seg-refs
diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td
index f2f9c20c9bc264..77c78ea0724da5 100644
--- a/clang/include/clang/Driver/Options.td
+++ b/clang/include/clang/Driver/Options.td
@@ -4291,6 +4291,10 @@ defm data_sections : BoolFOption<"data-sections",
PosFlag<SetTrue, [], [ClangOption, CC1Option],
"Place each data in its own section">,
NegFlag<SetFalse>>;
+defm call_graph_section : BoolFOption<"call-graph-section",
+ CodeGenOpts<"CallGraphSection">, DefaultFalse,
+ PosFlag<SetTrue, [], [CC1Option], "Emit a call graph section">,
+ NegFlag<SetFalse>>;
defm stack_size_section : BoolFOption<"stack-size-section",
CodeGenOpts<"StackSizeSection">, DefaultFalse,
PosFlag<SetTrue, [], [ClangOption, CC1Option],
diff --git a/clang/lib/CodeGen/BackendUtil.cpp b/clang/lib/CodeGen/BackendUtil.cpp
index bf9b04f02e9f44..4b81d229cda3ca 100644
--- a/clang/lib/CodeGen/BackendUtil.cpp
+++ b/clang/lib/CodeGen/BackendUtil.cpp
@@ -455,6 +455,7 @@ static bool initTargetOptions(DiagnosticsEngine &Diags,
Options.StackUsageOutput = CodeGenOpts.StackUsageOutput;
Options.EmitAddrsig = CodeGenOpts.Addrsig;
Options.ForceDwarfFrameSection = CodeGenOpts.ForceDwarfFrameSection;
+ Options.EmitCallGraphSection = CodeGenOpts.CallGraphSection;
Options.EmitCallSiteInfo = CodeGenOpts.EmitCallSiteInfo;
Options.EnableAIXExtendedAltivecABI = LangOpts.EnableAIXExtendedAltivecABI;
Options.XRayFunctionIndex = CodeGenOpts.XRayFunctionIndex;
diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp
index d3eec9fea0d498..d749f0fc5375a8 100644
--- a/clang/lib/Driver/ToolChains/Clang.cpp
+++ b/clang/lib/Driver/ToolChains/Clang.cpp
@@ -6609,6 +6609,10 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
CmdArgs.push_back(A->getValue());
}
+ if (Args.hasFlag(options::OPT_fcall_graph_section,
+ options::OPT_fno_call_graph_section, false))
+ CmdArgs.push_back("-fcall-graph-section");
+
Args.addOptInFlag(CmdArgs, options::OPT_fstack_size_section,
options::OPT_fno_stack_size_section);
diff --git a/clang/test/Driver/call-graph-section.c b/clang/test/Driver/call-graph-section.c
new file mode 100644
index 00000000000000..5832aa6754137f
--- /dev/null
+++ b/clang/test/Driver/call-graph-section.c
@@ -0,0 +1,5 @@
+// RUN: %clang -### -S -fcall-graph-section %s 2>&1 | FileCheck --check-prefix=CALL-GRAPH-SECTION %s
+// RUN: %clang -### -S -fcall-graph-section -fno-call-graph-section %s 2>&1 | FileCheck --check-prefix=NO-CALL-GRAPH-SECTION %s
+
+// CALL-GRAPH-SECTION: "-fcall-graph-section"
+// NO-CALL-GRAPH-SECTION-NOT: "-fcall-graph-section"
``````````
</details>
https://github.com/llvm/llvm-project/pull/117037
More information about the llvm-branch-commits
mailing list