Postgresql中的C/C++混编（JIT）
1 Postgresql编译JIT
整体上看使用了GCC、G++编译文件，最后用G++汇总：
GCC编译的三个.o文件llvmjit、llvmjit_deform、llvmjit_expr
llvmjit.c -> llvmjit.o
gcc -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wshadow=compatible-local -Wformat-security -fno-strict-aliasing -fwrapv -fexcess-precision=standard -Wno-format-truncation -Wno-stringop-truncation -g -ggdb -O0 -g3 -gdwarf-2 -Wno-deprecated-declarations -fPIC -fvisibility=hidden -D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_CONSTANT_MACROS -D_GNU_SOURCE -DBUILD_EXAMPLES -I/data02/mingjie/bin/llvm15/include -I../../../../src/include -D_GNU_SOURCE -I/usr/include/libxml2 -c -o llvmjit.o llvmjit.c -MMD -MP -MF .deps/llvmjit.Po
llvmjit_deform.c -> llvmjit_deform.o
gcc -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wshadow=compatible-local -Wformat-security -fno-strict-aliasing -fwrapv -fexcess-precision=standard -Wno-format-truncation -Wno-stringop-truncation -g -ggdb -O0 -g3 -gdwarf-2 -Wno-deprecated-declarations -fPIC -fvisibility=hidden -D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_CONSTANT_MACROS -D_GNU_SOURCE -DBUILD_EXAMPLES -I/data02/mingjie/bin/llvm15/include -I../../../../src/include -D_GNU_SOURCE -I/usr/include/libxml2 -c -o llvmjit_deform.o llvmjit_deform.c -MMD -MP -MF .deps/llvmjit_deform.Po
llvmjit_expr.c -> llvmjit_expr.o
gcc -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wshadow=compatible-local -Wformat-security -fno-strict-aliasing -fwrapv -fexcess-precision=standard -Wno-format-truncation -Wno-stringop-truncation -g -ggdb -O0 -g3 -gdwarf-2 -Wno-deprecated-declarations -fPIC -fvisibility=hidden -D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_CONSTANT_MACROS -D_GNU_SOURCE -DBUILD_EXAMPLES -I/data02/mingjie/bin/llvm15/include -I../../../../src/include -D_GNU_SOURCE -I/usr/include/libxml2 -c -o llvmjit_expr.o llvmjit_expr.c -MMD -MP -MF .deps/llvmjit_expr.Po
G++编译的三个.o文件llvmjit_error、llvmjit_inline、llvmjit_wrap
llvmjit_error.cpp -> llvmjit_error.o
g++ -Wall -Wpointer-arith -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wshadow=compatible-local -Wformat-security -fno-strict-aliasing -fwrapv -g -O2 -std=c++14 -fno-rtti -fPIC -fvisibility=hidden -fvisibility-inlines-hidden -D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_CONSTANT_MACROS -D_GNU_SOURCE -DBUILD_EXAMPLES -I/data02/mingjie/bin/llvm15/include -I../../../../src/include -D_GNU_SOURCE -I/usr/include/libxml2 -c -o llvmjit_error.o llvmjit_error.cpp -MMD -MP -MF .deps/llvmjit_error.Po
llvmjit_inline.cpp -> llvmjit_inline.o
g++ -Wall -Wpointer-arith -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wshadow=compatible-local -Wformat-security -fno-strict-aliasing -fwrapv -g -O2 -std=c++14 -fno-rtti -fPIC -fvisibility=hidden -fvisibility-inlines-hidden -D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_CONSTANT_MACROS -D_GNU_SOURCE -DBUILD_EXAMPLES -I/data02/mingjie/bin/llvm15/include -I../../../../src/include -D_GNU_SOURCE -I/usr/include/libxml2 -c -o llvmjit_inline.o llvmjit_inline.cpp -MMD -MP -MF .deps/llvmjit_inline.Po
llvmjit_wrap.cpp -> llvmjit_wrap.o
g++ -Wall -Wpointer-arith -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wshadow=compatible-local -Wformat-security -fno-strict-aliasing -fwrapv -g -O2 -std=c++14 -fno-rtti -fPIC -fvisibility=hidden -fvisibility-inlines-hidden -D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_CONSTANT_MACROS -D_GNU_SOURCE -DBUILD_EXAMPLES -I/data02/mingjie/bin/llvm15/include -I../../../../src/include -D_GNU_SOURCE -I/usr/include/libxml2 -c -o llvmjit_wrap.o llvmjit_wrap.cpp -MMD -MP -MF .deps/llvmjit_wrap.Po
G++汇总的llvmjit.so
llvmjit.so
g++ -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wshadow=compatible-local -Wformat-security -fno-strict-aliasing -fwrapv -fexcess-precision=standard -Wno-format-truncation -Wno-stringop-truncation -g -ggdb -O0 -g3 -gdwarf-2 -Wno-deprecated-declarations -fPIC -fvisibility=hidden -shared -o llvmjit.so llvmjit.o llvmjit_error.o llvmjit_inline.o llvmjit_wrap.o llvmjit_deform.o llvmjit_expr.o -L../../../../src/port -L../../../../src/common -L/data02/mingjie/bin/llvm15/lib -Wl,--as-needed -Wl,-rpath,'/data02/mingjie/pgroot99/pghome/lib',--enable-new-dtags -fvisibility=hidden -lLLVMOrcJIT -lLLVMPasses -lLLVMObjCARCOpts -lLLVMCoroutines -lLLVMipo -lLLVMVectorize -lLLVMLinker -lLLVMIRReader -lLLVMAsmParser -lLLVMFrontendOpenMP -lLLVMJITLink -lLLVMX86TargetMCA -lLLVMMCA -lLLVMX86Disassembler -lLLVMX86AsmParser -lLLVMX86CodeGen -lLLVMCFGuard -lLLVMGlobalISel -lLLVMX86Desc -lLLVMX86Info -lLLVMMCDisassembler -lLLVMSelectionDAG -lLLVMInstrumentation -lLLVMAsmPrinter -lLLVMInterpreter -lLLVMExecutionEngine -lLLVMRuntimeDyld -lLLVMOrcTargetProcess -lLLVMOrcShared -lLLVMCodeGen -lLLVMTarget -lLLVMScalarOpts -lLLVMInstCombine -lLLVMAggressiveInstCombine -lLLVMTransformUtils -lLLVMBitWriter -lLLVMAnalysis -lLLVMProfileData -lLLVMSymbolize -lLLVMDebugInfoPDB -lLLVMDebugInfoMSF -lLLVMDebugInfoDWARF -lLLVMObject -lLLVMTextAPI -lLLVMMCParser -lLLVMMC -lLLVMDebugInfoCodeView -lLLVMBitReader -lLLVMCore -lLLVMRemarks -lLLVMBitstreamReader -lLLVMBinaryFormat -lLLVMSupport -lLLVMDemangle -lrt -ldl -lpthread -lm -lz -lzstd -ltinfo -lxml2
1.1 GCC
gcc -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wshadow=compatible-local -Wformat-security -fno-strict-aliasing -fwrapv -fexcess-precision=standard -Wno-format-truncation -Wno-stringop-truncation -g -ggdb -O0 -g3 -gdwarf-2 -Wno-deprecated-declarations -fPIC -fvisibility=hidden
- llvm-config --cflags:
-D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_CONSTANT_MACROS -D_GNU_SOURCE -DBUILD_EXAMPLES -I/data02/mingjie/bin/llvm15/include
-I…/…/…/…/src/include -D_GNU_SOURCE -I/usr/include/libxml2 -c -o llvmjit.o llvmjit.c -MMD -MP -MF .deps/llvmjit.Po
1.2 G++
g++ -Wall -Wpointer-arith -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wshadow=compatible-local -Wformat-security -fno-strict-aliasing -fwrapv -g -O2 -std=c++14 -fno-rtti -fPIC -fvisibility=hidden -fvisibility-inlines-hidden
- llvm-config --cflags:
-D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_CONSTANT_MACROS -D_GNU_SOURCE -DBUILD_EXAMPLES -I/data02/mingjie/bin/llvm15/include
-I…/…/…/…/src/include -D_GNU_SOURCE -I/usr/include/libxml2 -c -o llvmjit_error.o llvmjit_error.cpp -MMD -MP -MF .deps/llvmjit_error.Po
1.3 llvmjit.so
g++ -Wall -Wmissing-prototypes -Wpointer-arith -Wdeclaration-after-statement -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wshadow=compatible-local -Wformat-security -fno-strict-aliasing -fwrapv -fexcess-precision=standard -Wno-format-truncation -Wno-stringop-truncation -g -ggdb -O0 -g3 -gdwarf-2 -Wno-deprecated-declarations -fPIC -fvisibility=hidden -shared -o llvmjit.so llvmjit.o llvmjit_error.o llvmjit_inline.o llvmjit_wrap.o llvmjit_deform.o llvmjit_expr.o -L…/…/…/…/src/port -L…/…/…/…/src/common
- llvm-config --ldflags:
-L/data02/mingjie/bin/llvm15/lib
-Wl,–as-needed -Wl,-rpath,‘/data02/mingjie/pgroot99/pghome/lib’,–enable-new-dtags -fvisibility=hidden
-
llvm-config --libs:
-lLLVMOrcJIT -lLLVMPasses -lLLVMObjCARCOpts -lLLVMCoroutines -lLLVMipo -lLLVMVectorize -lLLVMLinker -lLLVMIRReader -lLLVMAsmParser -lLLVMFrontendOpenMP -lLLVMJITLink -lLLVMX86TargetMCA -lLLVMMCA -lLLVMX86Disassembler -lLLVMX86AsmParser -lLLVMX86CodeGen -lLLVMCFGuard -lLLVMGlobalISel -lLLVMX86Desc -lLLVMX86Info -lLLVMMCDisassembler -lLLVMSelectionDAG -lLLVMInstrumentation -lLLVMAsmPrinter -lLLVMInterpreter -lLLVMExecutionEngine -lLLVMRuntimeDyld -lLLVMOrcTargetProcess -lLLVMOrcShared -lLLVMCodeGen -lLLVMTarget -lLLVMScalarOpts -lLLVMInstCombine -lLLVMAggressiveInstCombine -lLLVMTransformUtils -lLLVMBitWriter -lLLVMAnalysis -lLLVMProfileData -lLLVMSymbolize -lLLVMDebugInfoPDB -lLLVMDebugInfoMSF -lLLVMDebugInfoDWARF -lLLVMObject -lLLVMTextAPI -lLLVMMCParser -lLLVMMC -lLLVMDebugInfoCodeView -lLLVMBitReader -lLLVMCore -lLLVMRemarks -lLLVMBitstreamReader -lLLVMBinaryFormat -lLLVMSupport -lLLVMDemangle -
llvm-config --system-libs: -lrt -ldl -lpthread -lm -lz -lzstd -ltinfo -lxml2
2 Postgresql加载JIT
编译后生成llvmjit.so，在程序启动时不做加载，在运行时按需加载。
加载位置：jit.c提供的provider_init函数中，对llvmjit.so进行动态加载：
provider_init
load_external_function(path, "_PG_jit_provider_init", true, NULL)
internal_load_library
dlopen(file_scanner->filename, RTLD_NOW | RTLD_GLOBAL)
dlsym(lib_handle, funcname);
2.1 如何使用llvmjit.so的符号表？
G++编译的三个cpp文件会依赖llvm库，llvm是用C++实现的，所以llvmjit.so中会存在大量经过mangling的符号，GCC无法识别，例如：readelf -s llvmjit.so | head -n 100
那怎么让进程找到符号呢？用extern "C"标注那些不要做mangling的符号。
例如：llvmjit.h
（一般在.c中随便写，需要避免mangling的在.h中声明即可）
extern "C"
{
...
extern void llvm_enter_fatal_on_oom(void);
...
extern bool llvm_compile_expr(struct ExprState *state);
struct TupleTableSlotOps;
extern LLVMValueRef slot_compile_deform(struct LLVMJitContext *context, TupleDesc desc,
const struct TupleTableSlotOps *ops, int natts);
...
} /* extern "C" */
在符号表中可以看到这几个符号：
这类符号就可以被GCC编译的程序正常调用，例如：llvm_enter_fatal_on_oom函数在llvmjit_error.cpp中定义，在llvmjit_expr.c中使用。
3 思考
- link llvm后so过大（1.3GB），需要考虑加载时间，加载时IO大文件到内存中还是会严重拉低执行时间的。
- 如果避免每个子进程都要加载，在PG中让父进程加载一次即可。
- LLVM逻辑可以基本都在CPP中实现，给外部提供接口即可，接口只做钩子函数赋值。
- 如果是现存系统，LLVM适合做旁路逻辑，主逻辑还需要存在，需逐步演化。
mangling的一些细节：
声明时禁止mangling：
extern "C" {
int f1(int);
int f2(int);
int f3(int);
};
定义时也可以禁止mangling：
extern "C" {
void p(int){
/* not mangled */
}
};
多层时已最内层为准，会mangling：
extern "C" {
extern "C++" {
void func();
}
}