LLVM Link Time Optimization: design and implementation

The LLVM Link Time Optimizer seeks complete transparency, while doing intermodular optimization, in compiler tool chain. Its main goal is to let developer take advantage of intermodular optimizer without making any significant changes to their makefiles or build system. This is achieved through tight integration with linker. In this model, linker treates LLVM bytecode files like native objects file and allows mixing and matching among them. The linker uses LLVMlto, a dynamically loaded library, to handle LLVM bytecode files. This tight integration between the linker and LLVM optimizer helps to do optimizations that are not possible in other models. The linker input allows optimizer to avoid relying on conservative escape analysis.

Following example illustrates advantage of integrated approach that uses clean interface.

Input source file a.c is compiled into LLVM byte code form.
Input source file main.c is compiled into native object code.


--- a.h ---

extern int foo1(void);

extern void foo2(void);

extern void foo4(void);

--- a.c ---

#include "a.h"



static signed int i = 0;



void foo2(void) {

 i = -1;

}



static int foo3() {

foo4();

return 10;

}



int foo1(void) {

int data = 0;



if (i < 0) { data = foo3(); }



data = data + 42;

return data;

}



--- main.c ---

#include < stdio.h >

#include "a.h"



void foo4(void) {

 printf ("Hi\n");

}



int main() {

 return foo1();

}



--- command lines ---

 $ llvm-gcc4 --emit-llvm -c a.c -o a.o  # <-- a.o is LLVM bytecode file

 $ llvm-gcc4 -c main.c -o main.o # <-- main.o is native object file

 $ llvm-gcc4 a.o main.o -o main # <-- standard link command without any modifications

In this example, the linker recognizes that foo2() is a externally visible symbol defined in LLVM byte code file. This information is collected using readLLVMObjectFile() . Based on this information, linker completes its usual symbol resolution pass and finds that foo2() is not used anywhere. This information is used by LLVM optimizer and it removes foo2(). As soon as foo2() is removed, optimizer recognizes that condition i < 0 is always false, which means foo3() is never used. Hence, optimizer removes foo3() also. And this in turn, enables linker to remove foo4(). This example illustrates advantage of tight integration with linker. Here, optimizer can not remove foo3() without the linker's input.

Compiler driver invokes link time optimizer separately.

In this model link time optimizer is not able to take advantage of information collected during normal linker's symbol resolution phase. In above example, optimizer can not remove foo2() without linker's input because it is externally visible. And this in turn prohibits optimizer from removing foo3().
Use separate tool to collect symbol information from all object file.

In this model, this new separate tool or library replicates linker's capabilities to collect information for link time optimizer. Not only such code duplication is difficult to justify but it also has several other disadvantages. For example, the linking semantics and the features provided by linker on various platform are not unique. This means, this new tool needs to support all such features and platforms in one super tool or one new separate tool per platform is required. This increases maintance cost for link time optimizer significantly, which is not necessary. Plus, this approach requires staying synchronized with linker developements on various platforms, which is not the main focus of link time optimizer. Finally, this approach increases end user's build time due to duplicate work done by this separate tool and linker itself.

The linker collects information about symbol defininitions and uses in various link objects which is more accurate than any information collected by other tools during typical build cycle. The linker collects this information by looking at definitions and uses of symbols in native .o files and using symbol visibility information. The linker also uses user supplied information, such as list of exported symbol. LLVM optimizer collects control flow information, data flow information and knows much more about program structure from optimizer's point of view. Our goal is to take advantage of tight intergration between the linker and optimizer by sharing this information during various linking phases.

The linker first reads all object files in natural order and collects symbol information. This includes native object files as well as LLVM byte code files. In this phase, the linker uses readLLVMObjectFile() to collect symbol information from each LLVM bytecode files and updates its internal global symbol table accordingly. The intent of this interface is to avoid overhead in the non LLVM case, where all input object files are native object files, by putting this code in the error path of the linker. When the linker sees the first llvm .o file, it dlopen()s the dynamic library. This is to allow changes to LLVM part without relinking the linker.

In this stage, the linker resolves symbols using global symbol table information to report undefined symbol errors, read archive members, resolve weak symbols etc... The linker is able to do this seamlessly even though it does not know exact content of input LLVM bytecode files because it uses symbol information provided by readLLVMObjectFile() . If dead code stripping is enabled then linker collects list of live symbols.

After symbol resolution, the linker updates symbol information supplied by LLVM bytecode files appropriately. For example, whether certain LLVM bytecode supplied symbols are used or not. In the example above, the linker reports that foo2() is not used anywhere in the program, including native .o files. This information is used by LLVM interprocedural optimizer. The linker uses optimizeModules() and requests optimized native object file of the LLVM portion of the program.

In this phase, the linker reads optimized native object file and updates internal global symbol table to reflect any changes. Linker also collects information about any change in use of external symbols by LLVM bytecode files. In the examle above, the linker notes that foo4() is not used any more. If dead code striping is enabled then linker refreshes live symbol information appropriately and performs dead code stripping.
After this phase, the linker continues linking as if it never saw LLVM bytecode files.

LLVMlto is a dynamic library that is part of the LLVM tools, and is intended for use by a linker. LLVMlto provides an abstract C++ interface to use the LLVM interprocedural optimizer without exposing details of LLVM internals. The intention is to keep the interface as stable as possible even when the LLVM optimizer continues to evolve.

LLVMSymbol class is used to describe the externally visible functions and global variables, tdefined in LLVM bytecode files, to linker. This includes symbol visibility information. This information is used by linker to do symbol resolution. For example : function foo2() is defined inside a LLVM bytecode module and it is externally visible symbol. This helps linker connect use of foo2() in native object file with future definition of symbol foo2(). The linker will see actual definition of foo2() when it receives optimized native object file in Symbol Resolution after optimization phase. If the linker does not find any use of foo2(), it updates LLVMSymbol visibility information to notify LLVM intermodular optimizer that it is dead. The LLVM intermodular optimizer takes advantage of such information to generate better code.

readLLVMObjectFile() is used by the linker to read LLVM bytecode files and collect LLVMSymbol nformation. This routine also supplies list of externally defined symbols that are used by LLVM bytecode files. Linker uses this symbol information to do symbol resolution. Internally, LLVMlto maintains LLVM bytecode modules in memory. This function also provides list of external references used by bytecode file.