LLVM-GCC is a new compiler optimization technology based on the LLVM (Low-Level Virtual Machine) open source project. In Xcode 3.1.x, LLVM-GCC is manifested in the LLVM-GCC 4.2 compiler.
In LLVM-GCC, LLVM is used to provide an optimizer and code generator plug-in to GCC. This technology is used in the OpenGL software stack in Mac OS X v.10.5. You can take advantage of it in your projects.
You can learn more about LLVM at http://llvm.org.
To learn about the changes made to LLVM-GCC 4.2 in Xcode 3.1.1, see “Xcode 3.1.1 Additions”.
LLVM-GCC 4.2 Overview
LLVM-GCC 4.2 Installation
Using LLVM-GCC 4.2 in Xcode Projects
Using LLVM-GCC 4.2 in UNIX-Based Development Environments
Link-Time Optimization
GCC-4.2 Features Not Available in LLVM-GCC-4.2
LLVM Open-Source Community Information
Submitting Feedback to Apple
Xcode 3.1.1 Additions
LLVM-GCC 4.2 is a standalone compiler that combines the industry standard GCC 4.2 front end with the innovative LLVM optimizer and code generator. Therefore, you can use LLVM-GCC 4.2 as a replacement for GCC 4.2. The two compilers are largely compatible, supporting the same languages and using the same command-line option and architectures. Like GCC 4.2, LLVM-GCC 4.2 supports compilation of C, C++, Objective-C, and Objective-C++ on x86, x86-64, PowerPC, and PowerPC-64. See “GCC-4.2 Features Not Available in LLVM-GCC-4.2” for a list of GCC 4.2 features that LLVM-GCC 4.2 does not support.
The Xcode 3.1 installer places LLVM-GCC 4.2 in <Xcode>/usr/bin (<Xcode> is the directory into which you installed Xcode, which the installer sets as /Developer by default).
In Xcode projects, you use build rules to specify the compiler to use for particular source files. To use LLVM-GCC 4.2 instead of GCC 4.2 in an Xcode project:
Open the target editor by double-clicking the appropriate target.
In the Rules pane choose Rules Specific to Target from the pop-up menu.
Click the Add (+) button.
Configure the new rule:
Chose the appropriate language form the Process pop-up menu.
Choose “llvmgcc42” from the “using” pop-up menu.
You use LLVM-GCC 4.2 in UNIX-based development environments just as you use GCC 4.2. The majority of such environments respect the CC and CXX environment variables. For example, the following listing shows one way to set LLVM-GCC 4.2 as the compiler for C source files.
CC=<Xcode>/usr/bin/llvm-gcc-4.2 make |
Here, <Xcode> is the directory into which you installed Xcode.
If you have several copies of Xcode installed on your computer, may also use `xcode-select -print-path` in CC, CXX, or shell scripts to get the path to the Xcode release you have selected as the one to use for UNIX-based development.
CC=`xcode-select -print-path`/usr/bin/llvm-gcc-4.2 make |
See xcode-select for more information on this facility, but keep in mind that you cannot use all the features of xcode-select to locate Xcode directories on your file system. Only xcode-select -print-path is supported with LLVM-GCC 4.2.
Traditionally, compilers optimize individual source files while building a project. However, this model does not allow compilers to take advantage of symbol information available in separate source files, but the linker collects this information during the link stage. Link-Time Optimization allows the optimizer to look across object files in your program and optimize across file boundaries. For example, one possible optimization is the inlining of a function defined in one file into a function defined in another file. See Listing 1-1 and Listing 1-2.
Listing 1-1 sourcefile1.c: getNumber()
int getNumber() { |
return 42; |
} |
Listing 1-2 sourcefile2.c: processNo(int)
extern int getNumber(); |
int processNo(int i) { |
return i * getNumber(); |
} |
During the link-time optimization stage, the function getNumber() (defined in sourcefile1.c) is inlined inside the function processNo(int) (defined in sourcefile2.c).
To turn on Link-Time Optimization use -O4 command-line option or set the Optimization Level build setting to 4 in the build settings editor for the appropriate target.
LLVM-GCC 4.2 supports a wide range of interprocedural optimizations, including global variable optimization, inter modular inlining, interprocedural constant propagation and dead argument elimination.
When Link-Time Optimization is turned on, the object files LLVM-GCC 4.2 generates are not Mach-O object files. Instead they are object files that use the LLVM bitcode file format (http://llvm.org/docs/BitCodeFormat.html). These files contain LLVM bitcode that the linker optimizes using the LLVM optimizer. Specifically, the Xcode linker links LLVM files together. But it can also link LLVM bitcode files with Mach-O object files to generate an optimized Mach-O binary.
Let’s look at an example that illustrates how the linker links one LLVM bitcode file with a Mach-O object file. In this example, the source files Listing 1-3, Listing 1-4, and Listing 1-5 are compiled and linked with Listing 1-6.
Listing 1-3 header.h
extern int fn1(void); |
extern void fn2(void); |
extern void fn4(void); |
Listing 1-4 sourcefile3.c
#include "header.h" |
static signed int flag = 0; |
void fn2(void) { |
flag = -1; |
} |
static int fn3() { |
fn4(); |
return 10; |
} |
int fn1(void) { |
int data = 0; |
if (flag < 0) { |
data = fn3(); |
} |
data = data + 42; |
return data; |
} |
Listing 1-5 main.c
#include <stdio.h> |
#include "header.h" |
void fn4() { |
printf("Hi\n"); |
} |
int main() { |
return fn1(); |
} |
Listing 1-6 Compiling and linking sourcefile3.c and main.c
$ llvm-gcc-4.2 -O4 -c sourcefile3.c -o sourcefile3.o |
$ llvm-gcc-4.2 -c main.c -o main.o |
$ llvm-gcc-4.2 sourcefile3.o main.o -o main |
The linker recognizes that fn2() is an externally visible symbol defined in LLVM bitcode but unused. This information helps the LLVM optimizer remove fn2(), which, in turn, allows the optimizer to prove that flag is always 0. As a result, the optimizer removes the branch in fn1(), which results in the removal of fn3() because it’s now unused. With the removal of fn3(), the linker concludes that fn4() is also unused, which it then removes.
Note that sourcefile3.o is an LLVM bitcode file and main.o is a Mach-O object file. No additional linker command-line options are required to link them together and generate the optimized main executable.
Link-Time Optimization respects symbol visibility. The link-time optimizer does not remove a symbol that is marked as externally visible while building a dynamic library, even if the optimizer determines that the symbol is unused. If a function is not marked as externally visible, the optimizer is free to change its signature by removing dead arguments or removing the function entirely, for example.
In a dynamic library, symbols are considered externally visible by default, but in a standalone executable, symbols other than main() are internal by default. This behavior allows the optimizer to remove fn2() in the example in “Mixing LLVM Bitcode Files with Mach-O Files”. For more information on symbol visibility, see Controlling Symbol Visibility.
These are GCC 4.2 features that are not available in LLVM-GCC 4.2:
Profile-guided optimization
Auto vectorization
Stack canaries
Inline assembly that uses x87 registers
Generating debug information for optimized code
LLVM is supported by an active open-source community, whose home is located at http://llvm.org. You can interact directly with LLVM developers through the mailing lists available at this website.
Reminder: Please do not discuss Xcode information covered under your NDA with Apple in these mailing lists.
Apple Developer Connection members can submit bug reports and technical-support inquiries:
To submit bug reports, visit https://bugreport.apple.com.
Send technical-support inquiries to dts@apple.com.
If you’re not an ADC member, vist http://developer.apple.com to sign up.
For a list of Apple mailing lists, visit http://lists.apple.com.
These are the improvements made to LLVM-GCC 4.2 in Xcode 3.1.1:
OpenMP support (with the inclusion of atomic built-in functions)
Object-size checking; for example, -D_FORTIFY_SOURCE=2
LLVM bitcode format change to align files to 16 bytes and to contain the CPU type to fix a problem with link-time optimization (LTO)
Several ABI fixes:
In the 64-bit ABI, nonfragile Objective-C instance-variable sizes are calculated at runtime instead of compile time.
Fixed problem in which some Objective-C metadata is incorrectly generated for nonoptimized code.
Fixed minor vector-related ABI mismatches between GCC 4.2 and LLVM-GCC 4.2 on the x86-64 architecture.
Fixed a problem in which the alignment of a structure of a field assumes the alignment of the base of the structure.
Fixed a problem in which explicit trailing padding bytes of a structure with a trailing union didn’t fill out the structure completely.
Debugging improvements and bug fixes:
Improved compilation time when the -O0 and -g flags are used:
Unreferenced static functions are no longer emitted
Explicit names for variables in LLVM bitcode files are not used
DWARF output improvements to prevent unnecessary information duplication
Reduction of the amount of unnecessary work the compiler does when compiling for debugging
Improvement of the compiler’s memory management
Debug information is emitted for data-only files
The linker accepts some command-line options (such as -mllvm) to turn on LTO debugging, statistics dumping, and so on
Fixed memory corruption problem with precompiled headers (PCH)
Last updated: 2008-09-09