Edit

Share via

Profile-guided optimizations

  • Article

Profile-guided optimization (PGO) lets you optimize a whole executable file, where the optimizer uses data from test runs of the .exe or .dll file. The data represents the likely performance of the program in a production environment.

Profile-guided optimizations are only available for x86, x64, or ARM64 native targets. Profile-guided optimizations aren't available for executable files that run on the common language runtime. Even if you produce an assembly with mixed native and managed code (by using the /clr compiler option), you can't use profile-guided optimization on just the native code. If you attempt to build a project with these options set in the IDE, a build error results.

Note

Information that's gathered from profiling test runs overrides optimizations that would otherwise be in effect if you specify /Ob, /Os, or /Ot. For more information, see /Ob (Inline Function Expansion) and /Os, /Ot (Favor Small Code, Favor Fast Code).

Steps to optimize your app

To use profile-guided optimization, follow these steps to optimize your app:

  • Compile one or more source code files with /GL.

    Each module built with /GL can be examined during profile-guided optimization test runs to capture run-time behavior. Every module in a profile-guided optimization build doesn't have to be compiled with /GL. However, only those modules compiled with /GL are instrumented and later available for profile-guided optimizations.

  • Link using /LTCG and /GENPROFILE or /FASTGENPROFILE.

    Using both /LTCG and /GENPROFILE or /FASTGENPROFILE creates a .pgd file when the instrumented app is run. After test-run data is added to the .pgd file, it can be used as input to the next link step (creating the optimized image). When specifying /GENPROFILE, you can optionally add a PGD=filename argument to specify a nondefault name or location for the .pgd file. The combination of /LTCG and /GENPROFILE or /FASTGENPROFILE linker options replaces the deprecated /LTCG:PGINSTRUMENT linker option.

  • Profile the application.

    Each time a profiled EXE session ends, or a profiled DLL is unloaded, a appname!N.pgc file is created. A .pgc file contains information about a particular application test run. appname is the name of your app, and N is a number starting with 1 that's incremented based on the number of other appname!N.pgc files in the directory. You can delete a .pgc file if the test run doesn't represent a scenario you want to optimize.

    During a test run, you can force closure of the currently open .pgc file and the creation of a new .pgc file with the pgosweep utility (for example, when the end of a test scenario doesn't coincide with application shutdown).

    Your application can also directly invoke a PGO function, PgoAutoSweep, to capture the profile data at the point of the call as a .pgc file. It can give you finer control over the code covered by the captured data in your .pgc files. For an example of how to use this function, see the PgoAutoSweep documentation.

    When you create your instrumented build, by default, data collection is done in non-thread-safe mode, which is faster but may be imprecise. By using the EXACT argument to /GENPROFILE or /FASTGENPROFILE, you can specify data collection in thread-safe mode, which is more precise, but slower. This option is also available if you set the deprecated PogoSafeMode environment variable, or the deprecated /POGOSAFEMODE linker option, when you create your instrumented build.

  • Link using /LTCG and /USEPROFILE.

    Use both the /LTCG and /USEPROFILE linker options to create the optimized image. This step takes as input the .pgd file. When you specify /USEPROFILE, you can optionally add a PGD=filename argument to specify a non-default name or location for the .pgd file. You can also specify this name by using the deprecated /PGD linker option. The combination of /LTCG and /USEPROFILE replaces the deprecated /LTCG:PGOPTIMIZE and /LTCG:PGUPDATE linker options.

It's even possible to create the optimized executable file and later determine that additional profiling would be useful to create a more optimized image. If the instrumented image and its .pgd file are available, you can do additional test runs and rebuild the optimized image with the newer .pgd file, by using the same /LTCG and /USEPROFILE linker options.

Note

Both .pgc and .pgd files are binary file types. If stored in a source control system, avoid any automatic transformation that may be made to text files.

Optimizations performed by PGO

The profile-guided optimizations include these checks and improvements:

  • Inlining - For example, if a function A frequently calls function B, and function B is relatively small, then profile-guided optimizations inline function B in function A.

  • Virtual Call Speculation - If a virtual call, or other call through a function pointer, frequently targets a certain function, a profile-guided optimization can insert a conditionally executed direct call to the frequently targeted function, and the direct call can be inlined.

  • Register Allocation - Optimization based on profile data results in better register allocation.

  • Basic Block Optimization - Basic block optimization allows commonly executed basic blocks that temporally execute within a given frame to be placed in the same set of pages (locality). It minimizes the number of pages used, which minimizes memory overhead.

  • Size/Speed Optimization - Functions where the program spends the most execution time can be optimized for speed.

  • Function Layout - Based on the call graph and profiled caller/callee behavior, functions that tend to be along the same execution path are placed in the same section.

  • Conditional Branch Optimization - With the value probes, profile-guided optimizations can find if a given value in a switch statement is used more often than other values. This value can then be pulled out of the switch statement. The same can be done with if...else instructions where the optimizer can order the if...else so that either the if or else block is placed first, depending on which block is more frequently true.

  • Dead Code Separation - Code that isn't called during profiling is moved to a special section that's appended to the end of the set of sections. It effectively keeps this section out of the often-used pages.

  • EH Code Separation - Because EH code is only exceptionally executed, it can often be moved to a separate section. It's moved when profile-guided optimizations can determine that the exceptions occur only on exceptional conditions.

  • Memory Intrinsics - Whether to expand an intrinsic or not depends on whether it's called frequently. An intrinsic can also be optimized based on the block size of moves or copies.

Next steps

Read more about these environment variables, functions, and tools you can use in profile-guided optimizations:

Environment variables for profile-guided optimizations
These variables were used to specify run-time behavior of testing scenarios. They're now deprecated and replaced by new linker options. This document shows you how to move from the environment variables to the linker options.

PgoAutoSweep
A function you can add to your app to provide fine-grained .pgc file data capture control.

pgosweep
A command-line utility that writes all profile data to the .pgc file, closes the .pgc file, and opens a new .pgc file.

pgomgr
A command-line utility that adds profile data from one or more .pgc files to the .pgd file.

How to: Merge multiple PGO profiles into a single profile
Examples of pgomgr usage.

See also

Additional MSVC build tools