The article describes the main steps which should be performed to correctly port 32-bit Windows applications onto 64-bit Windows systems. Although the article is meant for developers using C/C++ in Visual Studio 2005/2008 environment, it will be also useful for developers who plan to port their applications onto 64-bit systems.
The article describes the main problems facing developers who plan to port 32-bit programs onto 64-bit systems. Of course the list of issues considered is not complete, but we hope that we'll offer a more detailed version of this article in future. The author would be glad to receive responses, comments, and questions which will help increase the informational value of this article.
Within the framework of a computer architecture by the term "64-bit", 64-bit integers and other 64-bit-sized data types are understood. By "64-bit" systems 64-bit microprocessor architectures (for example, EM64T, IA-64) or 64-bit operation system (for example, Windows XP Professional x64 Edition) can be understood [1].
AMD64 (or x86-64, Intel 64, EM64T, x64) is a 64-bit microprocessor architecture, and a corresponding set of instructions developed by AMD company [2]. This set of instructions was licensed by Intel company under the name of EM64T (Intel64). AMD64 architecture is an extension of x86 architecture with full backward compatibility. The architecture became widespread as a basis for personal computers and workstations.
IA-64 is a 64-bit microprocessor architecture developed together by Intel and Hewlett Packard companies. It is implemented in microprocessors Itanium and Itanium 2 [3]. The architecture is used mainly in multi-processor servers and cluster systems.
AMD64 and IA-64 are two different 64-bit architectures, which are incompatible with each other. This is why developers have to decide at once if they need to support both of the architectures, or only one of them. In most cases, if you don't develop highly tailored software for cluster systems, or don't implement your own high-performance DBMS, most likely you will have to implement support for only the AMD64 architecture, which is much more popular than IA-64. It especially concerns software for the PC market, which is nearly 100% occupied by AMD64 architecture.
Further in the article we'll speak only about AMD64 (EM64T, x64) architecture, as nowadays it is the most topical for application software developers.
Speaking about different architectures, we should mention the notion "Data model". By a data model we understand correlations between type sizes accepted within the framework of the development environment. There can be several development tools sticking to different data types for one operation system. But usually only one model dominates which corresponds to the hardware and software environment most. Such an example is 64-bit Windows, whose original data model is LLP64. But for compatibility purposes 64-bit Windows supports execution of 32-bit programs which operate in ILP32LL data model mode. Table 1 gives information about the basic data models.
Table 1. Data models.
The data model being used influences the process of developing 64-bit applications, as you need to keep in mind sizes of the data being used in programs' code [4].
You should begin mastering 64-bit systems with the question: "Do I really need to rebuild my project for a 64-bit system?" You give an answer to this question only after you have thought it over, carefully. On the one hand, you can lag behind your rivals if you don't offer 64-bit solutions. On the other hand, you might waste your time developing a 64-bit application which won't provide any competitive advantages.
Let's list the basic factors that will help you make up your mind.
You shouldn't create the 64-bit version of an application with a short life-cycle. Thanks to the WOW64 subsystem, old 32-bit applications operate rather well on 64-bit Windows systems, and this is why there is no sense in making a program 64-bit, for it won't be supported in 2 years [5]. Moreover, practice shows that porting onto 64-bit Windows versions has been delayed, and perhaps most of your users will use only the 32-bit version of your program solution in the short-term.
If you plan long-term development and support of a program product, you should begin to work over the 64-bit version of your solution. You can do this without haste, but keep in mind that the longer you don't have a complete 64-bit version, the more difficulties you will face in supporting this application installed on 64-bit Windows versions.
Recompilation of a program for a 64-bit system will allow it to use large sizes of main memory, and will also speed up its operation by 5-15%. Increase in 5-10% will be gained due to using the 64-bit processor's architectural abilities, for example a larger number of registers. The rest of the speed increase of 1-5% is explained by the absence of the WOW64 layer, which translates API calls between 32-bit applications and a 64-bit operation system.
If your program doesn't operate with large data sizes (more than 2GB), and the speed of its operation is not crucial, porting onto a 64-bit system will not be so urgent in the near future.
By the way, even simple 32-bit applications can gain advantages by being launched in a 64-bit environment. Perhaps you know that a program built with /LARGEADDRESSAWARE:YES key can allocate up to 3 GB of memory, if the 32-bit Windows is launched with /3gb key. This very 32-bit program launched on a 64-bit system can allocate nearly 4 GB of memory (in practice about 3.5 GB).
If you develop libraries, components, or other elements with the help of third-party developers who create their own software, you should act quickly while creating the 64-bit version of your product. Otherwise, your clients interested in release of 64-bit versions, will have to search for alternative solutions. For example, some developers of software-hardware security responded slowly in releasing 64-bit programs, and that made some clients search for other tools to protect their programs.
An additional advantage of releasing the 64-bit version of a library, is that you can sell it as a separate product. Thus, your clients wishing to create both 32-bit and 64-bit applications will have to buy 2 different licenses. For example, this policy is used by Spatial Corporation when selling Spatial ACIS library.
Before you plan your work on creation of the 64-bit version of your product, find out if there are 64-bit versions of libraries and components used in it. Besides this, learn about the pricing policy concerning the 64-bit version of a library. If there is no support provided, search for alternative solutions supporting 64-bit systems beforehand.
If your solutions still use 16-bit units, it is high time you got rid of them. 16-bit applications in 64-bit Windows versions are not supported.
We should explain one thing here concerning use of 16-bit installers. They are still used to install some 32-bit applications. There is a special mechanism which replaces some of the most popular 16-bit installers with their newer versions. It can lead to the false idea that 16-bit programs still operate in the 64-bit environment. Remember: it is not so.
Don't forget that using a large size of Assembler code can significantly increase the cost of creating the 64-bit version of an application.
Having thought all the listed factors over, and weighed all the pros and cons, decide if you need to port your project onto 64-bit systems. If the answer is yes, we can go further.
If you have decided to develop the 64-bit version of your product, and are ready to spend time on it, it is still not enough to guarantee success. The point is that you must possess the entire necessary toolkit, and here you may face some difficulties.
Absence of a 64-bit compiler can be the simplest but the most insuperable problem. The article is being written in 2009, but there is still no 64-bit C++ Builder compiler by Codegear. Its release is expected by the end of this year. It is impossible to avoid this problem, if only to rewrite the whole project using for example, Visual Studio. But if everything is clear about the absence of a 64-bit compiler, other similar problems can appear to be less transparent, and occur only at the stage of porting the project onto a new architecture. That's why we would like to advise you to find out beforehand if there are all the necessary components you will need, to implement the 64-bit version of your product. You may face unpleasant surprises.
Of course it is impossible to list everything you may need for a project here, but I will continue the list which will help you to orientate yourself, and perhaps to remember other things necessary to implement your 64-bit project:
There is hardly more to say about the importance of having a 64-bit compiler. It simply must be.
If you are planning to develop 64-bit applications using the latest (at the time the article is written) Visual Studio 2008 version, the following Table 2 will help you understand which of the Visual Studio editions you need.
Table 2. Abilities of different editions of Visual Studio 2008.
Of course you can use virtual machines for launching 64-bit applications on 32-bit computers, but it is too inconvenient, and won't provide the necessary level of tests. It is desirable that the machines have not less than 4-8 GB of main memory.
If libraries are presented in source codes, there must be a 64-bit configuration of the project. It can be a thankless and difficult task to update the library for a 64-bit system on your own, and the result can be unreliable and contain errors. Besides, you can violate license agreements by these actions. If you use libraries in the form of binary units, you should also find out if there are 64-bit units. You cannot use 32-bit DLL inside a 64-bit application. You can create a special tie through COM, but it will be a separate large, and difficult task [6]. Also keep in mind that you may need to spend some extra money to purchase the 64-bit version of the library.
Visual C++ doesn't support a 64-bit inline assembler. You must either use an external 64-bit assembler (for example, MASM), or possess an implementation with the same functionality in C/C++ [7].
It means considerable remaking of the testing methodology, update of unit-tests, and using new tools. We will speak about it in more detail further on, but don't forget to take it into account at the stage of evaluating time costs on migration of an application on a new system [8].
If you are developing resource-intensive applications using a large amount of main memory, you need to provide replenishment of the testing input data base. During load testing of 64-bit applications, it is desirable to excess the limits of 4 GB of the used memory. Many errors can occur only in these conditions.
The security system being used must provide full support of 64-bit systems. For example, Aladdin Company has released 64-bit drivers for support of hardware Hasp keys rather quickly. But for a long time there has been no system of automatic protection of 64-bit binary files (Hasp Envelop program). Thus, the security mechanism had to be implemented manually inside the program code, and that was one more difficult task demanding professionalism and time. Don't forget about such things relating to security, system updates, etc.
You need a new installer able to fully install 64-bit applications. We would like to warn you about one very typical mistake. It is creation of 64-bit installers for installing 32/64-bit program products. Preparing the 64-bit version of an application, developers often want to make "64-bit mode" in it absolute and create a 64-bit installer forgetting that those who use a 32-bit operation system won't simply be able to launch such an installation package. Pay attention that it is not the 32-bit application included into the distribution kit together with the 64-bit one, but the installer itself. For if the distribution kit is a 64-bit application, of course it won't operate on a 32-bit operation system. What is the most unpleasant is that a user won't be able to guess why it is happening. He will simply see an installation package which cannot be launched.
Creation of the 64-bit configuration of a project in Visual Studio 2005/2008 looks rather simple. Difficulties will begin at the stage of building a new configuration, and searching for errors in it. To create the 64-bit configuration itself, you need to perform the following 4 steps:
Launch the configuration manager, as shown in Figure 1:
Figure 1. Launch of the configuration manager.
In the configuration manager, choose support of the new platform (Figure 2):
Figure 2. Creation of a new configuration.
Choose the 64-bit platform (x64), and as a basis - settings from the 32-bit version (Figure 3). Those settings which influence the building mode will be corrected by Visual Studio automatically.
Figure 3. Choose x64 as a platform and use Win32 configuration as a basis.
Addition of a new configuration is complete, and now you can choose the 64-bit configuration version and begin compiling a 64-bit application. Choosing the 64-bit configuration for building is shown in Figure 4.
Figure 4. Now both 32-bit and 64-bit configurations are available.
If you are lucky you won't need to additionally set a 64-bit project. But it depends greatly on the project, its complexity, and the number of libraries used. The only thing you should change at once is the stack's size. If the stack's size in your project is set by default, which is 1 MB, you should define it as 2 MB for the 64-bit version. It is not necessary, but it is better to insure yourself beforehand. If you use a size different from that by default, there is sense in increasing it twice for the 64-bit version. To do this, find and change Stack Reserve Size and Stack Commit Size parameters in the project's settings.
Here we should tell you about typical problems occurring at the stage of compiling the 64-bit configuration, discuss what problems occur in third-party libraries, tell you that in the code relating to WinAPI functions the compiler won't permit placing of a pointer into LONG type, and you will have to update your code, and use LONG_PTG type. And there is a lot more to say. Unfortunately, there are so many problems, and errors are vary so much, that we cannot describe them all in one article, or even one even book. You will have to look through all the errors the compiler shows you, and all the new warnings which weren't there before by yourself and in each particular case, find out how to update the code.
Let's describe here only types which may be of interest for developers when porting applications. These types are shown in Table 3. Most recompilation errors will relate to using these very types.
|
Type |
Type's size on x86 / x64 platform |
Note |
|---|---|---|
|
int |
32 / 32 |
Basic type. On 64-bit systems remains 32-bit. |
|
long |
32 / 32 |
Basic type. On 64-bit Windows systems remains 32-bit. Keep in mind that in 64-bit Linux systems this type was extended to 64-bit. Don't forget about it if you develop code which should be compiled for Windows and Linux systems. |
|
size_t |
32 / 64 |
Basic unsigned type. The type's size is chosen in such a way that you could write the maximum size of a theoretically possible array into it. You can safely put a pointer into size_t type (except for pointers to class functions, but this is a special case). |
|
ptrdiff_t |
32 / 64 |
Similar to size_t type but this is a signed type. The result of the expression where one pointer is subtracted from the other (ptr1-ptr2) will have ptrdiff_t type. |
|
Pointer |
32 / 64 |
The size of the pointer directly depends on the platform's size. Be careful while converting pointers to other types. |
|
__int64 |
64 / 64 |
Signed 64-bit type. |
|
DWORD |
32 / 32 |
32-bit unsigned type. In WinDef.h is defined as: typedef unsigned long DWORD; |
|
DWORDLONG |
64 / 64 |
64-bit unsigned type. In WinNT.h is defined as: typedef ULONGLONG DWORDLONG; |
|
DWORD_PTR |
32 / 64 |
Unsigned type in which a pointer can be placed. In BaseTsd.h is defined as: typedef ULONG_PTR DWORD_PTR; |
|
DWORD32 |
32 / 32 |
32-bit unsigned type. In BaseTsd.h is defined as: typedef unsigned int DWORD32; |
|
DWORD64 |
64 / 64 |
64-bit unsigned type. In BaseTsd.h is defined as: typedef unsigned __int64 DWORD64; |
|
HALF_PTR |
16 / 32 |
A half of a pointer. In Basetsd.h is defined as: #ifdef _WIN64 typedef int HALF_PTR; #else typedef short HALF_PTR; #endif |
|
INT_PTR |
32 / 64 |
Signed type in which a pointer can be placed. In BaseTsd.h is defined as: #if defined(_WIN64) typedef __int64 INT_PTR; #else typedef int INT_PTR; #endif |
|
LONG |
32 / 32 |
Signed type which remained 32-bit. That's why in many cases LONG_PTR now should be used. In WinNT.h is defined as: typedef long LONG; |
|
LONG_PTR |
32 / 64 |
Signed type in which a pointer can be placed. In BaseTsd.h is defined as: #if defined(_WIN64) typedef __int64 LONG_PTR; #else typedef long LONG_PTR; #endif |
|
LPARAM |
32 / 64 |
Parameter for sending messages. In WinNT.h is defined as: typedef LONG_PTR LPARAM; |
|
SIZE_T |
32 / 64 |
Analog of size_t type. In BaseTsd.h is defined as: typedef ULONG_PTR SIZE_T; |
|
SSIZE_T |
32 / 64 |
Analog of ptrdiff_t type. In BaseTsd.h is defined as: typedef LONG_PTR SSIZE_T; |
|
ULONG_PTR |
32 / 64 |
Unsigned type in which a pointer can be placed. In BaseTsd.h is defined as: #if defined(_WIN64) typedef unsigned __int64 ULONG_PTR; #else typedef unsigned long ULONG_PTR; #endif |
|
WORD |
16 / 16 |
Unsigned 16-bit type. In WinDef.h is defined as: typedef unsigned short WORD; |
|
WPARAM |
32 / 64 |
Parameter for sending messages. In WinDef.h is defined as: typedef UINT_PTR WPARAM; |
Table N3. Types to be noted while porting 32-bit programs on 64-bit Windows systems.
If you think that after correcting all the compilation errors you will get a long-expected 64-bit application we have to disappoint you. The most difficult part is still ahead. At the compilation stage, you will correct the most explicit errors which the compiler has managed to detect, and which mostly relate to impossibility of implicit type conversion. But this is only a small part of the problem. Most errors are hidden. From the viewpoint of the abstract C++ language, these errors look safe and are disguised by explicit type conversions. The number of such errors is much larger than the number of errors detected at the compilation stage.
You shouldn't set your hopes on the /Wp64 key. This key is often presented as a wonderful means of searching for 64-bit errors. In reality the /Wp64 key simply allows you to get some warning messages concerning incorrectness of some code sections in 64-bit mode, while compiling 32-bit code. While compiling 64-bit code, these warnings will be shown anyway. And that's why the /Wp64 key is ignored when compiling a 64-bit application. And surely this key won't help in the search of hidden errors [9].
Let's consider several examples of hidden errors.
The simplest (but certainly not the easiest to detect) error class relates to explicit type conversions, when significant bits are cut. A popular example is conversion of pointers to 32-bit types when transferring them into functions such as SendMessage:
MyObj* pObj = ...
::SendMessage(hwnd, msg, (WORD)x, (DWORD)pObj);Here, the explicit type conversion is used to turn a pointer into a numeric type. For a 32-bit architecture this example is correct as the last parameter of SendMessage function has LPARAM type, which coincides with DWORD on a 32-bit architecture. For a 64-bit architecture, DWORD is incorrect and must be replaced with LPARAM. LPARAM type has sizes of 32 or 64 bits, depending on the architecture.
This is a simple case, but type conversion often looks more complicated and it is impossible to detect it using the compiler's warnings, or search through the program text. Explicit type conversions suppress the compiler's diagnosis, as they are intended for this very purpose - to tell the compiler that the type conversion is correct, and the programmer is responsible for the code's safety. Explicit search won't help either. Types can have non-standard names (defined by the programmer through typedef), and the number of methods to perform explicit type conversion is also large. To safely diagnose such errors you must use a special toolkit, such as Viva64 or PC-Lint analyzers.
The next example relates to implicit type conversion, when significant bits are also lost. fread function's code performs reading from the file, but it is incorrect when trying to read more than 2 GB on a 64-bit system.
size_t __fread(void * __restrict buf, size_t size,
size_t count, FILE * __restrict fp);
size_t
fread(void * __restrict buf, size_t size, size_t count,
FILE * __restrict fp)
{
int ret;
FLOCKFILE(fp);
ret = __fread(buf, size, count, fp);
FUNLOCKFILE(fp);
return (ret);
}__fread function returns size_t type, but int type is used to store the number of the bytes read. As a result, at large sizes of read data the function can return a false number of bytes.
You can say that it is an illiterate code for beginners, that the compiler will announce this type conversion, and that this code is actually easy to find and to correct. This is in theory. In practice everything may be quite different in cases of large projects. This example is taken from FreeBSD source code. The error was corrected only in December 2008! Note that the first (experimental) 64-bit version of FreeBSD was released in June 2003.
It is easy to make an error in the code while working with separate bits. The following error type relates to shift operations. Here is an example:
ptrdiff_t SetBitN(ptrdiff_t value, unsigned bitNum) {
ptrdiff_t mask = 1 << bitNum;
return value | mask;
}This code works well on a 32-bit architecture, and allows you to set bits with numbers 0 to 31 to unity. After porting the program onto a 64-bit platform, you will need to set bits 0 to 63. But this code will never set bits 32-63. Pay attention that "1" has int type, and when a shift at 32 positions occurs, an overflow will take place as shown in Figure 5. Whether we will get 0 (Figure 5-B) or 1 (Figure 5-C), as a result, depends on the compiler's implementation.
Figure 5. A - Correct setting of the 32nd bit in 32-bit code; B,C - error of setting of the 32nd bit on a 64-bit system (two ways of behavior)
To correct the code we need to make "1" constant of the same type as mask variable:
ptrdiff_t mask = ptrdiff_t(1) << bitNum;Also pay attention to the fact that the incorrect code leads to one more error. When setting 31 bits on a 64-bit system, the result of the function will be the value 0xffffffff80000000 (see Figure 6). The result of 1 << 31 expression is the negative number -2147483648. In a 64-bit integer variable, this number is presented as 0xffffffff80000000.
Figure 6. Error of setting of the 31th bit on a 64-bit system
Magic constants, i.e. numbers with the help of which the size of this or that type is defined, can cause a lot of troubles. The proper decision would be to use sizeof() operators for these purposes, but in a large program, an old code section can still be hidden where, as programmers believe, the pointer's size is 4 bytes and in size_t it is always 32 bits. Usually such errors look as follows:
size_t ArraySize = N * 4;
size_t *Array = (size_t *)malloc(ArraySize);Figure 4 shows the basic numbers with which you should be cautious working with, while migrating onto a 64-bit platform.
Table 4. Basic magic values which are dangerous while porting applications from a 32-bit platform onto 64-bit.
In programs processing large data sizes, errors relating to indexing large arrays or eternal loops may occur. The following example contains 2 errors:
const size_t size = ...;
char *array = ...;
char *end = array + size;
for (unsigned i = 0; i != size; ++i)
{
const int one = 1;
end[-i - one] = 0;
}The first error here, is that if the size of the data being processed excesses 4 GB (0xFFFFFFFF), an eternal loop may occur as 'i' variable has 'unsigned' type, and will never reach 0xFFFFFFFF value. I write deliberately that it can occur but not necessarily. It depends on what code the compiler will build. For example, in debug mode the eternal loop will be present, and in release-code there will be no loop as the compiler will decide to optimize the code using a 64-bit register for the counter, and the loop will be correct. All this adds much confusion, and the code which worked yesterday can fail to work today.
The second error relates to parsing the array from beginning to end for what negative indices' values are used. This code will operate well in 32-bit mode, but when executed on a 64-bit computer, access outside the array's limits will occur at the first iteration of the loop, and there will be a program crash. Let's study the reason of such a behavior.
According to C++ rules "-i - one" expression on a 32-bit system will be calculated as follows: (at the first step i = 0):
"-i" expression has unsigned type and has 0x00000000u value.
'one' variable will be extended from 'int' type to unsigned type, and will equal 0x00000001u. Note: int type is extended (according to C++ standard) up to 'unsigned' type if it participates in an operation where the second argument has unsigned type.
A subtraction operation takes place in which two values of unsigned type participate, and the result of the operation equals 0x00000000u - 0x00000001u = 0xFFFFFFFFu. Note that the result will have unsigned type.
On a 32-bit system access to the array by the index 0xFFFFFFFFu is the same as using -1 index. That is end[0xFFFFFFFFu], is an analog of end[-1]. As a result the array's items will be processed correctly.
In a 64-bit system, the situation will be quite different concerning the last point. Unsigned type will be extended to signed ptfdiff_t type, and the array's index will equal 0x00000000FFFFFFFFi64. As a result an overflow will occur.
To correct the code you should use ptrdiff_t and size_t types.
There are errors which are nobody's fault, but they are still errors. Imagine that long, long ago in a faraway galaxy (in Visual Studio 6.0), a project was developed which contained CSampleApp class - a successor of CWinApp. In the basic class there is a virtual function WinHelp. The successor overlaps this function, and performs all the necessary actions. This process is shown in Figure 7.
Figure 7. Efficient correct code created in Visual Studio 6.0
After that the project is ported on Visual Studio 2005, where the prototype of WinHelp function has changed, but nobody notices it because in 32-bit mode DWORD and DWORD_PTR types coincide, and the program continues operating correctly (Figure 8).
Figure 8. Incorrect but efficient 32-bit code
The error is waiting to reveal itself on a 64-bit system, where the types DWORD and DWORD_PTR have different sizes (see Figure 9). So it turns out that in the 64-bit mode, the classes contain two DIFFERENT WinHelp functions, which is surely incorrect. Keep in mind that such traps may hide not only in MFC, where some of the functions now have other argument types, but also in the code of your applications, and third-party libraries.
Figure 9. The error occurs in 64-bit code
There are a lot of examples of such 64-bit errors. Those who are interested in this topic, and who would like to know more about these errors see the article "20 issues of porting C++ code on the 64-bit platform" [10].
As you see the stage of searching hidden errors is a nontrivial task, and besides, many of them will occur irregularly and only at large data inputs. Static code analyzers are good for diagnosing such errors, as they can check the whole code of an application independently from the input data, and the frequency of its sections execution in real conditions. There is sense in using static analysis both at the stage of porting an application onto 64-bit platforms, in order to find most errors at the very beginning, and in further development of 64-bit solutions. Static analysis will warn, and teach, a programmer to better understand the peculiarities of errors relating to a 64-bit architecture, and to write more efficient code. The author of the article is a developer of one of such specialized code analyzer, named Viva64 [11]. To learn more about the tool and to download a demo version visit the site of OOO "Program Verification Systems" company.
For justice' sake we should say that Gimpel PC-Lint and Parasoft C++test code analyzers have sets of rules for diagnosing 64-bit errors. But, firstly, these are general-purpose analyzers, and the rules of diagnosing 64-bit errors are incomplete. Secondly, they are intended mostly for the LP64 data model used in the family of Linux operation systems, and so are not so useful for Windows programs where LLP64 data model is used [12].
The step of searching for errors in program code described in the previous section is necessary, but insufficient. None of the methods, including static code analysis, can guarantee detection of all errors, and the best result can be achieved only when combining different methods.
If your 64-bit program processes a larger data size than the 32-bit version, you need to extend tests to include processing data with a size of more than 4 GB. This is the border beyond which many 64-bit errors begin to occur. Such tests may take much more time, and you must be prepared for it. Usually tests are written in such a way that each test could process a small number of items, and thus make it possible to perform all the internal unit-tests in several minutes, while automatic tests (for example, using AutomatedQA TestComplete) could be performed in several hours. It is nearly certain that the sorting function sorting 100 items will behave correctly at 100000 items on a 32-bit system. But the same function can fail on a 64-bit system while trying to process 5 billion items. The speed of executing a unit-test can fall in million times. Don't forget about the cost of adapting tests while mastering 64-bit systems. A good solution is to divide unit-tests into quick (working with small memory sizes), and slow ones processing gigabytes, and executed for example, at nighttime. Automated testing of resource-intensive 64-bit programs can be organized on the basis of distributed calculations.
There is one more unpleasant thing. You will hardly succeed in using tools like BoundsChecker to search for errors in resource-intensive 64-bit programs consuming large memory size. The reason is a great slowdown of the programs being tested which makes this approach very inconvenient. In the mode of diagnosing all the errors relating to memory operation, the Parallel Inspector tool included in Intel Parallel Studio, will slow down execution of an application by 100 times, on average (Figure 10). It is very likely that you will have to leave the algorithm being tested for the night to see the results only the next day, while normally this algorithm operates in just 10 minutes. And still, I'm sure that Parallel Inspector is one of the most useful and convenient tools when working in the mode of searching memory-operation errors. You just need to be ready to change the practice of error diagnosing, and keep it in mind when planning to master 64-bit systems.
Figure 10. The settings window of the Parallel Inspector program before launching an application.
And the last thing. Don't forget to add tests checking compatibility of data formats between the 32-bit and 64-bit versions. Data compatibility is often violated during migration, because of writing of such types as size_t or long (in Linux systems) into files.
