Inside the Intel Compiler

Dale Schouten

Xinmin Tian

Aart Bik

Milind Girkar

Issue #106, February 2003

How did Intel's compiler beat gcc on Benchmarks? Intel's compiler developers explain IA-32 optimizations they use.

The increasing acceptance of Linux among developers and researchers has yet to be matched by a similar increase in the number of available development tools. The recently released Intel C++ and Fortran compilers for Linux aim to bridge this gap by providing application developers with highly optimizable compilers for the Intel IA-32 and Itanium processor families. These compilers provide strict ANSI support, as well as optional support for some popular extensions. This article focuses on the optimizations and features of the compiler for the Intel IA-32 processors. Throughout the rest of this article, we refer to the Intel C++ and Fortran compilers for Linux on IA-32 collectively as “the Intel compiler”.

The Intel compiler optimizes a program at all levels, from high-level loop and interprocedural optimizations to standard compiler data flow optimizations, in addition to efficient low-level optimizations, such as instruction scheduling, basic block layout and register allocation. In this article, we mainly focus on compiler optimizations unique to the Intel compiler. For completeness, however, we also include a brief overview of some of the more traditional optimizations supported by the Intel compiler.

Traditional Compiler Optimizations

Decreasing the number of instructions that are dynamically executed and replacing instructions with faster equivalents are perhaps the two most obvious ways to improve performance. Many traditional compiler optimizations fall into this category: copy and constant propagation, redundant expression elimination, dead code elimination, peephole optimizations, function inlining, tail recursion elimination and so forth.

The Intel compiler provides a rich variety of both types of optimizations. Many local optimizations are based on the static-single-assignment (SSA) form. Redundant (or partially redundant) expressions, for example, are eliminated according to Chow's algorithm (see Resource 6), where an expression is considered redundant if it is unnecessarily calculated more than once on an execution path. For instance, in the statement:

x[i] += a[i+j*n] + b[i+j*n];

the expression i+j*n is redundant and needs to be calculated only once. Partial redundancy occurs when an expression is redundant on some paths but not necessarily all paths. In the code:

if (c) {
    x = y+a*b;
} else {
    x = a;
}
z = a*b;
the expression a*b is partially redundant. If the else branch is taken, a*b is only calculated once; but if the then branch is taken, it is calculated twice. The code can be modified as follows:
t = a*b;
if (c) {
    x = y+t;
} else {
    x = a;
}
z = t;
so there is only one calculation of a*b, no matter which path is taken.

Clearly, this transformation must be used judiciously as the increase in temporary values, ideally stored in registers, can increase lifetimes and, hence, register pressure. An algorithm similar to Chow's algorithm (see Resource 9) is used to eliminate dead stores, in which a store is succeeded by another store to the same location before a fetch, and partially dead stores, which are dead along some but not necessarily all paths. Other optimizations based on the SSA form are constant propagation (see Resource 7) and the propagation of conditions. Consider the following example:

if (x>0) {
    if (y>0) {
        . . .
        if (x == 0) {
            . . .
        }
    }
}

Since x>0 holds within the outmost if, unless x is changed, we know that x != 0, and therefore the code within the inner if is dead. Although this and the previous example may seem contrived, such situations are actually quite common in the presence of address calculations, macros or inlined functions.

Powerful memory disambiguation (see Resource 8) is used by the Intel compiler to determine whether memory references might overlap. This analysis is important to enhance, for instance, register allocation and to enable the detection and exploitation of implicit parallelism in the code, as discussed in the following sections. The Intel compiler also provides extensive interprocedural optimizations, including manual and automatic function inlining, partial inlining where only the hot parts of a routine are inlined, interprocedural constant optimizations and exception-handling optimizations. With the optional “whole program” analysis, the data layout of certain data structures, such as COMMON BLOCKS in Fortran, may be modified to enhance memory accesses on various processors. For example, the data layout could be padded to provide better data alignment. In addition, in order to make decisions that are more intelligent about when and where to inline, the Intel compiler relies on two types of profiling information: static profiling and dynamic profiling. Static profiling refers to information that can be deduced or estimated at compile time. Dynamic profiling is information gathered from actual executions of a program. These two types of profiling are discussed in the next section.

Profiling Optimizations

First, we will look at static profiling. Consider the following code fragment:

g();
for (i=0; i<10; i++) {
    g();
}

Obviously, the call inside the loop executes ten times more often than the call outside the loop. In many cases, however, there is no way to make a good estimate. In the following code:

for (i=0; i<10; i++) {
    if (condition) {
        g();
    } else {
        h();
    }
}
it is difficult to say whether one condition is more likely to occur than another. If h() happened to be an exit or some other routine that was known not to return, it would be safe to assume the then branch was more likely taken and inlining g() may be worthwhile. Without such information, however, the decision of whether to inline one call or the other (or both) gets more complicated. Another option is to use dynamic profiling.

Dynamic profiling gathers information from actual executions of a program. This allows the compiler to take advantage of the way a program actually runs in order to optimize it. In a three-step process, the application is first built with profiling instrumentation embedded in it. Then the resulting application is run with a representative sample (or samples) of data, which yields a database for the compiler to use in a subsequent build of the application. Finally, the information in this database is used to guide optimizations such as code placement or grouping frequently executed basic blocks together, function or partial inlining and register allocation. Register allocation in the Intel compiler is based on graph fusion (see Resource 5), which breaks the code into regions. These regions are typically loop bodies or other cohesive units. With profile information, the regions can be selected more effectively and are based on the actual frequency of the blocks instead of syntactic guesses. This allows spills to be pushed into less frequently executed parts of the program.

Intra-Register Vectorization

Exploiting parallelism is an important way to increase application performance in modern architectures. The Intel compiler can be key in the effort to exploit potential parallelism in a program by facilitating such optimizations as automatic vectorization, automatic parallelization and support for OpenMP directives. Let's look at the automatic conversion of serial loops into a form that takes advantage of the instructions provided by the Intel MMX technology or SSE/SSE2 (Streaming-SIMD-extensions), a process we refer to as “intra-register vectorization” (see Resource 1). For example, given the function:

void vecadd(float a[], float b[], float c[], int n)
{
  int i;
  for (i = 0; i < n; i++) {
      c[i] = a[i] + b[i];
  }
}

the Intel compiler will transform the loop to allow four single-precision floating-point additions to occur simultaneously using the addps instruction. Simply put, using a pseudo-vector notation, the result would look something like this:

for (i = 0; i < n; i+=4) {
    c[i:i+3] = a[i:i+3] + b[i:i+3];
}
A scalar cleanup loop would follow to execute the remainder of the instructions if the trip count n is not exactly divisible by four. Several steps are involved in this process. First, because it is possible that no information exists about the base addresses of the arrays, runtime code must be inserted to ensure that the arrays do not overlap (dynamic dependence testing) and that the bulk of the loop runs with each vector iteration having addresses aligned along 16-byte boundaries (dynamic loop peeling for alignment). In order to vectorize efficiently, only loops of sufficient size are vectorized. If the number of iterations is too small, a simple serial loop is used instead. Besides simple loops, the vectorizer also supports loops with reductions (such as summing an array of numbers or searching for the max or min in an array, conditional constructs, saturation arithmetic and other idioms. Even the vectorization of loops with trigonometric mathematical functions is supported by means of a vector math library.

To give a taste of a realistic performance improvement that can be obtained by intra-register vectorization, we report some performance numbers for the double-precision version of the Linpack benchmark (available in both Fortran and C at www.netlib.org/benchmark). This benchmark reports the performance of a linear equation solver that uses the routines DGEFA and DGESL for the factorization and solve phase, respectively. Most of the runtime of this benchmark results from repetitively calling the Level 1 BLAS routine DAXPY for different subcolumns of the coefficient matrix during factorization. Under generic optimizations (switch -O2), this benchmark reports 1,049 MFLOPS for solving a 100×100 system on a 2.66GHz Pentium 4 processor. When intra-register vectorization for the Pentium 4 processor is enabled (switch -xW), the performance goes up to 1,292 MFLOPS, boosting the performance by about 20%.

OpenMP and Auto-Parallelization

The OpenMP standard for C/C++ and Fortran (www.openmp.org) has recently emerged as the de facto standard for shared-memory parallel programming. It allows the user to specify parallelism without getting involved in the details of iteration partitioning, data sharing, thread scheduling and synchronization. Based on these directives, the Intel compiler will transform the code to generate multithreaded code automatically. The Intel compiler supports the OpenMP C++ 2.0 and OpenMP Fortran 2.0 standard directives for explicit parallelization. Applications can use these directives to increase performance on multiprocessor systems by exploiting both task and data parallelism.

The following is an example program, illustrating the use of OpenMP directives with the Intel C++ Linux OpenMP compiler:

#define N 10000
void   ploop(void)
{
  int k, x[N], y[N], z[N];
  #pragma omp parallel for private(k) shared(x,y,z)
  for (k=0;  k<N; k++) {
    x[k] = x[k] * y[k] + workunit(z[k]);
  }
}

The for loop will be executed in parallel by a team of threads that divide the iterations in the loop body amongst themselves. Variable k is marked private—each thread will have its own copy of k—while the arrays x, y and z are shared among the threads.

The resulting multithreaded code is illustrated below. The Intel compiler generates OpenMP runtime library calls for thread creation and management, as well as synchronization (see Resources 1 and 2):

#define N 10000
void  ploop(void)
{
    int k, x[N], y[N], z[N];
    __kmpc_fork_call(loc,
                     3,
                     T-entry(_ploop_par_loop),
                     x, y, z)
    goto L1:
    T-entry _ploop_par_loop(loc, tid,
                            x[], y[], z[]) {
       lower_k = 0;
       upper_k = N;
       __kmpc_for_static_init(loc, tid, STATIC,
                              &lower_k,
                              &upper_k, ...);
       for (local_k=lower_k;  local_k<=upper_k;
            local_k++)  {
          x[local_k] = x[local_k] * y[local_k]
                       + workunit(z[local_k]);
       }
       __kmpc_for_static_fini(loc, tid);
       T-return;
    }
L1: return;
}

The multithreaded code generator inserts the thread invocation call __kmpc_fork_call with the T-entry point and data environment (for example, thread id tid) for each loop. This call into the Intel OpenMP runtime library forks a number of threads that execute the iterations of the loop in parallel.

The serial loops annotated with the OpenMP directive are converted to multithreaded code by localizing the lower- and upper-loop bounds and by privatizing the iteration variable. Finally, multithreading runtime initialization and synchronization code is generated for each T-region defined by a [T-entry, T-ret] pair. The call __kmpc_for_static_init computes the localized loop lower-bound, upper-bound and stride for each thread according to a scheduling policy. In this example, the generated code uses static scheduling. The library call __kmpc_for_static_fini informs the runtime system that the current thread has completed one loop chunk.

Rather than performing source-to-source transformations, as is done in other compilers such as OpenMP NanosCompiler and OdinMP, the Intel compiler performs these transformations internally. This allows tight integration of the OpenMP implementation with other advanced, high-level compiler optimizations for improved uniprocessor performance such as vectorization and loop transformations.

Besides the compiler support for exploiting the OpenMP directive-guided explicit parallelism, users also can try auto-parallelization by using the option -parallel. Under this option, the compiler automatically analyzes the loops in the program to detect those that have no loop-carried dependency and can be executed in parallel profitably. The auto-parallelization phase in the compiler relies on the advanced memory disambiguation techniques for its analysis, as well as the profiling information for its heuristics in deciding when to parallelize.

CPU-Dispatch

One of the unique features of the Intel compiler is CPU-Dispatch, which allows the user to target a single object for multiple IA-32 architectures by means of either manual CPU-Dispatch or Auto-CPU-Dispatch. Manual CPU-Dispatch allows the user to write multiple versions of a single function. Each function either is assigned a specific IA-32 architecture platform or is considered generic, meaning it can run on any IA-32 architecture. The Intel compiler generates code that dynamically determines on which architecture the code is running and accordingly chooses the particular version of the function that will actually execute. This runtime determination allows programmers to take advantage of architecture-specific optimizations, such as SSE and SSE2, without sacrificing flexibility, allowing execution of the same binary on architectures that do not support newer instructions.

Auto-CPU_Dispatch is similar but with the added benefit that the compiler automatically generates multiple versions of a given function. During compilation, the compiler decides which routines will gain from architecture-specific optimizations. These routines are then automatically duplicated to produce architecture-specific optimized versions, as well as generic versions. The benefit of this feature is, it does not require any rewrite by the programmer. A normal source file can take advantage of the Auto-CPU-Dispatch feature by the simple use of a command-line option. For example, given the function:

void init(float b[], double c[], int n)
{
  int i;
  for (i = 0; i < n; i++) {
      b[i] = (float)i;
  }
  for (i = 0; i < n; i++) {
      c[i] = (double)i;
  }
}

the Intel compiler can produce up to three versions of the function. A generic version of the function is generated that will run on any IA-32 processor. Another version would be tuned for the Pentium III processor by vectorizing the first loop with SSE instructions. A third version would be optimized for the Pentium 4 processor by vectorizing both loops to take advantage of SSE2 instructions.

The resulting function begins with dispatch code like this:

.L1  testl     $-512, __intel_cpu_indicator
     jne       init.J
     testl     $-128, __intel_cpu_indicator
     jne       init.H
     testl     $-1, __intel_cpu_indicator
     jne       init.A
     call      __intel_cpu_indicator_init
     jmp       .L1

Where init.A, init.H and init.J are the generic, SSE and SSE2 optimized versions, respectively.

Language Extensions

While the Intel compiler is strictly ANSI-compliant, there are options to cover many GCC extensions, such as long long int, zero-length arrays or macros with variable number of arguments. GCC-style inline assembly code is also supported. DWARF2 debugging information is provided to use with standard debuggers such as GDB. Certain Microsoft extensions are also enabled, such as __declspec attributes, along with support for Microsoft-style inline assembly code.

In addition to inline assembly code, the Intel compiler also supports MMX and SSE/SSE2 intrinsics. These allow access to the processor-specific extensions without the performance and correctness problems often caused by using inline assembly that can interfere with the analysis and transformations of the Intel compiler. By using the provided intrinsics, the programmer can take advantage of specific instructions but still receive the benefits of register allocation, scheduling and other optimizations.

Conclusions

The Intel compiler for Linux is a state-of-the-art compiler that delivers performance among the best in the industry, using sophisticated techniques to enable advanced features of Intel IA-32 architectures. More information can be found at developer.intel.com/software/products/compilers.

Acknowledgements

Thanks to Zia Ansari and David Kreitzer for their help in describing some of the technical details of the compiler. We also thank all the other members of the Intel compiler team.

Intel, Pentium, Itanium and MMX are trademarks or registered trademarks of Intel Corporation or its subsidiaries in the United States and other countries.

Resources

Dale Schouten (Dale.A.Schouten@intel.com) works at the Intel Compiler Lab. He has a PhD from the University of Illinois. In his other life, Dale is an unprofessional musician and the father of two exceptional children.

Xinmin Tian (Xinmin.Tian@intel.com) works at the Intel Compiler Lab at Intel Corp. He manages the OpenMP Parallelization group. He holds BSc, MSc and PhD degrees in Computer Science from Tsinghua University.

Aart Bik (Aart.Bik@intel.com) received his MSc degree in Computer Science from Utrecht University, The Netherlands, and his PhD degree from Leiden University, The Netherlands. He is currently working on vectorization and parallelization at the Intel Compiler Lab.

Milind Girkar (Milind.Girkar@intel.com) received a PhD degree in Computer Science from the University of Illinois at Urbana-Champaign. Currently, he manages the IA-32 Compiler Development group at the Intel Compiler Lab.