Why doesn't GCC optimize a*a*a*a*a*a to (a*a*a)*(a*a*a)?

Question

I am doing some numerical optimization on a scientific application. One thing I noticed is that GCC will optimize the call pow(a,2) by compiling it into a

Answer 1

Another similar case: most compilers won't optimize a + b + c + d to (a + b) + (c + d) (this is an optimization since the second expression can be pipelined better) and evaluate it as given (i.e. as (((a + b) + c) + d)). This too is because of corner cases:

float a = 1e35, b = 1e-5, c = -1e35, d = 1e-5;
printf("%e %e\n", a + b + c + d, (a + b) + (c + d));

This outputs 1.000000e-05 0.000000e+00

Answer 2

No posters have mentioned the contraction of floating expressions yet (ISO C standard, 6.5p8 and 7.12.2). If the FP_CONTRACT pragma is set to ON, the compiler is allowed to regard an expression such as a*a*a*a*a*a as a single operation, as if evaluated exactly with a single rounding. For instance, a compiler may replace it by an internal power function that is both faster and more accurate. This is particularly interesting as the behavior is partly controlled by the programmer directly in the source code, while compiler options provided by the end user may sometimes be used incorrectly.

The default state of the FP_CONTRACT pragma is implementation-defined, so that a compiler is allowed to do such optimizations by default. Thus portable code that needs to strictly follow the IEEE 754 rules should explicitly set it to OFF.

If a compiler doesn't support this pragma, it must be conservative by avoiding any such optimization, in case the developer has chosen to set it to OFF.

GCC doesn't support this pragma, but with the default options, it assumes it to be ON; thus for targets with a hardware FMA, if one wants to prevent the transformation a*b+c to fma(a,b,c), one needs to provide an option such as -ffp-contract=off (to explicitly set the pragma to OFF) or -std=c99 (to tell GCC to conform to some C standard version, here C99, thus follow the above paragraph). In the past, the latter option was not preventing the transformation, meaning that GCC was not conforming on this point: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=37845

Answer 3

Because a 32-bit floating-point number - such as 1.024 - is not 1.024. In a computer, 1.024 is an interval: from (1.024-e) to (1.024+e), where "e" represents an error. Some people fail to realize this and also believe that * in a*a stands for multiplication of arbitrary-precision numbers without there being any errors attached to those numbers. The reason why some people fail to realize this is perhaps the math computations they exercised in elementary schools: working only with ideal numbers without errors attached, and believing that it is OK to simply ignore "e" while performing multiplication. They do not see the "e" implicit in "float a=1.2", "a*a*a" and similar C codes.

Should majority of programmers recognize (and be able to execute on) the idea that C expression a*a*a*a*a*a is not actually working with ideal numbers, the GCC compiler would then be FREE to optimize "a*a*a*a*a*a" into say "t=(a*a); t*t*t" which requires a smaller number of multiplications. But unfortunately, the GCC compiler does not know whether the programmer writing the code thinks that "a" is a number with or without an error. And so GCC will only do what the source code looks like - because that is what GCC sees with its "naked eye".

... once you know what kind of programmer you are, you can use the "-ffast-math" switch to tell GCC that "Hey, GCC, I know what I am doing!". This will allow GCC to convert a*a*a*a*a*a into a different piece of text - it looks different from a*a*a*a*a*a - but still computes a number within the error interval of a*a*a*a*a*a. This is OK, since you already know you are working with intervals, not ideal numbers.

Answer 4

Library functions like "pow" are usually carefully crafted to yield the minimum possible error (in generic case). This is usually achieved approximating functions with splines (according to Pascal's comment the most common implementation seems to be using Remez algorithm)

fundamentally the following operation:

pow(x,y);

has a inherent error of approximately the same magnitude as the error in any single multiplication or division.

While the following operation:

float a=someValue;
float b=a*a*a*a*a*a;

has a inherent error that is greater more than 5 times the error of a single multiplication or division (because you are combining 5 multiplications).

The compiler should be really carefull to the kind of optimization it is doing:

1. if optimizing pow(a,6) to a*a*a*a*a*a it may improve performance, but drastically reduce the accuracy for floating point numbers.
2. if optimizing a*a*a*a*a*a to pow(a,6) it may actually reduce the accuracy because "a" was some special value that allows multiplication without error (a power of 2 or some small integer number)
3. if optimizing pow(a,6) to (a*a*a)*(a*a*a) or (a*a)*(a*a)*(a*a) there still can be a loss of accuracy compared to pow function.

In general you know that for arbitrary floating point values "pow" has better accuracy than any function you could eventually write, but in some special cases multiple multiplications may have better accuracy and performance, it is up to the developer choosing what is more appropriate, eventually commenting the code so that noone else would "optimize" that code.

The only thing that make sense (personal opinion, and apparently a choice in GCC wichout any particular optimization or compiler flag) to optimize should be replacing "pow(a,2)" with "a*a". That would be the only sane thing a compiler vendor should do.

Answer 5

I would not have expected this case to be optimized at all. It can't be very often where an expression contains subexpressions that can be regrouped to remove entire operations. I would expect compiler writers to invest their time in areas which would be more likely to result in noticeable improvements, rather than covering a rarely encountered edge case.

I was surprised to learn from the other answers that this expression could indeed be optimized with the proper compiler switches. Either the optimization is trivial, or it is an edge case of a much more common optimization, or the compiler writers were extremely thorough.

There's nothing wrong with providing hints to the compiler as you've done here. It's a normal and expected part of the micro-optimization process to rearrange statements and expressions to see what differences they will bring.

While the compiler may be justified in considering the two expressions to deliver inconsistent results (without the proper switches), there's no need for you to be bound by that restriction. The difference will be incredibly tiny - so much so that if the difference matters to you, you should not be using standard floating point arithmetic in the first place.

Answer 6

gcc actually can do this optimization, even for floating-point numbers. For example,

double foo(double a) {
  return a*a*a*a*a*a;
}

becomes

foo(double):
    mulsd   %xmm0, %xmm0
    movapd  %xmm0, %xmm1
    mulsd   %xmm0, %xmm1
    mulsd   %xmm1, %xmm0
    ret

with -O -funsafe-math-optimizations. This reordering violates IEEE-754, though, so it requires the flag.

Signed integers, as Peter Cordes pointed out in a comment, can do this optimization without -funsafe-math-optimizations since it holds exactly when there is no overflow and if there is overflow you get undefined behavior. So you get

foo(long):
    movq    %rdi, %rax
    imulq   %rdi, %rax
    imulq   %rdi, %rax
    imulq   %rax, %rax
    ret

with just -O. For unsigned integers, it's even easier since they work mod powers of 2 and so can be reordered freely even in the face of overflow.