I totally understand the problems associated with floating points, but I have seen a very interesting behavior that I can't explain.
float x = 1028.25478;
long int y = 102825478;
float z = y/(float)100000.0;
printf("x = %f ", x);
printf("z = %f",z);
The output is:
x = 1028.254761 z = 1028.254780
Now if floating numbers failed to represent that specific random value (1028.25478) when I assigned that to variable x. Why isn't it the same in case of variable z?
P.S. I'm using pellesC IDE to test the code (C11 compiler).
I am pretty sure that what happens here is that the latter floating point variable is elided and instead kept in a double-precision register; and then passed as is as an argument to printf. Then the compiler will believe that it is safe to pass this number at double precision after default argument promotions.
I managed to produce a similar result using GCC 7.2.0, with these switches:
-Wall -Werror -ffast-math -m32 -funsafe-math-optimizations -fexcess-precision=fast -O3
The output is
x = 1028.254761 z = 1028.254800
The number is slightly different there^.
The description for -fexcess-precision=fast says:
-fexcess-precision=style
This option allows further control over excess precision on
machines where floating-point operations occur in a format with
more precision or range than the IEEE standard and interchange
floating-point types. By default, -fexcess-precision=fast is in
effect; this means that operations may be carried out in a wider
precision than the types specified in the source if that would
result in faster code, and it is unpredictable when rounding to
the types specified in the source code takes place. When
compiling C, if -fexcess-precision=standard is specified then
excess precision follows the rules specified in ISO C99; in
particular, both casts and assignments cause values to be rounded
to their semantic types (whereas -ffloat-store only affects
assignments). This option [-fexcess-precision=standard] is enabled by default for C if a
strict conformance option such as -std=c99 is used. -ffast-math
enables -fexcess-precision=fast by default regardless of whether
a strict conformance option is used.
This behaviour isn't C11-compliant
Restricting this to IEEE754 strict floating point, the answers should be the same.
1028.25478 is actually 1028.2547607421875. That accounts for x.
In the evaluation of y / (float)100000.0;, y is converted to a float, by C's rules of argument promotion. The closest float to 102825478 is 102825480. IEEE754 requires the returning of the the best result of a division, which should be 1028.2547607421875 (the value of z): the closest number to 1028.25480.
So my answer is at odds with your observed behaviour. I put that down to your compiler not implementing floating point strictly; or perhaps not implementing IEEE754.
Code acts as if z was a double and y/(float)100000.0 is y/100000.0.
float x = 1028.25478;
long int y = 102825478;
double z = y/100000.0;
// output
x = 1028.254761 z = 1028.254780
An important consideration is FLT_EVAL_METHOD. This allows select floating point code to evaluate at higher precision.
#include <float.h>
#include <stdio.h>
printf("FLT_EVAL_METHOD %d\n", FLT_EVAL_METHOD);
Except for assignment and cast ..., the values yielded by operators with floating operands and values subject to the usual arithmetic conversions and of floating constants are evaluated to a format whose range and precision may be greater than required by the type. The use of evaluation formats is characterized by the implementation-defined value of FLT_EVAL_METHOD.
-1 indeterminable;
0 evaluate all operations and constants just to the range and precision of the
type;
1 evaluate ... type float and double to the
range and precision of the double type, evaluate long double
... to the range and precision of the long double
type;
2 evaluate all ... to the range and precision of the
long double type.
Yet this does not apply as z with float z = y/(float)100000.0; should lose all higher precision on the assignment.
I agree with #Antti Haapala that code is using a speed optimization that has less adherence to the expected rules of floating point math.
Related
I'm wondering if there are any circumstances where code like this will be incorrect due to floating point inaccuracies:
#include <math.h>
// other code ...
float f = /* random but not NAN or INF */;
int i = (int)floorf(f);
// OR
int i = (int)ceilf(f);
Are there any guarantees about these values? If I have a well-formed f (not NAN or INF) will i always be the integer that it rounds to, whichever way that is.
I can image a situation where (with a bad spec/implementation) the value you get is the value just below the true value rather than just above/equal but is actually closer. Then when you truncate it actually rounds down to the next lower value.
It doesn't seem possible to me given that integers can be exact values in ieee754 floating point but I don't know if float is guaranteed to be that standard
The C standard is sloppy in specifying floating-point behavior, so it is technically not completely specified that floorf(f) produces the correct floor of f or that ceilf(f) produces the correct ceiling of f.
Nonetheless, no C implementations I am aware of get this wrong.
If, instead of floorf(some variable), you have floorf(some expression), there are C implementations that may evaluate the expression in diverse ways that will not get the same result as if IEEE-754 arithmetic were used throughout.
If the C implementation defines __STDC_IEC_559__, it should evaluate the expressions using IEEE-754 arithmetic.
Nonetheless, int i = (int)floorf(f); is of course not guaranteed to set i to the floor of f if the floor of f is out of range of int.
I tried running this simple code on ideone.com
#include<stdio.h>
int main()
{
double a = 0.7f; // Notice: f for float
double b = 0.7;
if (a == b)
printf("Identical\n");
else
printf("Differ\n");
return 0;
}
With gcc-5.1 the output is Identical
With clang 3.7 the output is Differ
So it seems gcc ignores the f in 0.7f and treats it as a double while clang treats it as a float.
Is this a bug in one of the compilers or is this implementation dependent per standard?
Note: This is not about floating point numbers being inaccurate. The point is that gcc and clang treats this code differently.
The C standard allows floating point operations use higher precision than what is implied by the code during compilation and execution. I'm not sure if this is the exact clause in the standard but the closest I can find is §6.5 in C11:
A floating expression may be contracted, that is, evaluated as though it were a single operation, thereby omitting rounding errors implied by the source code and the expression evaluation method
Not sure if this is it, or there's a better part of the standard that specifies this. There was a huge debate about this a decade ago or two (the problem used to be much worse on i386 because of the internally 40/80 bit floating point numbers in the 8087).
The compiler is required to convert the literal into an internal representation which is at least as accurate as the literal. So gcc is permitted to store floating point literals internally as doubles. Then when it stores the literal value in 'a' it will be able to store the double. And clang is permitted to store floats as floats and doubles as doubles.
So it's implementation specific, rather than a bug.
Addendum: For what it is worth, something similar can happen with ints as well
int64_t val1 = 5000000000;
int64_t val2 = 5000000000LL;
if (val1 != val2) { printf("Different\n"); } else { printf("Same\n"); }
can print either Different or Same depending on how your compiler treats integer literals (though this is more particularly an issue with 32 bit compilers)
Closed. This question needs details or clarity. It is not currently accepting answers.
Want to improve this question? Add details and clarify the problem by editing this post.
Closed 7 years ago.
Improve this question
There are numerous reference on the subject (here or here). However I still fails to understand why the following is not considered UB and properly reported by my favorite compiler (insert clang and/or gcc) with a neat warning:
// f1, f2 and epsilon are defined as double
if ( f1 / f2 <= epsilon )
As per C99:TC3, 5.2.4.2.2 §8: we have:
Except for assignment and cast (which remove all extra range and
precision), the values of operations with floating operands and values
subject to the usual arithmetic conversions and of floating constants
are evaluated to a format whose range and precision may be greater
than required by the type. [...]
Using typical compilation f1 / f2 would be read directly from the FPU. I've tried here using gcc -m32, with gcc 5.2. So f1 / f2 is (over-here) on an 80 bits (just a guess dont have the exact spec here) floating point register. There is not type promotion here (per standard).
I've also tested clang 3.5, this compiler seems to cast the result of f1 / f2 back to a normal 64 bits floating point representation (this is an implementation defined behavior but for my question I prefer the default gcc behavior).
As per my understanding the comparison will be done in between a type for which we don't know the size (ie. format whose range and precision may be greater) and epsilon which size is exactly 64 bits.
What I really find hard to understand is equality comparison with a well known C types (eg. 64bits double) and something whose range and precision may be greater. I would have assumed that somewhere in the standard some kind of promotion would be required (eg. standard would mandates that epsilon would be promoted to a wider floating point type).
So the only legitimate syntaxes should instead be:
if ( (double)(f1 / f2) <= epsilon )
or
double res = f1 / f2;
if ( res <= epsilon )
As a side note, I would have expected the litterature to document only the operator <, in my case:
if ( f1 / f2 < epsilon )
Since it is always possible to compare floating point with different size using operator <.
So in which cases the first expression would make sense ? In other word, how could the standard defines some kind of equality operator in between two floating point representation with different size ?
EDIT: The whole confusion here, was that I assumed it was possible to compare two float of different size. Which cannot possibly happen. (thanks #DevSolar!).
<= is well-defined for all possible floating point values.
There is one exception though: the case when at least one of the arguments is uninitialised. But that's more to do with reading an uninitialised variable being UB; not the <= itself
I think you're confusing implementation-defined with undefined behavior. The C language doesn't mandate IEEE 754, so all floating point operations are essentially implementation-defined. But this is different from undefined behavior.
After a bit of chat, it became clear where the miscommunication came from.
The quoted part of the standard explicitly allows an implementation to use wider formats for floating operands in calculations. This includes, but is not limited to, using the long double format for double operands.
The standard section in question also does not call this "type promotion". It merely refers to a format being used.
So, f1 / f2 may be done in some arbitrary internal format, but without making the result any other type than double.
So when the result is compared (by either <= or the problematic ==) to epsilon, there is no promotion of epsilon (because the result of the division never got a different type), but by the same rule that allowed f1 / f2 to happen in some wider format, epsilon is allowed to be evaluated in that format as well. It is up to the implementation to do the right thing here.
The value of FLT_EVAL_METHOD might tell what exactly an implementation is doing exactly (if set to 0, 1, or 2 respectively), or it might have a negative value, which indicates "indeterminate" (-1) or "implementation-defined", which means "look it up in your compiler manual".
This gives an implementation "wiggle room" to do any kind of funny things with floating operands, as long as at least the range / precision of the actual type is preserved. (Some older FPUs had "wobbly" precisions, depending on the kind of floating operation performed. The quoted part of the standard caters for exactly that.)
In no case may any of this lead to undefined behaviour. Implementation-defined, yes. Undefined, no.
The only case where you would get undefined behavior is when a large floating point variable gets demoted to a smaller one which cannot represent the contents. I don't quite see how that applies in this case.
The text you quote is concerned about whether or not floats may be evaluated as doubles etc, as indicated by the text you unfortunately didn't include in the quote:
The use of evaluation formats is characterized by the
implementation-defined value of FLT_EVAL_METHOD:
-1 indeterminable;
0 evaluate all operations and constants just to the range and precision of the type;
1 evaluate operations and constants of type float and double to the range and precision of the double type, evaluate long double operations and constants to the range and precision of the long double type;
2 evaluate all operations and constants to the range and precision of the long double type.
However, I don't believe this macro overwrites the behavior of the usual arithmetic conversions. The usual arithmetic conversions guarantee that you can never compare two float variables of different size. So I don't see how you could run into undefined behavior here. The only possible issue you would have is performance.
In theory, in case FLT_EVAL_METHOD == 2 then your operands could indeed get evaluated as type long double. But please note that if the compiler allows such implicit promotions to larger types, there will be a reason for it.
According to the text you cited, explicit casting will counter this compiler behavior.
In which case the code if ( (double)(f1 / f2) <= epsilon ) is nonsense. By the time you cast the result of f1 / f2 to double, the calculation is already done and have been carried out on long double. The calculation of the result <= epsilon will however be carried out on double since you forced this with the cast.
To avoid long double entirely, you would have to write the code as:
if ( (double)((double)f1 / (double)f2) <= epsilon )
or to increase readability, preferably:
double div = (double)f1 / (double)f2;
if( (double)div <= (double)epsilon )
But again, code like this does only make sense if you know that there will be implicit promotions, which you wish to avoid to increase performance. In practice, I doubt you'll ever run into that situation, as the compiler is most likely far more capable than the programmer to make such decisions.
Ignoring why I would want to do this, the 754 IEEE fp standard doesn't define the behavior for the following:
float h = NAN;
printf("%x %d\n", (int)h, (int)h);
Gives: 80000000 -2147483648
Basically, regardless of what value of NAN I give, it outputs 80000000 (hex) or -2147483648 (dec). Is there a reason for this and/or is this correct behavior? If so, how come?
The way I'm giving it different values of NaN are here:
How can I manually set the bit value of a float that equates to NaN?
So basically, are there cases where the payload of the NaN affects the output of the cast?
Thanks!
The result of a cast of a floating point number to an integer is undefined/unspecified for values not in the range of the integer variable (±1 for truncation).
Clause 6.3.1.4:
When a finite value of real floating type is converted to an integer type other than _Bool, the fractional part is discarded (i.e., the value is truncated toward zero). If the value of the integral part cannot be represented by the integer type, the behavior is undefined.
If the implementation defines __STDC_IEC_559__, then for conversions from a floating-point type to an integer type other than _BOOL:
if the floating value is infinite or NaN or if the integral part of the floating value exceeds the range of the integer type, then the "invalid" floating-
point exception is raised and the resulting value is unspecified.
(Annex F [normative], point 4.)
If the implementation doesn't define __STDC_IEC_559__, then all bets are off.
There is a reason for this behavior, but it is not something you should usually rely on.
As you note, IEEE-754 does not specify what happens when you convert a floating-point NaN to an integer, except that it should raise an invalid operation exception, which your compiler probably ignores. The C standard says the behavior is undefined, which means not only do you not know what integer result you will get, you do not know what your program will do at all; the standard allows the program to abort or get crazy results or do anything. You probably executed this program on an Intel processor, and your compiler probably did the conversion using one of the built-in instructions. Intel specifies instruction behavior very carefully, and the behavior for converting a floating-point NaN to a 32-bit integer is to return 0x80000000, regardless of the payload of the NaN, which is what you observed.
Because Intel specifies the instruction behavior, you can rely on it if you know the instruction used. However, since the compiler does not provide such guarantees to you, you cannot rely on this instruction being used.
First, a NAN is everything not considered a float number according to the IEEE standard.
So it can be several things. In the compiler I work with there is NAN and -NAN, so it's not about only one value.
Second, every compiler has its isnan set of functions to test for this case, so the programmer doesn't have to deal with the bits himself. To summarize, I don't think peeking at the value makes any difference. You might peek the value to see its IEEE construction, like sign, mantissa and exponent, but, again, each compiler gives its own functions (or better say, library) to deal with it.
I do have more to say about your testing, however.
float h = NAN;
printf("%x %d\n", (int)h, (int)h);
The casting you did trucates the float for converting it to an int. If you want to get the
integer represented by the float, do the following
printf("%x %d\n", *(int *)&h, *(int *)&h);
That is, you take the address of the float, then refer to it as a pointer to int, and eventually take the int value. This way the bit representation is preserved.
I don't think most compilers care if, for instance, you don't append the f to a variable of float type. But just because I want to be as explicit and accurate as I can, I want to express the correct type.
What is the type of the result of two literal operands of different types, or does it depend on the circumstances? e.g.:
int i=1.0f/1;
float f=1.0f/1;
The compiler wouldn't complain in both these instances, is it because of its tolerant view of literal types or because the type of the result of the operation is always converted according to the context?
First, compilers do care about the f suffix. 1.0 is a double. 1.0f is a float. For instance:
printf("%.15f %.15f\n", 0.1f, 0.1);
produces 0.100000001490116 0.100000000000000 (note that the second number is not the real 0.1 any more than the first, it is just closer to it).
This matters not only to determine what numbers can and cannot be represented but also to determine the type of the operations this constant is involved in. For instance:
printf("%.15f %.15f\n", 1.0f/10.0f, 1.0/10.0);
produces the same output as before. What is important to notice here is that 1 and 10 are both representable exactly as float as well as as double. What we are seeing is not rounding taking place at the literal level, but the type of the operation being decided from the type of the operands.
Compilers are not being lenient about your examples. They are strictly applying rules that you can find in the relevant standard, especially sections 6.3.1.4 and 6.3.1.5.
The first one divides a float by an int, which always results in a float (as does dividing an int by a float). This float result is then converted into an int to be stored in i.
The second one takes the float that resulted from the operation, and assigns it to a float variable, preserving the floatness.
There are two separate issues here: How literals with or without suffixes are interpreted by the compiler, and the result of operations performed on mixed numeric types.
The type of literals is important in other contexts, for example (assuming f is a float):
f = f + 1.0; // Cast `f` to double, do a double-precision add, and case the result back.
f = f + 1.0f; // Single-precision addition.
Note that some compilers provide a relaxed mode, where both of them will generate the same code.