https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108281
Bug ID: 108281
Summary: float value range estimation missing (vs. integer)
Product: gcc
Version: 12.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: aleks at physik dot tu-berlin.de
Target Milestone: ---
"gcc -O3" and optional: "-funsafe-math-optimizations" (isnan)
GCC ignores ranges of float numbers for optimization, tested via
https://godbolt.org/
For many frequent used functions (or even all in math.h) gcc could know a range
limit and concidering it for comparisons.
#include <math.h>
int ranges(float x) {
if (x!=x) return -1; // optional NAN check
if (cos(x) < -1.0f) return -2;
if (sin(x) > 1.0f) return -3;
if (fabs(x) < 0.0f) return -4;
if (atan(x) < -2.0f) return -5; // +-PI/2
if (exp(x) < 0.0f) return -6;
if (sqrt(x) < 0.0f) return -7;
if (log(x) < 90.0f) return -8; // ln(FLT_MAX)=88.8
// ln(DBL_MAX) = 709.8
return 0; // the only valid return (beside -1)
}
int sqr2(float x) { // squares give non-negative results
return x*x < 0.0f; // == false
}
int ax2(float x) {
return fabs(x) > -1.0f; // == true
}
int cmp_sqrt(float x, float y) { // similar (!sadly very often seen!)
//x = fabs(x); y = fabs(y); // optional sign removal line
return sqrtf(x) < sqrtf(y); // == (x < y), hint: sqrt=slow
}
