This is the mail archive of the
gcc-help@gcc.gnu.org
mailing list for the GCC project.
Extreme memory demand compiling not-too-complicated expression: gcc bug?
- From: Bob McKay <urilabob at gmail dot com>
- To: gcc-help at gcc dot gnu dot org
- Date: Wed, 11 Apr 2012 20:00:32 +0900
- Subject: Extreme memory demand compiling not-too-complicated expression: gcc bug?
Dear gcc-ers,
I need to compile fairly complex (automatically generated)
expressions using gcc. An example is attached. To allow the code to
run reasonably fast, I assumed it would be reasonable to macro-ise the
functions where possible (most operators need to be protected against
overflow, since I can't control the expressions that are generated),
and inline the rest. Inlining works fine (the code compiles in a
fraction of a second), but macro-ising doesn't. This version of the
code ran up to 14GB of (real) memory and ~15 minutes compilation time
before I killed it. While it doesn't matter much to me (the inlined
version is still a huge speed-up on the interpreter I was using till
now), it does seem a little surprising that compiling a relatively
simple expression should generate such overheads. I could understand
it, I guess, if I were using aggressive optimisations; but the same
overheads appeared even under minimum optimisation. So back to the
question: does this look like a bug to you? Should I be reporting it?
Even better, can anyone suggest a combination of switches that will
allow it to compile in a reasonable time and memory?
Thanks and Best Wishes
Bob
#include "math.h"
#define VCB_V_alk 98
#define VCB_V_ba 76
#define VCB_V_bz 75
#define VCB_V_cd 73
#define VCB_V_do 3
#define VCB_V_lgt 77
#define VCB_V_n 53
#define VCB_V_p 16
#define VCB_V_ph 2
#define VCB_V_sd 97
#define VCB_V_si 102
#define VCB_V_tb 15
#define VCB_V_tmp 115
#define INFPLUS exp(700)
#define INFMINUS -exp(700)//#define EXPOVL 650
#define PVAL(y) ((y==HUGE_VAL)?(INFPLUS):((y==-HUGE_VAL)?(-INFMINUS):(y)))
#define DIVISION_BY_ZERO_MULTIPLIER 100000 // When division by zero happens, this number is multiplied to the numerator.
#define DIVISION_BY_ZERO_EPSILON 0.000001
#define FN_PLUS(X,Y) (PVAL(X+Y))
#define FN_MINUS(X,Y) (PVAL(X-Y))
#define FN_MULTIPLY(X,Y) (PVAL(X*Y))
// Can't #define DIVIDE(X,Y) because protected - must inline
#define FN_SIN(X) (sin(X)) // range -1..1, no protection needed
#define FN_COS(X) (cos(X)) // range -1..1, no protection needed
#define FN_CARET(X) (PVAL(exp(X)))
// Can't #define LOG(X) because protected - must inline
inline double FN_DIVIDE(double x, double y) {
if (fabs(y - 0) > DIVISION_BY_ZERO_EPSILON) { // Protected division
return PVAL(x/y);
} else {
return x * DIVISION_BY_ZERO_MULTIPLIER;
}
}
inline double FN_LOG(double x) { //log (base e)
if (x > 0) { // Protected logarithm
return log(fabs(x));
} else {
return 1;
}
}
double ba_eval(double vval[]) {
double retvalue;
retvalue = FN_MINUS(FN_MULTIPLY(vval[VCB_V_ba],FN_MINUS(FN_MULTIPLY(FN_DIVIDE(vval[VCB_V_lgt],30),FN_MULTIPLY(FN_DIVIDE(1,FN_PLUS(1,FN_CARET(FN_MULTIPLY(-1.5,FN_PLUS(vval[VCB_V_si],-2.5))))),FN_MULTIPLY(FN_DIVIDE(1,FN_PLUS(1,FN_CARET(FN_MULTIPLY(-25,FN_PLUS(vval[VCB_V_p],-0.1))))),FN_MULTIPLY(FN_DIVIDE(1,FN_PLUS(1,FN_CARET(FN_MULTIPLY(-0.2,FN_PLUS(vval[VCB_V_n],-2))))),FN_MULTIPLY(0.0334,FN_MULTIPLY(FN_CARET(FN_MINUS(1,FN_DIVIDE(vval[VCB_V_lgt],30))),FN_PLUS(FN_MULTIPLY(FN_DIVIDE(vval[VCB_V_tmp],27),FN_CARET(FN_MINUS(1,FN_DIVIDE(vval[VCB_V_tmp],27)))),FN_MULTIPLY(FN_DIVIDE(vval[VCB_V_tmp],4),FN_CARET(FN_MINUS(1,FN_DIVIDE(vval[VCB_V_tmp],4))))))))))),FN_MULTIPLY(0.0284681,FN_CARET(FN_MULTIPLY(0.038,vval[VCB_V_tmp]))))),FN_MULTIPLY(vval[VCB_V_bz],FN_MULTIPLY(0.00375,FN_DIVIDE(vval[VCB_V_ba],FN_PLUS(250,vval[VCB_V_ba])))));
return retvalue;
}