Re: Strange floating point problems on SH4 with gcc 4.1.0

[Andrew de Quincey <adq_dvb at lidskialf dot net>]

Sorry this is taking a while, I'm still catching up on where I got to two 
weeks ago. Anyway, the original error I am debugging occurs in the following 
line of nsComponentManager.cpp/nsComponentManagerImpl::Init().

       PL_DHashTableSetAlphaBounds(&mFactories,
                                    0.875,
                                    PL_DHASH_MIN_ALPHA(&mFactories, 2));

In fact its the PL_DHASH_MIN_ALPHA macro, which is defined as:

       #define PL_DHASH_MIN_ALPHA(table, k) \
    ((float)((table)->entrySize / sizeof(void *) - 1) / ((table)->entrySize / 
sizeof(void *) + (k)))

As I'm getting a floating point arithmetic exception, I immediately 
thought "oh division by 0". However I have verified that none of the values 
being used here are 0:

table->entrySize is 8
sizeof(void*) is 4
and k is 2 obviously.

The error seems to occur when the int is cast to a float... it somehow becomes 
0.

Looking at the ASM in gdb shows the error happens at:

0x297b5030 <_ZN22nsComponentManagerImpl4InitEPK18nsStaticModuleInfoj+472>:      
fdiv    fr1,fr2

Which would indicate that for some reason fr2 is unexepectedly 0. GDB refuses 
to dump out floating point registers on this platform so I cannot confirm 
directly :(


Originally, I added some debug tracing printfs before that to dump out 
variables as follows:

       int x = 20;
       float k = (float) x;

       float y = (float)(mFactories.entrySize / sizeof(void *) - 1);
       printf("%f\n", k);
       printf("%i %i %i\n", mFactories.entrySize, sizeof(nsFactoryTableEntry), 
sizeof(void*));
       printf("%f\n", y);
       printf("%i\n", x);
       printf("%f %f\n", k, (float) x);
       printf("%f\n", y / (float) x);
       fflush(stdout);

The first printout reports the value of 'k' to be 0.007812... I know people 
report rounding problems erroneously as gcc errors, but 20 rounded to 
0.007812 ain't right :)

I get a floating point error from the second printf in the __udivsi3_i4 
function again.

I know it sounds like a bug in glibc's printf for this platform, but the 
original error I am trying to fix does not use printf().