Please do not reply to this email. Use the web interface provided at: http://bugs.ecos.sourceware.org/show_bug.cgi?id=1001490
--- Comment #6 from Bernd Edlinger <[email protected]> 2012-08-09 08:49:57 BST --- (In reply to comment #5) > It's okay for ("%.18f\n", 3.14e-11) > 0.000000000031400000 but not quite OK for these: ("%.18f\n", 3.1415926E-11) eCos: "0.000000000031400000" glib: "0.000000000031415926" ("%.18f\n", DBL_MAX*2) eCos: "inf000" glib: "inf" > But the padding/zeroing will be wrong for %e, %E, when requested prec > > MAXPREC. but also for %g: this does work like %e, when the value is >=10^prec or <=10^-4 > Well, it looks like my fix (Suzuki did talk about the same point which I > found in GDB), but my workaround was > if (prec > MAXFRACT) { > if ((ch == 'f' && ch == 'F') || (flags&ALT)) { > fpprec = prec - MAXFRACT; > prec = MAXFRACT; > } > } else if (prec == -1) ok, but with this patch there will be a crash in printf("%.999e", x) limiting prec MAXFRACT helps to avoid the buffer overrun in "cvt" however with DBL_MAX the buffer size BUF 2 characters too small as I said. I tried to solve it this way: diff -Nur ecos-cvs-120723/packages/language/c/libc/stdio/current/src/output/vfnprintf.cxx ecos/packages/language/c/libc/stdio/current/src/output/vfnprintf.cxx --- ecos-cvs-120723/packages/language/c/libc/stdio/current/src/output/vfnprintf.cxx 2009-08-20 18:09:18.000000000 +0200 +++ ecos/packages/language/c/libc/stdio/current/src/output/vfnprintf.cxx 2012-08-07 10:16:48.809576300 +0200 @@ -107,7 +107,7 @@ # define MAXFRACT DBL_DIG # define MAXEXP DBL_MAX_10_EXP -# define BUF (MAXEXP+MAXFRACT+1) /* + decimal point */ +# define BUF (MAXEXP+MAXFRACT+3) /* + decimal point + rounding */ # define DEFPREC 6 static int @@ -420,7 +420,7 @@ * zeroes later, so buffer size stays rational. */ if (prec > MAXFRACT) { - if ((ch != 'g' && ch != 'G') || (flags&ALT)) + if (ch == 'f' || ch == 'F') fpprec = prec - MAXFRACT; prec = MAXFRACT; } else if (prec == -1) This way there are no buffer overruns, and the added zeros are at least never in the exponent. That would at least be a interim solution... But I start to think that the "cvt" function will need a complete re-write for strict conformance. And another point would be, that when you look at the vfnprintf function in the assembler (ARM9, eCosCentric GNU tools 4.3.2-sw) vfnprintf:stmdb r13!,{r4-r11,r14} sub r13,r13,#0x30C This function needs 816 bytes on the stack, even if you do not use any %f formats! Maybe reducing this number could be worth the effort. Bernd. -- Configure bugmail: http://bugs.ecos.sourceware.org/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.
