KHMan
keinh****@gmail*****
Thu Mar 1 11:05:03 JST 2018
On 2/16/2018 5:36 AM, John Beale wrote: > [snip] > I am using the MINGW32 shell and gcc -v reports (MinGW.org > GCC-6.3.0-1) 6.3.0 I can get different results (44/45) using the same compiler running different settings. Here are some -S outputs. -O2 will optimize away the divide at compile time: movl $45, 4(%esp) movl $LC0, (%esp) call _printf leave -O0 will either use x87 instructions (80 bits precision internally) or SSE2 (64 bits) depending on other options: fldl LC0 fstpl 56(%esp) fldl LC1 fstpl 48(%esp) fldl 56(%esp) fdivl 48(%esp) fnstcw 30(%esp) movzwl 30(%esp), %eax movb $12, %ah movw %ax, 28(%esp) fldcw 28(%esp) fistpl 44(%esp) fldcw 30(%esp) The x87 control word is changed. The $12 is for setting RC=11 (truncate). Then FISTPL will convert with truncation. And you may get a 44 as the result. At -O2, changing a and b to volatile gets me 44 too. Different precisions (64 vs 80) may give you slightly different results (1 ULP or so) and then truncate, boom. There are gcc math options and math library features that we can play with, but IMHO relying on them in this case is the wrong solution. Prefer robust over fragile techniques. Please get expert guidance if you are unsure of these things. > CODE: SELECT ALL > <https://www.raspberrypi.org/forums/viewtopic.php?f=63&t=205567#> > > |#include <stdio.h> void main() { double a = 4.5; double b = 0.1; > int i = a / b; printf("%d\n",i); }| > > Ubuntu 16.04 on VirtualBox/Win10 result: 45 > Raspberry Pi Raspbian (ARM) result: 45 > Online compiler demo: at www.onlinegdb.com/online_c_compiler > <https://www.onlinegdb.com/online_c_compiler> : 45 > MinGW32 gcc 6.3.0-1 result: 44 -- Cheers, Kein-Hong Man (esq.) Selangor, Malaysia
Fork