[Bug c/85805] New: Improper code generation for 64 bit comparisons on avr-gcc

["sandor.zsuga at jubatian dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85805

            Bug ID: 85805
           Summary: Improper code generation for 64 bit comparisons on
                    avr-gcc
           Product: gcc
           Version: 4.8.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: sandor.zsuga at jubatian dot com
  Target Milestone: ---

GCC version:

avr-gcc (GCC) 4.8.1

Compiling with either -O1 or -O2 optimizations enabled, tested with some ATMega
and XMega targets. Test case:


#include <stdint.h>

uint8_t volatile tmp;

__attribute__((noinline)) void test_64(uint64_t d64)
{
 if ((d64 & 0xFF800000UL) == 0xFF800000UL){
  tmp ++;
 }
}

__attribute__((noinline)) void test_32(uint32_t d32)
{
 if ((d32 & 0xFF800000UL) == 0xFF800000UL){
  tmp ++;
 }
}

int main(void)
{
 test_64(0);
 test_32(0);

 while(1);
}


A cut from the output assembly, showing the critical part (the code generated
for test_64, test_32 and main):


00000228 <test_64>:
 228:   08 95           ret

0000022a <test_32>:
 22a:   66 27           eor r22, r22
 22c:   77 27           eor r23, r23
 22e:   80 78           andi    r24, 0x80   ; 128
 230:   61 15           cp  r22, r1
 232:   71 05           cpc r23, r1
 234:   80 48           sbci    r24, 0x80   ; 128
 236:   9f 4f           sbci    r25, 0xFF   ; 255
 238:   09 f0           breq    .+2         ; 0x23c <test_32+0x12>
 23a:   08 95           ret
 23c:   80 91 00 20     lds r24, 0x2000
 240:   8f 5f           subi    r24, 0xFF   ; 255
 242:   80 93 00 20     sts 0x2000, r24
 246:   08 95           ret

00000248 <main>:
 248:   20 e0           ldi r18, 0x00   ; 0
 24a:   30 e0           ldi r19, 0x00   ; 0
 24c:   40 e0           ldi r20, 0x00   ; 0
 24e:   50 e0           ldi r21, 0x00   ; 0
 250:   60 e0           ldi r22, 0x00   ; 0
 252:   70 e0           ldi r23, 0x00   ; 0
 254:   80 e0           ldi r24, 0x00   ; 0
 256:   90 e0           ldi r25, 0x00   ; 0
 258:   0e 94 14 01     call    0x228   ; 0x228 <test_64>
 25c:   60 e0           ldi r22, 0x00   ; 0
 25e:   70 e0           ldi r23, 0x00   ; 0
 260:   cb 01           movw    r24, r22
 262:   0e 94 15 01     call    0x22a   ; 0x22a <test_32>
 266:   ff cf           rjmp    .-2         ; 0x266 <main+0x1e>


It seems like the compiler incorrectly determines that the "if" is always false
in the 64 bit case, and produces a corresponding result (a function doing
nothing). A native version of GCC produces the expected code (the comparison
being performed).