https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90340
--- Comment #4 from Martin Liška <marxin at gcc dot gnu.org> ---
Simplified test-case:
$ cat /tmp/csibe.c
int a;
int foo(char c) {
switch (c) {
case 'c':
return a;
case 's':
return 3;
case 'n':
return 1;
case '%':
return -2;
case 'o':
return a + 2;
break;
case 'X':
case 'x':
return 2222;
case 'd':
case 'i':
case 'u':
return 3333;
default:
return 0;
}
}
$ ~/Programming/bloaty/bloaty bad.o -- good.o
VM SIZE FILE SIZE
-------------- --------------
[ = ] 0 .rela.rodata +528 [NEW]
[NEW] +176 .rodata +176 [NEW]
[ = ] 0 [ELF Headers] +128 +15%
[ = ] 0 .rela.text +24 +50%
[ = ] 0 .symtab +24 +10.0%
[ = ] 0 .shstrtab +13 +15%
[ = ] 0 [Unmapped] +5 +62%
[ = ] 0 .comment -22 -33.8%
-20.7% -28 .text -28 -20.7%
+81% +148 TOTAL +848 +56%
Where bad.o is GCC 9.1 and good.o is GCC 8.3.1.
Reason is that GCC 9 can newly generate jump tables for a fraction of switch
statement and we allow:
/* Max growth ratio for code that is optimized for size. */
static const unsigned HOST_WIDE_INT max_ratio_for_size = 3;
;; GIMPLE switch case clusters: 37 88 JT(values:8 comparisons:8 range:22
density: 36.36%):99-120
I can imagine shrinking the parameter to 2 for -Os.
Can you please help me how to run all CSIBE benchmarks?
