https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112413
--- Comment #3 from Vincent Riviere <vincent.riviere at freesbee dot fr> ---
(In reply to Andrew Pinski from comment #1)
> I don't see any issues with the output of gcc. Are you sure this is not a
> binutils gnu as issue where the offsets are done incorrectly there.
Yes, I'm sure it's a gcc bug.
With the testcase I initially provided, by chance it's the favourable case.
But if I artificially add a misalignment with a nop, for example, the wrong
result appears:
$ cat swi2.c
int g;
void f(int i)
{
asm("nop");
switch (i)
{
case 0: g = 6082; break;
case 1: g = 9332; break;
case 2: g = 5642; break;
case 3: g = 1423; break;
case 4: g = 2152; break;
case 5: g = 6779; break;
case 6: g = 7074; break;
case 7: g = 8280; break;
}
}
$ m68k-linux-gcc -Os -c swi2.c -mlong-jump-table-offsets -malign-int
$ m68k-linux-objdump -d swi2.o
swi2.o: file format elf32-m68k
Disassembly of section .text:
00000000 <f>:
0: 202f 0004 movel %sp@(4),%d0
4: 4e71 nop
6: 7207 moveq #7,%d1
8: b280 cmpl %d0,%d1
a: 6536 bcss 42 <f+0x42>
c: e588 lsll #2,%d0
e: 203b 0808 movel %pc@(18 <f+0x18>,%d0:l),%d0 |right offset
12: 4efb 0802 jmp %pc@(16 <f+0x16>,%d0:l) |wrong offset
16: 284c moveal %a4,%a4 |harmful filler
18: 0000 0020 orib #32,%d0
1c: 0000 002c orib #44,%d0
20: 0000 0038 orib #56,%d0
See that:
- actual jump table starts at offset 0x18
- at offset 0x16, a useless "moveal %a4,%a4" instruction is inserted as filler
- at offset 0xe, offset 0x18 is used appropriately for label .L4. So the right
jump table entry is properly read.
- but at offset 0x12, a *wrong* offset 0x16 is used for the jump. That's
actually the offset of the filler, while it should be 0x18 for label .L4.
This can't work:
- the jump table offsets are computed from the start of the jump table
- but jmp, with that "2" hardcoded as offset, expects offsets being relative to
the address right after itself.
So if a filler is inserted between jmp and actual table contents, as in the
example above, the jump occurs to a wrong address.