On Fri, Aug 26, 2011 at 8:28 AM, Pekka Enberg <penberg@xxxxxxxxxx> wrote:
> On Fri, Aug 26, 2011 at 6:59 AM, Christopher Li <sparse@xxxxxxxxxxx> wrote:
>> On Mon, Aug 22, 2011 at 6:57 AM, Pekka Enberg <penberg@xxxxxxxxxx> wrote:
>>> The value of 'ctype->bit_size' is set to 1 for booleans which confuses the i386
>>> backend:
>>>
>>> ./compile allocate.c
>>> compile: compile-i386.c:1406: emit_binop: Assertion `0' failed.
>>> Aborted
>>>
>>> Looking at the code, we assume that "bit_size / 8" gives a sane result on
>>> various places. This patch fixes the problem by bumping bit_size to 8 for
>>> booleans. This also makes sizeof(_Bool) return 1 which is consistent with what
>>> GCC 4.4.3, for example, does.
>>>
>>> diff --git a/target.c b/target.c
>>> index 17b228a..6a535bc 100644
>>> --- a/target.c
>>> +++ b/target.c
>>> @@ -14,7 +14,7 @@ int max_alignment = 16;
>>> /*
>>> * Integer data types
>>> */
>>> -int bits_in_bool = 1;
>>> +int bits_in_bool = 8;
>>
>> I object this part. I consider the sizeof(_Bool) == 1 as external behaviour.
>> But internally we should know that the real useful part of bool is in just one
>> bit, not any bit of that 1 byte storage.
>
> You missed the most important part of my reasoning: sparse code
> already expects "bit_size / 8" to return a non-zero number and it's
> not just compile-i386.c! So while I don't disagree with you that we
> should internally know that a bool is just one bit, I don't consider
> that to be relevant for this particular patch.
Oh, I see the same problem with sparse-llvm when generating code for OP_CAST.
This little C function:
int sete(int x, int y)
{
return x == y;
}
is compiled by GCC to:
00000170 <setne>:
170: 55 push %ebp
171: 89 e5 mov %esp,%ebp
173: 8b 45 0c mov 0xc(%ebp),%eax
176: 39 45 08 cmp %eax,0x8(%ebp)
179: 5d pop %ebp
17a: 0f 95 c0 setne %al
17d: 0f b6 c0 movzbl %al,%eax
180: c3 ret
Sparse-llvm compiles the code to this which seems wrong:
00000140 <sete>:
140: 31 c9 xor %ecx,%ecx
142: 8b 44 24 08 mov 0x8(%esp),%eax
146: 39 44 24 04 cmp %eax,0x4(%esp)
14a: b8 ff ff ff ff mov $0xffffffff,%eax
14f: 0f 45 c1 cmovne %ecx,%eax
152: c3 ret
However, if I bump up 'bits_in_bool' to 32:
diff --git a/target.c b/target.c
index 17b228a..2b83498 100644
--- a/target.c
+++ b/target.c
@@ -14,7 +14,7 @@ int max_alignment = 16;
/*
* Integer data types
*/
-int bits_in_bool = 1;
+int bits_in_bool = 32;
int bits_in_char = 8;
int bits_in_short = 16;
int bits_in_int = 32;
I get this from sparse-llvm:
00000140 <sete>:
140: 8b 44 24 08 mov 0x8(%esp),%eax
144: 39 44 24 04 cmp %eax,0x4(%esp)
148: 0f 94 c0 sete %al
14b: c3 ret
I don't know sparse well enough to know what's the right thing to do
here. Linus, Jeff, Chris, someone, anyone, help!!!
Pekka
--
To unsubscribe from this list: send the line "unsubscribe linux-sparse" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
