On Jun 8, 2012, at 4:16 PM, Linus Torvalds wrote:
>
> Looking at the LLVM API docs that Xi referred to, the LLVM preference
> is actually pretty weak.
>
> So I'd almost say continue with the current IR, and just do the
> pointer arithmetic by hand. If LLVM doesn't treat GEP *that*
> specially, doing a separate pointer arithmetic level is quite possibly
> going to just cause problems: simplification of combinations of GEP
> and arithmetic is potentially quite annoying and has problems that
> simplification of perfectly regular addition does not.
I tried a simple strategy to emit an offset:
1) cast the pointer to i8* (char*);
2) emit GEP with the offset computed by sparse;
3) cast the pointer back to its original type.
There is no pointer-integer conversion at all. This seems to work
well with both current IR and LLVM's optimizer.
Here are some examples, using my experimental backend "splay".
struct X { int a, b; };
void szero(struct X *x)
{
x->a = 0;
x->b = 0;
}
splay (with opt -O2) emits:
%0 = getelementptr inbounds %struct.X* %x, i64 0, i32 0
store i32 0, i32* %0, align 4
%1 = getelementptr inbounds %struct.X* %x, i64 0, i32 1
store i32 0, i32* %1, align 4
LLVM's optimizer is even able to "recover" GEP x, 0, 1 for x->b,
from what splay (without opt) actually emits:
%4 = bitcast %struct.X* %x to i32*
%5 = bitcast i32* %4 to i8*
%6 = getelementptr inbounds i8* %5, i64 4
Here's an example of variable offset.
void izero(int *x, size_t n)
{
x[n] = 0;
}
splay (with opt -O2) emits:
%0 = shl nsw i64 %n, 2
%1 = bitcast i32* %x to i8*
%2 = getelementptr inbounds i8* %1, i64 %0
%3 = bitcast i8* %2 to i32*
store i32 0, i32* %3, align 4
This time the result doesn't look as perfect as Clang, which emits
GEP %x, %n, but the assembly code is the same:
movl $0, (%rdi,%rsi,4)
Here goes a more complex one.
void azero(int *x, size_t n)
{
size_t i;
for (i = 0; i != n; ++i)
x[i] = 0;
}
splay (with opt -O2) emits:
...
%1 = bitcast i32* %x to i8*
%2 = shl i64 %n, 2
call void @llvm.memset.p0i8.i64(i8* %1, i8 0, i64 %2, i32 4, i1 false)
...
LLVM's optimizer recognizes the in-loop GEP and rewrites the whole
thing using memset.
- xi
--
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
