On Tue, Oct 16, 2012 at 08:59:27PM +0300, Pekka Enberg wrote:
> On Fri, Oct 12, 2012 at 9:25 PM, Pekka Enberg <penberg@xxxxxxxxxx> wrote:
> > Sounds plausible but I'm still uneasy with the idea that LLVM backend
> > needs to reshuffle instructions like this.
Actually, the situation of Phi nodes in LLVM is actually slightly more
complex: They require "one pair (of value and BB) for each predecessor
basic block of the current block"[1]. This mean that we'll sometimes
need to insert phi nodes into BBs that don't directly use a value.
Consider the following piece of C code:
extern int done(void);
extern void foo(int);
static void test(void) {
int i;
for (i = 0; ; i++) {
if (done())
break;
foo(i);
}
}
Running it through test-linearize exhibits the problem:
[ I renamed the basic blocks to L0-L3 to increase readability. ]
test:
.L0:
<entry-point>
phisrc.32 %phi2(i) <- $0
br .L1
.L1:
call.32 %r1 <- done
br %r1, .L3, .L2
.L2:
phi.32 %r2(i) <- %phi2(i), %phi3(i)
call foo, %r2(i)
add.32 %r4 <- %r2(i), $1
phisrc.32 %phi3(i) <- %r4
br .L1
.L3:
ret
To comply with the "LLVM rules" we'd need to move the phi intruction up
into "L1".
regards,
Jonathan
--
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
