Extended Description

This simple C program:

void g(int *);

void f() {
int array[20];
int i;
for (i = 0; i<100; i++) {
array [0] = 0;
}
g(array);
}

is optimized down to:

[...]
bb:
%i.04 = phi i32 [ 0, %entry ], [ %1, %bb ]
store i32 0, i32* %0, align 4
%1 = add nsw i32 %i.04, 1
%exitcond = icmp eq i32 %1, 100
br i1 %exitcond, label %bb2, label %bb
[...]

by llvm-gcc -O2. If we were to run another opt -licm over this then it sinks the store instruction. The problem is that we have more complicated IR at the point of time when LICM actually runs:

[...]
bb:
%3 = phi i32* [ %1, %bb.nph ], [ %2, %bb1 ]
%storemerge1 = phi i32 [ 0, %bb.nph ], [ %storemerge, %bb1 ]
store i32 0, i32* %3, align 4
%4 = add nsw i32 %storemerge1, 1
br label %bb1
bb1:
%storemerge = phi i32 [ %4, %bb ]
%5 = icmp slt i32 %storemerge, 100
br i1 %5, label %bb, label %bb1.bb2_crit_edge
[...]

The full .ll that LICM isn't getting is attached.