gvn does not remove loads after loop

[tobiasgrosser]

Bugzilla Link	18848
Version	trunk
OS	Linux
Attachments	The LLVM-IR file to reproduce this bug
CC	@hfinkel

Extended Description

When computing the following C code:

void f(float *restrict A, float *restrict B, long N) {
B[0] = A[0];
for (long i = 1; i < N - 1; i += 1) {
B[i - 1] += A[i];
B[i] = A[i]; // Last iteration: B[N-2] = A[N-2];
}
B[N - 2] += A[N - 1];
}

global value numbering successfully eliminates redundant loads from B within the loop body, but it does not eliminate the redundant load of B right after the loop body.

The attached test case was slightly simplified to ensure the loop body dominates the statement after it.

When running the following command 'opt -basicaa store.preopt.ll -gvn -scalar-evolution -analyze' we get the following information:

Index of store in the loop body:

%arrayidx6 = getelementptr inbounds float* %B, i64 %storemerge9
--> {(4 + %B),+,4}<%for.body> Exits: (-8 + (4 * %N) + %B)

Index of load after the loop:
%arrayidx11 = getelementptr inbounds float* %B, i64 %sub10
--> (-8 + (4 * %N) + %B)

Both registers have the same value after the loop. We should eliminate both the load that comes right after the store as well as the unnecessary index computation.