Here's a testcase in the regression suite:
test/Regression/CFrontend/2004-02-13-StringInit.c.tr

Here's another testcase:

struct T { int A, B; };
struct S { struct T X, Y; };

int test1(int x) {
struct S A = { {x, x+1}, {0, 12 } };
}

int test2() {
struct S A[] = {
{ { 0, 1}, {2, 3} },
{ { 2, 1}, {2, 3} },
{ { 3, 1}, {2, 3} },
{ { 4, 1}, {2, 3} },
{ { 5, 1}, {2, 3} },
};
}

int test3(int x) {
const int array[] = {
17, x, 123, 123, 49, 17, 23,
};
return array[x];
}

Here is the example from comment #​1:

// RUN: %llvmgcc -xc %s -c -o - | llvm-dis | grep llvm.memset
char test(int X) {
char str[10000] = "abc"; // tail should be memset.
return str[X];
}

From comment #​1, we memset the first two arrays. However, we use a bunch of SSE instructions to initialize the "Pairs" variable:

xorps	%xmm0, %xmm0
movaps	%xmm0, -24(%ebp)
movaps	%xmm0, -40(%ebp)
movaps	%xmm0, -56(%ebp)
movaps	%xmm0, -72(%ebp)
movaps	%xmm0, -88(%ebp)
movl	$1, -88(%ebp)
movl	$2, -84(%ebp)

We don't care about llvm-gcc anymore, so stealing this for Clang. After r120645, clang now generates the right code for the second Arr testcase in Comment 0, and also handles hte Pairs example nicely from Comment #​0. It handles the examples in comment #​2, and the "str" case from comment 3. The example in comment 1 doesn't exist anymore.

The one remaining issue is that it does not handle this well from comment #​0:

int foo(int X) {
int Arr[100] = { X }; // Should use memset
}

Finished off in r120692. We now generate a memset + 1 store.