I know that there is some recent discussions about the optimizations on 
volatile loads and stores, and the conclusion is that LLVM is doing the right 
thing. But this time the story seems different. 

See below for details. The test case will be posted shortly. 

=======================================================================
For my first example, LLVM is able to preserve both volatile load and store 
correctly. 

void PassThrough(volatile int* DataIn, volatile int* DataOut)
{
  int i;
  int buffer[64];

  for (i = 0; i < 64; i++)
    buffer[i] = *DataIn;

  for (i = 0; i < 64; i++)
    *DataOut = buffer[i];
}

=======================================================================
For the second one, however, LICM will move the volatile store outside of the 
loop, which is apparently wrong. 

void Transpose(volatile int* DataIn, volatile int* DataOut)
{
  int i, j;
  int buffer[64];

  for (i = 0; i < 64; i++)
    buffer[i] = *DataIn;

  for (i = 0; i < 8; i++)
    for (j = 0; j < 8; j++)
      *DataOut = buffer[j * 8 + i];
}

=========================================================
What I did:

llvm-gcc -O0 -c -emit-llvm test_volatile.c -o test.bc
opt -mem2reg -licm test.bc -o new.bc -f