Prevent clever optimization to prematurely quit loop in safe memcmp
The previous version of `mbedtls_ssl_safer_memcmp` did not qualify the
pointers to the arrays to be compared as volatile, theoretically
opening the possibility for the compiler to notice that the loop
operation `diff |= A[i] ^ B[i]` is pointless if `diff = -1`. This
commit changes this. It also declares the stack variable `diff` as
volatile, to force read and write in every loop; omitting that, the
compiler would still be allowed to get away with reading `A[i]` and
`B[i]` but not doing the XOR and not updating `diff`.
diff --git a/include/mbedtls/ssl_internal.h b/include/mbedtls/ssl_internal.h
index 756360b..8d3ab61 100644
--- a/include/mbedtls/ssl_internal.h
+++ b/include/mbedtls/ssl_internal.h
@@ -600,9 +600,9 @@
static inline int mbedtls_ssl_safer_memcmp( const void *a, const void *b, size_t n )
{
size_t i;
- const unsigned char *A = (const unsigned char *) a;
- const unsigned char *B = (const unsigned char *) b;
- unsigned char diff = 0;
+ volatile const unsigned char *A = (volatile const unsigned char *) a;
+ volatile const unsigned char *B = (volatile const unsigned char *) b;
+ volatile unsigned char diff = 0;
for( i = 0; i < n; i++ )
diff |= A[i] ^ B[i];