The following code generates suboptimal code on i686-pc-linux-gnu: with Interfaces; use Interfaces; with System.Storage_Elements; package A is Var : unsigned_8; for Var'Address use System.Storage_Elements.To_Address (1000); pragma Volatile (Var); end A; with A; with Interfaces; use Interfaces; procedure U is begin A.Var := A.Var or 1; A.Var := A.Var or 2; end U; With -O3 -fomit-frame-pointer, we get: _ada_u: movl a__var, %edx <=== indirection instead of... movzbl (%edx), %eax <=== ...direct access here orl $1, %eax movb %al, (%edx) movl a__var, %edx <=== useless reloading of a__var movzbl (%edx), %eax orl $2, %eax movb %al, (%edx) ret If the variable is declared in the same compilation unit, we get: _ada_u: movzbl 1000, %eax orl $1, %eax movb %al, 1000 movzbl 1000, %eax orl $2, %eax movb %al, 1000 ret We should probably get the same code in both cases.
We don't try to optimize volatile accesses.
Well, in the second example I give (variable declared in the same compilation unit), the access is optimized. What I don't get is the difference between both cases, while the compiler does have the same information at its disposal.