Bug report: Inlining of functions returning a 'call-clobbered' 64-bit register variable assigned by the output of GNU-style inline assembly may generate wrong code
Compiler: Intel icc v11.1.056 (flags -O2 -use-asm)
Platform: Linux Intel64 (Debian 5.0, gcc 4.3.2)
[cpp]// Sample code: compile with -O2 -use-asm
#include
long gnu_asm (void)
{
// defining a specific register is usually a bad idea, but GNU gcc allows this
register long regvar __asm__ ("rax");
__asm__ (
" movq $12345, %0 "
: "=r"(regvar)
);
// rax is a 'call-clobbered' register, but there are no function calls between
// assignment and return statement. Therefore the return value must be 12345.
return regvar;
}
int main(int argc, char *argv[])
{
long k;
k = gnu_asm();
printf("A printf() function call to clobber the rax register.n");
printf("Output here should be 12345, but it's the return value of previous printf(): '%ld'n", k);
return 0;
}
[/cpp]
With -O2 flag the icc compiler inlines gnu_asm() and generates the following code:
[plain] ...
movq $12345, %rax
movl $_2__STRING.0, %edi
xorl %eax, %eax
call printf
movl $_2__STRING.1, %edi
movq %rax, %rsi
xorl %eax, %eax
call printf
...
[/plain]
The icc compiler generates code like 'k' would be a 'call-clobbered' register variable (in that case it would be legal to overwrite the previous assignment of k by the first printf call). The use of rdi as the register also results in wrong code.
