[Bug optimization/14824] removing constant assignments from loops not done even for simple conditional expressions

[niemayer at isg dot de]

------- Additional Comments From niemayer at isg dot de  2004-04-02 15:20 -------
As a followup to myself, the following example may show more visibly why I'm so
confused about the non-optimization: The following code is compiled more
efficiently with "-O3 -fno-inline" than with "-O3" alone, because once the
"internals" of the function with the __attribute__((const)) become invisible to
the compiler, the assignment _is_ removed from the loop body.

The source: 
--------------------------------------------------------------------
extern int y;

const int foo(const int x) __attribute__((const));

const int test(void) {
	
	int i;
	int z;
	const int x = y;
	
	asm volatile ("testlabel1: ");
	for (i = 0; i < x; i++) {
		/*
		if (x) z = 4;
		else z = 5;
		*/
		z = foo(x);
	}
	asm volatile ("testlabel2: ");

	return z;	
}

const int foo(const int x) {
		if (x) return 4;
		return 5;	
}
------------------------------------------------

compiled with "-O3", you'll fiend the same unneccessary x-times-assignment as
described above. But once you compile it with "-O3 -fno-inline", the code looks
like:

------------------------------------------------
        testlabel1: 
#NO_APP
        testl   %ebx, %ebx
        jg      .L9
.L8:
#APP
        testlabel2: 
#NO_APP
        movl    -4(%ebp), %ebx
        movl    %ebp, %esp
        popl    %ebp
        ret
        .p2align 4,,7
.L9:
        movl    %ebx, (%esp)
        call    foo
        movl    %ebx, %edx
        .p2align 4,,15
.L6:
        decl    %edx
        jne     .L6
        jmp     .L8
--------------------------------------------

which is much more efficient for large values of x. 

We certainly don't want inlining to hurt performance like this, do we? :-)





-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14824