This could cause problems when asm blocks contained instructions that the ORCA/C native code optimizer didn’t know about, as in the example below. It might also be possible to trigger this bug without asm blocks (particularly with the large memory model), but I haven’t run into a case that does.
The new approach conservatively assumes that unknown instructions block the optimization. This should be equivalent to the old code with respect to the instructions defined in CGI.pas, except that m_bit_imm should have been treated as blocking the optimization but was not. There are still some other potential problem cases with applying this lda-elimination optimization to arbitrary assembly code, but fixing them might interfere with the optimization in useful cases, so I’m leaving those alone for now.
Here is an example of a program with an asm block affected by this problem:
#pragma optimize 74
int x,y;
/* should print 2 when invoked with argc==1 */
int main(int argc, char **argv)
{
x = argc;
y = argc + 6;
asm {
lda #1
pha
eor >x
bne done
inc argc
done: pla
}
printf("%i\n", argc);
}
This generated invalid code in instances like the following. The code generated for "s = s->u.next" would update the most significant word of s first, then use an indirect load with the half-updated pointer value to update the least significant word of s. This would generally corrupt the result if the new and old pointers had different bank bytes.
#pragma optimize 79
#include <stdio.h>
struct S {
int i;
union {
struct S * next;
} u;
} s1 = {0, 0};
int main (void)
{
struct S * s = &s1;
s = s->u.next;
if (s != 0)
puts("compiler bug detected\n"); /* May not always be triggered, depending on memory contents. */
}
Previously, such initializations would sometimes generate a garbage value pointing up to 65535 bytes beyond the start of the string constant. (This was due to a lack of sign-extension in the object code generation.)
Computing a pointer to before the start of an object invokes undefined behavior, so the previous behavior wasn't technically wrong, but it was unintuitive and served no useful purpose. The new behavior should at least be easier to understand and debug.
