On Thu, Jul 24, 2025 at 2:39 AM magic0826gc via Gcc <gcc@gcc.gnu.org> wrote:
>
> Dear GCC Developers,
> I'm writing to report an observation regarding loop unswitching behavior when
> compiling the attached code with GCC 15 using -O2 -funswitch-loops
> optimization flags. Here's a detailed analysis of the issue:
> Code Overview:
> int
> foo(double *a, double *b, double *c, double *d, double *r, int size, int
> order)
> {
> for (int i = 0; i < size; i++)
> {
> double tmp;
> if (order < 3)
> tmp = -8 * a[i];
> else
> tmp = -4 * b[i];
> double x = 3 * tmp + d[i] + tmp;
> if (5 > order)
> x = 2;
> if (order == 12345)
> x = 5;
> for (int j = 0; j < size; j++)
> {
> double y = 3.4f * tmp + d[i];
> r[j] += x + y;
> }
> }
> return 0;
> }
> Outer loop with multiple conditional checks on 'order' parameter.
> Complex control flow with three independent conditionals.
> Inner loop containing runtime-invariant computations.
> Observed Behavior:
> The compiler fails to unswitch the outer loop despite the presence of
> multiple loop-invariant conditions. This prevents:
> Constant propagation of 'x' variable. Elimination of dead code paths (order
> == 12345 and 5 > order check).
> Current unswitching (tree_unswitch_outer_loop) appears limited to inner loop
> guards. What prevents handling nested loops with complex control flow?
> Potential Constraints:
> Register pressure considerations in multi-versioned loops. Code size
> explosion vs. runtime tradeoffs. Interaction with subsequent optimization
> passes.
> Could you help clarify:
> Architectural constraints in current unswitching implementation?
> Planned improvements for nested loop optimization?
Unswitching does currently not implement unswitching of non-innermost
loops. A quick
search in bugzilla doesn't reveal a bug where the summary notes this
so I suggest
you file a bugzilla for this.
> Recommended workarounds for such patterns?
You'd have to manually unswitch the outer loop.
Richard.
> Thank you for your time and expertise. I'd be happy to provide additional
> details if needed.
> Best regards,
