On 5/15/2026 2:04 AM, Richard Biener wrote:
On Tue, 12 May 2026, Xin Wang wrote:
When split_loop does iteration space splitting, split_at_bb_p may
swap the guard condition so that operand 0 is always the loop IV
and operand 1 is the invariant. For example, "t < i" (LT_EXPR)
becomes "i > t" (GT_EXPR). This can cause initial_true to be
false, meaning loop1 handles iterations where the guard is false
and loop2 handles iterations where the guard is true.
The profile update code used true_edge->probability for loop1 and
its inverse for loop2. That is correct only when loop1 keeps the
true branch. When initial_true is false, loop1 keeps the false
branch and loop2 keeps the true branch, so all profile quantities
must follow those semantic edges instead of the raw true/false edge
names.
Derive loop1_edge and loop2_edge once from initial_true and use them
consistently for loop_version's then probability, the split-loop BB
count scaling, and the iteration estimate scaling. Also make
fix_loop_bb_probability take loop1/loop2 edges explicitly, rather
than assuming its arguments are true/false edges.
The bug caused BB counts in the split loops to be swapped when
initial_true is false: the loop body whose guard is forced false
(loop1, executing fewer iterations) would get the higher profile
count, and vice versa. It could also leave the precondition edge
probabilities and iteration estimates based on the wrong split edge.
OK - much better to follow after the renaming.
And I've pushed it to the trunk on Xin Wang's behalf.
jeff
