https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79571
Bernd Schmidt <bernds at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |bernds at gcc dot gnu.org
--- Comment #8 from Bernd Schmidt <bernds at gcc dot gnu.org> ---
I was also playing with this before I noticed Jakub was investigating. As an
experiment, I came up with the following, trying to recognize situations where
picking one alternative would cause an infinite cycle:
Index: lra-constraints.c
===================================================================
--- lra-constraints.c (revision 245685)
+++ lra-constraints.c (working copy)
@@ -1899,6 +1899,7 @@ process_alt_operands (int only_alternati
int reload_nregs, reload_sum;
bool costly_p;
enum reg_class cl;
+ bool must_not_reload_op1 = false;
/* Calculate some data common for all alternatives to speed up the
function. */
@@ -1932,6 +1933,15 @@ process_alt_operands (int only_alternati
= regno_reg_rtx[hard_regno[nop]];
}
+ /* See if we have an insn of the form
+ (set (pseudo) (something)
+ If yes, then we should not try to reload operand 1 into a pseudo,
+ because this would cause an infinite cycle. */
+ if (curr_insn_set != NULL_RTX
+ && operand_reg[0] != NULL_RTX
+ && hard_regno[0] < 0)
+ must_not_reload_op1 = true;
+
/* The constraints are made of several alternatives. Each operand's
constraint looks like foo,bar,... with commas separating the
alternatives. The first alternatives for all operands go
@@ -2193,7 +2203,10 @@ process_alt_operands (int only_alternati
switch (get_constraint_type (cn))
{
case CT_REGISTER:
- cl = reg_class_for_constraint (cn);
+ if (nop == 1 && must_not_reload_op1)
+ cl = NO_REGS;
+ else
+ cl = reg_class_for_constraint (cn);
if (cl != NO_REGS)
goto reg;
break;
Sadly, it seems to be ineffective (or at least incomplete), it goes into a
different infinite cycle.