[RFC PATCH 2/9] LRA: Swap base_term and index_term for some case

[Kaz Kojima <kkojima at rr dot iij4u dot or dot jp>]

This was discussed in PR55212

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55212#c52
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55212#c53
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55212#c55

and is to fix ICE in assign_by_spills.
In that case, rtlanal.c:decompose_mem_address guesses base and index
regs when INDEX_REG_CLASS is assigned to that base candidate already.
This will be OK for almost targets but is bad for SH which has only
one register r0 in INDEX_REG_CLASS.  The patch fix-ups it with
swapping these candidates.

--
	* lra-constraints.c (process_address_1): Swap base_term and
	index_term if INDEX_REG_CLASS is assigned to base_term already
	when INDEX_REG_CLASS is a single register class.

diff --git a/lra-constraints.c b/lra-constraints.c
index a67bf8a..47cbbec 100644
--- a/lra-constraints.c
+++ b/lra-constraints.c
@@ -2828,6 +2828,20 @@ process_address_1 (int nop, bool check_only_p,
     decompose_mem_address (&ad, SUBREG_REG (op));
   else
     return false;
+  /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
+     index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
+     when INDEX_REG_CLASS is a single register class.  */
+  if (ad.base_term != NULL
+      && ad.index_term != NULL
+      && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
+      && REG_P (*ad.base_term)
+      && REG_P (*ad.index_term)
+      && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
+      && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
+    {
+      std::swap (ad.base, ad.index);
+      std::swap (ad.base_term, ad.index_term);
+    }
   if (! check_only_p)
     change_p = equiv_address_substitution (&ad);
   if (ad.base_term != NULL