commit 50e1e417b5cc68a6fd280e0a67c5c4eba524b70b
Author: Richard Sandiford <richard.sandiford@arm.com>
Date:   Sun Mar 22 18:51:48 2020 +0000

    rs6000: Allow FPRs to change between SDmode and DDmode [PR94254]
    
    g:497498c878d48754318e486428e2aa30854020b9 caused lra to cycle
    on some SDmode reloads for power6.  As explained in more detail
    in the PR comments, the problem was a conflict between two target
    hooks: rs6000_secondary_memory_needed_mode required SDmode FPR
    reloads to use DDmode memory (rightly, since using SDmode memory
    wouldn't make progress) but rs6000_can_change_mode_class didn't
    allow FPRs to change from SDmode to DDmode.  Previously lra
    ignored that and changed the mode anyway.
    
    From what Segher says, it sounds like the "from_size < 8 || to_size < 8"
    check is mostly there for SF<->64-bit subregs, and that SDmode is stored
    in the way that target-independent code expects.  This patch therefore
    allows SD<->DD changes.
    
    I wondered about checking for SD<->64-bit changes instead, but that
    seemed like an unnecessary generalisation for this stage.
    
    2020-03-23  Richard Sandiford  <richard.sandiford@arm.com>
    
    gcc/
            PR target/94254
            * config/rs6000/rs6000.c (rs6000_can_change_mode_class): Allow
            FPRs to change between SDmode and DDmode.

--- a/src/gcc/config/rs6000/rs6000.cc
+++ b/src/gcc/config/rs6000/rs6000.cc
@@ -13668,6 +13668,15 @@ rs6000_can_change_mode_class (machine_mo
 	      || (to == SDmode && from == DDmode))
 	    return true;
 
+	  /* Allow SD<->DD changes, since SDmode values are stored in
+	     the low half of the DDmode, just like target-independent
+	     code expects.  We need to allow at least SD->DD since
+	     rs6000_secondary_memory_needed_mode asks for that change
+	     to be made for SD reloads.  */
+	  if ((to == DDmode && from == SDmode)
+	      || (to == SDmode && from == DDmode))
+	    return true;
+
 	  if (from_size < 8 || to_size < 8)
 	    return false;