[RS6000] Fix PR44075, miscompilation of __builtin_eh_return

[Alan Modra <amodra at gmail dot com>]

The problem here is that rs6000_emit_eh_reg_restore, called from an
after-reload splitter, emits rtl depending on the state of
rs6000_stack_info()->lr_save_p.  If lr_save_p is not set, you get a
direct SET of lr.  Any future calls of rs6000_stack_info() then see
the SET via rs6000_ra_ever_killed and return lr_save_p true.  We add
the prologue/epilogue code later, so get an epilogue that restores lr,
wiping out the SET that we want.

Bootstrap and regression test in progress.  OK to apply assuming no
regressions?

	PR target/44075
	* gcc/config/rs6000/rs6000.c (struct machine_function): Reorder
	fields for better packing.  Add lr_save_state.
	(rs6000_ra_ever_killed): Return lr_save_state if set.
	(rs6000_emit_eh_reg_restore): Set lr_save_state.

Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c	(revision 159214)
+++ gcc/config/rs6000/rs6000.c	(working copy)
@@ -115,14 +115,16 @@ typedef struct rs6000_stack {
    This is added to the cfun structure.  */
 typedef struct GTY(()) machine_function
 {
-  /* Flags if __builtin_return_address (n) with n >= 1 was used.  */
-  int ra_needs_full_frame;
   /* Some local-dynamic symbol.  */
   const char *some_ld_name;
   /* Whether the instruction chain has been scanned already.  */
   int insn_chain_scanned_p;
+  /* Flags if __builtin_return_address (n) with n >= 1 was used.  */
+  int ra_needs_full_frame;
   /* Flags if __builtin_return_address (0) was used.  */
   int ra_need_lr;
+  /* Cache lr_save_p after expansion of builtin_eh_return.  */
+  int lr_save_state;
   /* Offset from virtual_stack_vars_rtx to the start of the ABI_V4
      varargs save area.  */
   HOST_WIDE_INT varargs_save_offset;
@@ -17729,6 +17731,9 @@ rs6000_ra_ever_killed (void)
   if (cfun->is_thunk)
     return 0;
 
+  if (cfun->machine->lr_save_state)
+    return cfun->machine->lr_save_state - 1;
+
   /* regs_ever_live has LR marked as used if any sibcalls are present,
      but this should not force saving and restoring in the
      pro/epilogue.  Likewise, reg_set_between_p thinks a sibcall
@@ -17898,6 +17903,12 @@ rs6000_emit_eh_reg_restore (rtx source, 
     }
   else
     emit_move_insn (gen_rtx_REG (Pmode, LR_REGNO), operands[0]);
+
+  /* Freeze lr_save_p.  We've just emitted rtl that depends on the
+     state of lr_save_p so any change from here on would be a bug.  In
+     particular, stop rs6000_ra_ever_killed from considering the SET
+     of lr we may have added just above.  */ 
+  cfun->machine->lr_save_state = info->lr_save_p + 1;
 }
 
 static GTY(()) alias_set_type set = -1;

-- 
Alan Modra
Australia Development Lab, IBM