[gcc(refs/vendors/ARM/heads/morello)] c: Ensure intcap save_expr operands are folded

[Alex Coplan <acoplan@gcc.gnu.org>]

https://gcc.gnu.org/g:aa8dde4799379b90301785bf0beff2c20935eacc

commit aa8dde4799379b90301785bf0beff2c20935eacc
Author: Alex Coplan <alex.coplan@arm.com>
Date:   Mon Dec 12 18:17:10 2022 +0000

    c: Ensure intcap save_expr operands are folded
    
    The commment above c_fully_fold notes the following requirement:
    
       Function arguments have already been folded
       before calling this function, as have the contents of SAVE_EXPR,
       TARGET_EXPR, BIND_EXPR, VA_ARG_EXPR, OBJ_TYPE_REF and
       C_MAYBE_CONST_EXPR.
    
    When we recently introduced save_exprs to avoid duplicate side-effects
    when forming .REPLACE_ADDRESS_VALUEs in the C front-end, we didn't take
    this requirement into account.
    
    As the testcase added with the patch shows, this leads to
    c_maybe_const_expr nodes escaping the C front-end, and subsequent ICEs
    in the gimplifier.  This patch fixes the issue by calling c_fully_fold
    prior to calling save_expr on intcap operands.

Diff:
---
 gcc/c/c-typeck.c                                 |  6 +++---
 gcc/testsuite/c-c++-common/torture/intptr-fold.c | 10 ++++++++++
 2 files changed, 13 insertions(+), 3 deletions(-)

diff --git a/gcc/c/c-typeck.c b/gcc/c/c-typeck.c
index 1fcaa7dbebb..817b8ac844b 100644
--- a/gcc/c/c-typeck.c
+++ b/gcc/c/c-typeck.c
@@ -4401,7 +4401,7 @@ build_unary_op (location_t location, enum tree_code code, tree xarg,
 	default:
 	  intcap = unary_op_get_intcap_provenance (arg);
 	  if (arg == intcap)
-	    intcap = arg = save_expr (arg);
+	    intcap = arg = save_expr (c_fully_fold (arg, false, NULL));
 	  arg = drop_capability (arg);
 	}
     }
@@ -11970,9 +11970,9 @@ build_binary_op_1 (location_t location, enum tree_code code,
      c + cv --> REPLACE_ADDRESS_VALUE (c, (non-cap type) c + cv)
      we need to avoid evaluating c twice.  */
   if (intcap == op0)
-    intcap = op0 = save_expr (op0);
+    intcap = op0 = save_expr (c_fully_fold (op0, false, NULL));
   else if (intcap == op1)
-    intcap = op1 = save_expr (op1);
+    intcap = op1 = save_expr (c_fully_fold (op1, false, NULL));
 
   if (common_ic_type)
     {
diff --git a/gcc/testsuite/c-c++-common/torture/intptr-fold.c b/gcc/testsuite/c-c++-common/torture/intptr-fold.c
new file mode 100644
index 00000000000..f62b398ae17
--- /dev/null
+++ b/gcc/testsuite/c-c++-common/torture/intptr-fold.c
@@ -0,0 +1,10 @@
+/* { dg-do compile } */
+/* This would ICE in the gimplifier on capability targets due to
+   c_maybe_const_expr escaping the C front-end.  */
+typedef __INTPTR_TYPE__ T;
+typedef struct {
+  T a;
+} S;
+T f(T x) {
+  return -(x ? ((S *)x)->a : x);
+}