Hi!
Our documentation says that paradoxical subregs shouldn't appear
in strict_low_part:
'(strict_low_part (subreg:M (reg:N R) 0))'
This expression code is used in only one context: as the
destination operand of a 'set' expression. In addition, the
operand of this expression must be a non-paradoxical 'subreg'
expression.
but on the testcase below that triggers UB at runtime
store_integral_bit_field emits exactly that.
The following patch fixes it by ensuring the requirement is satisfied.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2021-07-23 Jakub Jelinek <ja...@redhat.com>
PR rtl-optimization/101562
* expmed.c (store_integral_bit_field): Only use movstrict_optab
if the operand isn't paradoxical.
* gcc.c-torture/compile/pr101562.c: New test.
--- gcc/expmed.c.jj 2021-03-04 19:38:00.000000000 +0100
+++ gcc/expmed.c 2021-07-22 11:13:00.996420515 +0200
@@ -921,7 +921,10 @@ store_integral_bit_field (rtx op0, opt_s
}
subreg_off = bitnum / BITS_PER_UNIT;
- if (validate_subreg (fieldmode, GET_MODE (arg0), arg0, subreg_off))
+ if (validate_subreg (fieldmode, GET_MODE (arg0), arg0, subreg_off)
+ /* STRICT_LOW_PART must have a non-paradoxical subreg as
+ operand. */
+ && !paradoxical_subreg_p (fieldmode, GET_MODE (arg0)))
{
arg0 = gen_rtx_SUBREG (fieldmode, arg0, subreg_off);
--- gcc/testsuite/gcc.c-torture/compile/pr101562.c.jj 2021-07-22
11:22:55.745962043 +0200
+++ gcc/testsuite/gcc.c-torture/compile/pr101562.c 2021-07-22
11:22:15.839529580 +0200
@@ -0,0 +1,21 @@
+/* PR rtl-optimization/101562 */
+
+struct S { char c; };
+void baz (struct S a, struct S b);
+
+void
+foo (void)
+{
+ struct S x[1];
+ *(short *)&x[0] = 256;
+ baz (x[0], x[1]);
+}
+
+void
+bar (void)
+{
+ struct S x[1];
+ x[0].c = 0;
+ x[1].c = 1;
+ baz (x[0], x[1]);
+}
Jakub
