Hi!
The following testcase ICEs in cross to riscv64-linux. The problem is
that we have a DImode integral constant (that doesn't fit into SImode),
which is pushed into a constant pool and later access just the first half of
it using a MEM. When plus_constant is called on such a MEM, if the constant
has mode, we verify the mode, but if it doesn't, we don't and ICE later on
when we think the CONST_INT is a valid SImode constant.
Fixed thusly, tested with cross to riscv64-linux and bootstrapped/regtested
on x86_64-linux and i686-linux, ok for trunk?
2020-03-03 Jakub Jelinek <ja...@redhat.com>
PR rtl-optimization/94002
* explow.c (plus_constant): Punt if cst has VOIDmode and
get_pool_mode is different from mode.
* gcc.dg/pr94002.c: New test.
--- gcc/explow.c.jj 2020-01-12 11:54:36.564411130 +0100
+++ gcc/explow.c 2020-03-02 22:09:19.544380020 +0100
@@ -128,6 +128,9 @@ plus_constant (machine_mode mode, rtx x,
cst = gen_lowpart (mode, cst);
gcc_assert (cst);
}
+ else if (GET_MODE (cst) == VOIDmode
+ && get_pool_mode (XEXP (x, 0)) != mode)
+ break;
if (GET_MODE (cst) == VOIDmode || GET_MODE (cst) == mode)
{
tem = plus_constant (mode, cst, c);
--- gcc/testsuite/gcc.dg/pr94002.c.jj 2020-03-02 22:05:58.508338170 +0100
+++ gcc/testsuite/gcc.dg/pr94002.c 2020-03-02 22:05:32.864715503 +0100
@@ -0,0 +1,13 @@
+/* PR rtl-optimization/94002 */
+/* { dg-do compile } */
+/* { dg-options "-O1 -fno-tree-dce -fno-tree-reassoc" } */
+/* { dg-additional-options "-fPIC" { target fpic } } */
+
+unsigned a, b;
+
+void
+foo (void)
+{
+ __builtin_sub_overflow (b, 44852956282LL, &a);
+ a += ~b;
+}
Jakub
