Hi!
handle_operand_addr for INTEGER_CSTs uses wi::min_precision (UNSIGNED
for non-negative constants, SIGNED for negative ones) and from that
computes mp as minimum number of limbs which can represent that value,
and in some cases creates a test BITINT_TYPE with that precision to
categorize it and decide based on that what types to use on the constant
emitted into memory. For the actual precisions (what will be passed
to libgcc) it actually already uses MAX/MIN to adjust the corner cases:
*prec = MAX (min_prec, 1);
...
*prec = MIN ((int) -min_prec, -2);
but for integer_zerop min_prec will be 0,
mp = CEIL (min_prec, limb_prec) * limb_prec;
will be also 0 and we ICE trying to build unsigned BITINT_TYPE with
0 precision.
Fixed thusly by noting even 0 has to be encoded at least as one limb.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2023-12-01 Jakub Jelinek <ja...@redhat.com>
PR middle-end/112771
* gimple-lower-bitint.cc (bitint_large_huge::handle_operand_addr):
Use mp = limb_prec if it is zero.
* gcc.dg/bitint-44.c: New test.
--- gcc/gimple-lower-bitint.cc.jj 2023-11-30 17:24:59.828026693 +0100
+++ gcc/gimple-lower-bitint.cc 2023-11-30 17:42:07.172487347 +0100
@@ -2179,6 +2179,8 @@ bitint_large_huge::handle_operand_addr (
*prec = MIN ((int) -min_prec, -2);
}
mp = CEIL (min_prec, limb_prec) * limb_prec;
+ if (mp == 0)
+ mp = limb_prec;
if (mp >= (unsigned) TYPE_PRECISION (TREE_TYPE (op)))
type = TREE_TYPE (op);
else
--- gcc/testsuite/gcc.dg/bitint-44.c.jj 2023-11-30 17:45:02.421017172 +0100
+++ gcc/testsuite/gcc.dg/bitint-44.c 2023-11-30 17:39:51.968393081 +0100
@@ -0,0 +1,10 @@
+/* PR middle-end/112771 */
+/* { dg-do compile { target bitint575 } } */
+/* { dg-options "-std=c23" } */
+
+_BitInt(575)
+foo (_BitInt(575) a)
+{
+ a /= 0; /* { dg-warning "division by zero" } */
+ return a;
+}
Jakub
