Hi! builtin_memref::extend_offset_range when it sees a NOP_EXPR from INTEGRAL_TYPE (to INTEGRAL_TYPE of sizetype/ptrdifftype precision given the callers) uses wi::to_offset on TYPE_{MIN,MAX}_VALUE of the rhs1 type. This ICEs with large BITINT_TYPEs - to_offset is only supported for precisions up to the offset_int precision - but it even doesn't make any sense to do such thing for narrowing conversions, their range means the whole sizetype/ptrdifftype range and so the normal handling done later on (largest sized supported object) is the way to go in that case.
So, the following patch just restrict this to non-narrowing conversions. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2024-01-19 Jakub Jelinek <ja...@redhat.com> PR tree-optimization/113463 * gimple-ssa-warn-restrict.cc (builtin_memref::extend_offset_range): Only look through NOP_EXPRs if rhs1 doesn't have wider type than lhs. * gcc.dg/bitint-74.c: New test. --- gcc/gimple-ssa-warn-restrict.cc.jj 2024-01-03 11:51:27.705784291 +0100 +++ gcc/gimple-ssa-warn-restrict.cc 2024-01-18 16:00:02.519483821 +0100 @@ -391,7 +391,8 @@ builtin_memref::extend_offset_range (tre tree type; if (is_gimple_assign (stmt) && (type = TREE_TYPE (gimple_assign_rhs1 (stmt))) - && INTEGRAL_TYPE_P (type)) + && INTEGRAL_TYPE_P (type) + && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (offset))) { tree_code code = gimple_assign_rhs_code (stmt); if (code == NOP_EXPR) --- gcc/testsuite/gcc.dg/bitint-74.c.jj 2024-01-18 16:14:05.523599054 +0100 +++ gcc/testsuite/gcc.dg/bitint-74.c 2024-01-18 16:13:30.150099638 +0100 @@ -0,0 +1,16 @@ +/* PR tree-optimization/113463 */ +/* { dg-do compile { target bitint } } */ +/* { dg-options "-std=c23 -O2" } */ + +extern char *a, *b; +#if __BITINT_MAXWIDTH__ >= 129 +_BitInt(129) o; +#else +_BitInt(63) o; +#endif + +void +foo (void) +{ + __builtin_memcpy (a + o, b, 4); +} Jakub