So this is a minor bug in the riscv move expanders. It has a special cases for
extraction from vector objects which makes assumptions that it can use
gen_lowpart unconditionally. That's not always the case.
We can just bypass that special code for cases where we can't use gen_lowpart
and let the more generic code run. If gen_lowpart_common indicates we've got a
case that can't be handled we just bypass the special extraction code.
Tested on riscv64-elf and riscv32-elf. Waiting for pre-commit CI to do its
thing.
PR target/119275
gcc/
* config/riscv/riscv.cc (riscv_legitimize_move): Avoid calling
gen_lowpart for cases where it'll fail. Just use standard expander
paths for those cases.
gcc/testsuite/
* gcc.target/riscv/pr119275.c: New test.
/* This test can fail if (for example) we want a HF and Z[v]fh is
not enabled. In that case we just want to let the standard
expansion path run. */
- if (riscv_vector::get_vector_mode (smode, nunits).exists (&vmode))
+ if (riscv_vector::get_vector_mode (smode, nunits).exists (&vmode)
+ && gen_lowpart_common (vmode, SUBREG_REG (src)))
{
rtx v = gen_lowpart (vmode, SUBREG_REG (src));
rtx int_reg = dest;
--- /dev/null
+/* { dg-do compile { target int128 } } */
+/* { dg-options "-march=rv64gv -mabi=lp64d -mrvv-vector-bits=zvl" { target { rv64 } } } */
+
+__int128 h, j;
+int y;
+double d;
+void *p;
+char *q;
+char x;
+long u;
+
+char *bar(int, int);
+
+typedef __attribute__((__vector_size__ (2))) char V;
+
+void
+foo(V v)
+{
+ x += *bar (0, 0);
+ for(;;) {
+ __builtin_strcat (p, 7 + q);
+ d += __builtin_stdc_rotate_left (
+ (unsigned __int128) u | h << *__builtin_strcat (p, 7 + q), j);
+ u += (long) __builtin_memmove (&y, &v, 2);
+ }
+}