RISC-V: Expand shift count in Xmode in interleave pattern.

Hi,

currently ssa-dse-1.C ICEs because expand_simple_binop returns NULL
when building the scalar that is used to IOR two interleaving
sequences.

That's because we try to emit a shift in HImode.  This patch shifts in
Xmode and then lowpart-subregs the result to HImode.

Regtested on rv64gcv_zvl512b.

Regards
 Robin

gcc/ChangeLog:

	* config/riscv/riscv-v.cc (expand_const_vector): Shift in Xmode.

diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
index 9926a7b63bb55bda8db538361d31728dd238b0a6..3735a5ba65943e5d39cbda426184ec3c460d8a32 100644 (file)
--- a/gcc/config/riscv/riscv-v.cc
+++ b/gcc/config/riscv/riscv-v.cc
@@ -1505,10 +1505,11 @@ expand_const_vector (rtx target, rtx src)
                {
                  /* { 1, 1, 2, 1, ... }.  */
                  rtx scalar = expand_simple_binop (
-                   new_smode, ASHIFT,
-                   gen_int_mode (rtx_to_poly_int64 (base2), new_smode),
-                   gen_int_mode (builder.inner_bits_size (), new_smode),
+                   Xmode, ASHIFT,
+                   gen_int_mode (rtx_to_poly_int64 (base2), Xmode),
+                   gen_int_mode (builder.inner_bits_size (), Xmode),
                    NULL_RTX, false, OPTAB_DIRECT);
+                 scalar = simplify_gen_subreg (new_smode, scalar, Xmode, 0);
                  rtx tmp2 = gen_reg_rtx (new_mode);
                  rtx ior_ops[] = {tmp2, tmp1, scalar};
                  emit_vlmax_insn (code_for_pred_scalar (IOR, new_mode),