On Mon, Nov 6, 2023 at 2:02 PM Maxim Blinov <maxim.a.bli...@gmail.com> wrote:
>
> From: Maxim Blinov <maxim.bli...@imgtec.com>
>
> This patch is based on and intended for the
> vendors/riscv/gcc-13-with-riscv-opts branch - please apply if looks OK.
>
> Fixes the following ICEs that I'm seeing:
>
> FAIL: gcc.dg/vect/O3-pr49087.c (internal compiler error: in
> vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/no-scevccp-pr86725-1.c (internal compiler error: in
> vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/no-scevccp-pr86725-2.c (internal compiler error: in
> vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/no-scevccp-pr86725-3.c (internal compiler error: in
> vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/no-scevccp-pr86725-4.c (internal compiler error: in
> vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/pr94443.c (internal compiler error: in
> vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/pr94443.c -flto -ffat-lto-objects (internal compiler error:
> in vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/slp-50.c (internal compiler error: in vect_transform_loops,
> at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/slp-50.c -flto -ffat-lto-objects (internal compiler error:
> in vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/vect-cond-13.c (internal compiler error: in
> vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/vect-cond-13.c -flto -ffat-lto-objects (internal compiler
> error: in vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/vect-live-6.c (internal compiler error: in
> vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.dg/vect/vect-live-6.c -flto -ffat-lto-objects (internal compiler
> error: in vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.target/riscv/rvv/autovec/partial/live-1.c (internal compiler error:
> in vect_transform_loops, at tree-vectorizer.cc:1032)
> FAIL: gcc.target/riscv/rvv/autovec/partial/live-2.c (internal compiler error:
> in vect_transform_loops, at tree-vectorizer.cc:1032)
>
> -- >8 --
>
> When we create a VEC_EXPAND gimple stmt:
>
> /* SCALAR_RES = VEC_EXTRACT <VEC_LHS, LEN + BIAS - 1>. */
> tree scalar_res
> = gimple_build (&stmts, CFN_VEC_EXTRACT, TREE_TYPE (vectype),
> vec_lhs_phi, last_index);
>
> Under the hood we are really just creating a GIMPLE_CALL stmt. Later
> on, when we `gsi_insert_seq_before` our stmts:
>
> if (stmts)
> {
> gimple_stmt_iterator exit_gsi = gsi_after_labels (exit_bb);
> gsi_insert_seq_before (&exit_gsi, stmts, GSI_SAME_STMT);
>
> We eventually run into tree-ssa-operands.cc:1147:
>
> operands_scanner (fn, stmt).build_ssa_operands ();
>
> Since VEC_EXPAND is *not* marked with ECF_NOVOPS, ECF_CONST, or
> ECF_PURE flags in internal-fn.def, when
I see
DEF_INTERNAL_OPTAB_FN (VEC_EXTRACT, ECF_CONST | ECF_NOTHROW,
vec_extract, vec_extract)
?
> `operand_scanner::parse_ssa_operands` comes across our
> VEC_EXTRACT-type GIMPLE_CALL, it generates a `gimple_vop()` artificial
> variable.
>
> `operand_scanner::finalize_ssa_defs` then picks this up, so our final
> stmt goes from
>
> _73 = .VEC_EXTRACT (vect_last_9.56_71, _72);
>
> to
>
> # .MEM = VDEF <>
> _73 = .VEC_EXTRACT (vect_last_9.56_71, _72);
>
> But more importantly it marks us as `ssa_renaming_needed`, in
> tree-ssa-operands.cc:420:
>
> /* If we have a non-SSA_NAME VDEF, mark it for renaming. */
> if (gimple_vdef (stmt)
> && TREE_CODE (gimple_vdef (stmt)) != SSA_NAME)
> {
> fn->gimple_df->rename_vops = 1;
> fn->gimple_df->ssa_renaming_needed = 1;
> }
>
> This then proceeds to crash the compiler when we are about to leave
> `vect_transform_loops`:
>
> if (need_ssa_update_p (cfun))
> {
> gcc_assert (loop_vinfo->any_known_not_updated_vssa);
> fun->gimple_df->ssa_renaming_needed = false;
> todo |= TODO_update_ssa_only_virtuals;
> }
>
> Since,
>
> - `need_ssa_update_p (cfun)` is true (it was set when we generated a
> memory vdef)
> - `loop_vinfo->any_known_not_updated_vssa` is false
>
> As the code currently stands, creating a gimple stmt containing a
> VEC_EXTRACT should always generate a memory vdef, therefore we should
> remember to mark `loop_vinfo->any_known_not_updated_vssa` afterwards.
>
> gcc/ChangeLog:
>
> * tree-vect-loop.cc (vectorizable_live_operation): Remember to
> assert loop_vinfo->any_known_not_updated_vssa if we are inserting
> a call to VEC_EXPAND.
> ---
> gcc/tree-vect-loop.cc | 5 +++++
> 1 file changed, 5 insertions(+)
>
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index c8df2c88575..53c3a31d2a8 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -10155,6 +10155,11 @@ vectorizable_live_operation (vec_info *vinfo,
> = gimple_build (&stmts, CFN_VEC_EXTRACT, TREE_TYPE (vectype),
> vec_lhs_phi, last_index);
>
> + /* We've expanded SSA at this point, and since VEC_EXTRACT
> + will generate vops, make sure to tell GCC that we need to
> + update SSA. */
> + loop_vinfo->any_known_not_updated_vssa = true;
> +
> /* Convert the extracted vector element to the scalar type. */
> new_tree = gimple_convert (&stmts, lhs_type, scalar_res);
> }
> --
> 2.34.1
>
