[Bug target/107920] [13 Regression] ICE in execute_todo, at passes.cc:2140

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107920

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #7)
> Something like this should fix the issue (note setting the location should
> be done also):
> diff --git a/gcc/config/aarch64/aarch64-sve-builtins-base.cc
> b/gcc/config/aarch64/aarch64-sve-builtins-base.cc
> index 6347407555f..d0a0a3a0a46 100644
> --- a/gcc/config/aarch64/aarch64-sve-builtins-base.cc
> +++ b/gcc/config/aarch64/aarch64-sve-builtins-base.cc
> @@ -1232,6 +1232,15 @@ public:
>         tree mem_ref_op = fold_build2 (MEM_REF, access_type, arg1, zero);
>         gimple *mem_ref_stmt
>           = gimple_build_assign (mem_ref_lhs, mem_ref_op);
> +
> +       // Set the location of load
> +       gimple_set_location (mem_ref_stmt, gimple_location (f.call));
> +
> +       // Set the vuse for the load from the call
> +       tree reaching_vuse = gimple_vuse (f.call);
> +       gcc_assert (reaching_vuse && TREE_CODE (reaching_vuse) == SSA_NAME);
> +       gimple_set_vuse (mem_ref_stmt, reaching_vuse);
> +
>         gsi_insert_before (f.gsi, mem_ref_stmt, GSI_SAME_STMT);
> 
>         int source_nelts = TYPE_VECTOR_SUBPARTS (access_type).to_constant ();

All this should build a sequence of the replacement stmts and then use
gsi_replace_with_seq.

Note transfering EH will not work during folding since that would require
splitting the block.  Does the builtin even throw with non-call EH?  Why
does it throw without?