On Tue, Jun 11, 2024 at 01:20:50PM +0200, Richard Biener wrote:
> When the operand is gimplified to an extract of a register or a
> register we have to disallow memory as we otherwise fail to
> gimplify it properly. Instead of
>
> __asm__("" : "=rm" __imag <r>);
>
> we want
>
> __asm__("" : "=rm" D.2772);
> _1 = REALPART_EXPR <r>;
> r = COMPLEX_EXPR <_1, D.2772>;
>
> otherwise SSA rewrite will fail and generate wrong code with 'r'
> left bare in the asm output.
>
> Bootstrap and regtest in progress on x86_64-unknown-linux-gnu.
>
> I've made the testcase hopefully generic enough (the bug used =X
> which I'm not sure is portable - I've used _Complex int so 'r'
> has a chance to work).
> --- a/gcc/gimplify.cc
> +++ b/gcc/gimplify.cc
> @@ -7044,6 +7044,22 @@ gimplify_asm_expr (tree *expr_p, gimple_seq *pre_p,
> gimple_seq *post_p)
> ret = tret;
> }
>
> + /* If the gimplified operand is a register we do not allow memory. */
> + if (allows_mem
> + && (is_gimple_reg (TREE_VALUE (link))
> + || (handled_component_p (TREE_VALUE (link))
> + && is_gimple_reg (TREE_OPERAND (TREE_VALUE (link), 0)))))
> + {
> + if (allows_reg)
> + allows_mem = 0;
> + else
> + {
> + error ("impossible constraint in %<asm%>");
> + error ("non-memory output %d must stay in memory", i);
> + return GS_ERROR;
Does this else part ever trigger or could it be just gcc_assert (allows_reg)?
E.g. C FE build_asm_expr has
/* If the operand is going to end up in memory,
mark it addressable. */
if (!allows_reg && !c_mark_addressable (output))
Or C++ FE finish_asm_stmt:
/* If the operand is going to end up in memory,
mark it addressable. */
if (!allows_reg && !cxx_mark_addressable (*op))
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/pr115426.c
> @@ -0,0 +1,9 @@
> +/* { dg-do compile } */
> +/* { dg-options "-std=gnu11" } */
> +
> +_Complex int fcs()
> +{
> + _Complex int r;
It would be useful to initialize r or at least __real__ r
before the asm as we return it whole and I think the bug
should trigger with that too.
> + __asm__("" : "=rm" (__imag__ r));
> + return r;
> +}
Also, it would be nice to cover also the "=m" case in another
function to make sure that still works.
Jakub
