Re: Problem in hfa_element_mode for IA64

Re: Problem in hfa_element_mode for IA64

[Richard Kenner]

    > But Arr1 is SImode since layout_type sets its mode to an integer mode.

    This seems weird.

Why?  We do it for records.

Suppose we have an array of two chars with alignment of two bytes.  Shouldn't
it be HImode?

Re: Problem in hfa_element_mode for IA64

[Florian Weimer]

kenner@vlsi1.ultra.nyu.edu (Richard Kenner) writes:

>     > But Arr1 is SImode since layout_type sets its mode to an integer mode.
>
>     This seems weird.
>
> Why?  We do it for records.
>
> Suppose we have an array of two chars with alignment of two bytes.  Shouldn't
> it be HImode?

Wouldn't this mean that two consecutive writes to both characters are
combined into a single halfword write?

Re: Problem in hfa_element_mode for IA64

[Richard Kenner]

    > Suppose we have an array of two chars with alignment of two bytes.
    > Shouldn't it be HImode?

    Wouldn't this mean that two consecutive writes to both characters are
    combined into a single halfword write?

I suppose it's related to that, but the major effect is that the array, as a
whole, could live in a register.

Re: Problem in hfa_element_mode for IA64

[Richard Henderson]

On Sun, Mar 31, 2002 at 03:11:37PM -0500, Richard Kenner wrote:
> But Arr1 is SImode since layout_type sets its mode to an integer mode.

This seems weird.


r~

Problem in hfa_element_mode for IA64

[Richard Kenner]

Consider the following tiny Ada program (this can't be reproduced in C
since we don't have arguments of ARRAY_TYPE):

procedure Tbg0331 is
   type Arr1 is array (1..1) of Float;
   procedure Foo (X: Arr1);
   pragma Import (Ada, Foo);
   X: Arr1;
begin
   Foo (X);
end Tbg0331;

This blows up because hfa_element_mdoe when applied to type Arr1 says
it's SFmode since it's an array of SFmode.

But Arr1 is SImode since layout_type sets its mode to an integer mode.

So we try to call emit_move_insn with SFmode output and SImode input
and that aborts.

Who's wrong here?