Re: debian-i386 target (Was: ppc64le added to buildbot)

[Mark Wielaard <mark@klomp.org>]

Hi Philip,

On Wed, Jun 02, 2021 at 11:22:43AM +0100, Philip Herron wrote:
> > Analyzing compilation unit
> > /home/mark/gccrs/gcc/testsuite/rust.test/compile/methods1.rs: In function
> > ‘Rectangle_sum_x’:
> > /home/mark/gccrs/gcc/testsuite/rust.test/compile/methods1.rs:26:5: error:
> > type mismatch in binary expression
> >    26 |     fn sum_x(self) -> f64 {
> >       |     ^
> > f64
> >
> > <float:80>
> >
> > <float:80>
> >
> > D.227 = _2 + _4;
> > /home/mark/gccrs/gcc/testsuite/rust.test/compile/methods1.rs:26:5:
> > internal compiler error: ‘verify_gimple’ failed
> > 0x8b114e4 verify_gimple_in_seq(gimple*)
> >         ../../gccrs/gcc/tree-cfg.c:5157
> > 0x881a9e3 gimplify_body(tree_node*, bool)
> >         ../../gccrs/gcc/gimplify.c:15401
> > 0x881abb2 gimplify_function_tree(tree_node*)
> >         ../../gccrs/gcc/gimplify.c:15472
> > 0x8675c88 cgraph_node::analyze()
> >         ../../gccrs/gcc/cgraphunit.c:670
> > 0x8678a58 analyze_functions
> >         ../../gccrs/gcc/cgraphunit.c:1236
> > 0x8679641 symbol_table::finalize_compilation_unit()
> >         ../../gccrs/gcc/cgraphunit.c:2514
> 
> This is a good find I will open an issue to track this, I have a local
> Debian 32 bit machine at home so I can try this on later in the week. Off
> the top of my head I am wondering if there is something wrong with how we
> are building up the integer tree nodes. The relevant code starts in
> gcc/rust/backend/rust-compile-tyty.h inside ```  void visit (TyTy::IntType
> &type) override```.
> 
> Let me know if you get anywhere with this before Friday since I will should
> be able to look into this issue then.

I looked a bit but am still struggling with how to debug the gimple generation.
The following is a simpler variant that shows the issue:

fn sum(x: f64, y: f64) -> f64
{
  x + y
}

fn main()
{
  let _sum = sum(3.0, 4.0);
}

Interestingly the following, using an mutable temporary variable to hold the result does pass:

fn sum(x: f64, y: f64) -> f64
{
   let mut z:f64 = x;
   z += y;
   z
}

fn main()
{
  let _sum = sum(3.0, 4.0);
}

As far as I can tell the second variant still uses float:80 for the
temporaries, but because z is f64 it gets cast to (f64) when
returning, making the gimple check pass.

So I assumed it was because the arguments got the wrong type somehow.
But the following also doesn't pass:

fn sum(x: f64, y: f64) -> f64
{
   let a:f64 = x;
   let b:f64 = y;
   a + b
}

fn main()
{
  let _sum = sum(3.0, 4.0);
}

So it is as if the plus expression gets its variable types wrong.
I don't know where that is generated though, or why it is using float:80.

Cheers,

Mark