[rs6000] Do not generate sibling call to nested function

[rs6000] Do not generate sibling call to nested function

[Eric Botcazou]

[-- Attachment #1: Type: text/plain, Size: 481 bytes --]

Hi,

more precisely, to a nested function that requires a static chain.  The reason 
is that the sibling call epilogue may clobber the static chain register r11.

Tested on PowerPC/Linux, OK for the mainline?


2018-09-17  Eric Botcazou  <ebotcazou@adacore.com>

	* config/rs6000/rs6000.c (rs6000_function_ok_for_sibcall): Return false
	if the call takes a static chain.


2018-09-17  Eric Botcazou  <ebotcazou@adacore.com>

	* gcc.dg/nested-func-11.c: New test.

-- 
Eric Botcazou

[-- Attachment #2: nested-func-11.c --]
[-- Type: text/x-csrc, Size: 443 bytes --]

/* { dg-do run } */
/* { dg-options "-O2 -fno-omit-frame-pointer" } */

int __attribute__((noipa)) foo (int i)
{
  int a;

  void __attribute__((noipa)) nested2 (int i)
  {
    a = i;
  }

  void  __attribute__((noipa)) nested1 (int i)
  {
    int b[32];

    for (int j = 0; j < 32; j++)
      b[j] = i + j;

    nested2 (b[i]);
  }

  nested1 (i);

  return a;
}

int main (void)
{
  if (foo (4) != 8)
    __builtin_abort ();

  return 0;
}

[-- Attachment #3: p.diff --]
[-- Type: text/x-patch, Size: 453 bytes --]

Index: config/rs6000/rs6000.c
===================================================================
--- config/rs6000/rs6000.c	(revision 264342)
+++ config/rs6000/rs6000.c	(working copy)
@@ -24332,6 +24332,10 @@ rs6000_function_ok_for_sibcall (tree dec
 {
   tree fntype;
 
+  /* The sibcall epilogue may clobber the static chain register.  */
+  if (CALL_EXPR_STATIC_CHAIN (exp))
+    return false;
+
   if (decl)
     fntype = TREE_TYPE (decl);
   else

Re: [rs6000] Do not generate sibling call to nested function

[Segher Boessenkool]

Hi!

On Mon, Sep 17, 2018 at 11:11:53AM +0200, Eric Botcazou wrote:
> more precisely, to a nested function that requires a static chain.  The reason 
> is that the sibling call epilogue may clobber the static chain register r11.
> 
> Tested on PowerPC/Linux, OK for the mainline?
> 
> 
> 2018-09-17  Eric Botcazou  <ebotcazou@adacore.com>
> 
> 	* config/rs6000/rs6000.c (rs6000_function_ok_for_sibcall): Return false
> 	if the call takes a static chain.


> --- config/rs6000/rs6000.c	(revision 264342)
> +++ config/rs6000/rs6000.c	(working copy)
> @@ -24332,6 +24332,10 @@ rs6000_function_ok_for_sibcall (tree dec
>  {
>    tree fntype;
>  
> +  /* The sibcall epilogue may clobber the static chain register.  */
> +  if (CALL_EXPR_STATIC_CHAIN (exp))
> +    return false;
> +
>    if (decl)
>      fntype = TREE_TYPE (decl);
>    else

We could probably make sibcalls work with static chain, but that is most
likely not worth it.  Could you change the comment to say something like
that?

Approved for trunk and backports.  Thanks!


Segher

Re: [rs6000] Do not generate sibling call to nested function

[Kyrill Tkachov]

Hi Eric,

On 18/09/18 15:56, Segher Boessenkool wrote:
> Hi!
>
> On Mon, Sep 17, 2018 at 11:11:53AM +0200, Eric Botcazou wrote:
> > more precisely, to a nested function that requires a static chain.  The reason
> > is that the sibling call epilogue may clobber the static chain register r11.
> >
> > Tested on PowerPC/Linux, OK for the mainline?
> >
> >
> > 2018-09-17  Eric Botcazou <ebotcazou@adacore.com>
> >
> >        * config/rs6000/rs6000.c (rs6000_function_ok_for_sibcall): Return false
> >        if the call takes a static chain.
>
>
> > --- config/rs6000/rs6000.c    (revision 264342)
> > +++ config/rs6000/rs6000.c    (working copy)
> > @@ -24332,6 +24332,10 @@ rs6000_function_ok_for_sibcall (tree dec
> >  {
> >    tree fntype;
> >
> > +  /* The sibcall epilogue may clobber the static chain register.  */
> > +  if (CALL_EXPR_STATIC_CHAIN (exp))
> > +    return false;
> > +
> >    if (decl)
> >      fntype = TREE_TYPE (decl);
> >    else
>
> We could probably make sibcalls work with static chain, but that is most
> likely not worth it.  Could you change the comment to say something like
> that?
>
> Approved for trunk and backports.  Thanks!
>

I'm seeing a testism on the GCC 7 branch due to the noipa attribute in the testcase.
I think it doesn't exist in GCC 7 so we're getting a test for excess errors FAIL:

warning: 'noipa' attribute directive ignored [-Wattributes]

Any chance we can adjust the testcase?

Thanks,
Kyrill

>
> Segher

Re: [rs6000] Do not generate sibling call to nested function

[Eric Botcazou]

> I'm seeing a testism on the GCC 7 branch due to the noipa attribute in the
> testcase. I think it doesn't exist in GCC 7 so we're getting a test for
> excess errors FAIL:
> 
> warning: 'noipa' attribute directive ignored [-Wattributes]
> 
> Any chance we can adjust the testcase?

Yes, I did it yesterday.

-- 
Eric Botcazou