Fix mode-switching problem for SH SJLJ exceptions

Fix mode-switching problem for SH SJLJ exceptions

[Richard Sandiford]

Functions that return a value usually end with the sequence:

    A: (clobber (reg H))
    ...
    B: (set (reg H) (reg P))
    ...
    C: (use (reg H))

where H is the hard return register and P is the pseudo register
to which return values are assigned.

mode-switching.c uses a specially-split exit block to enforce MODE_EXIT
(the mode required on exit from a function).  The code that creates this
block -- create_pre_exit -- tries to find insn B and starts the new block
immediately before (or sometimes immediately after) it.

However, B can be deleted if P is never initialised.  In this case,
create_pre_exit splits the block immediately before A instead.
The code only searches for A and B in the fallthrough predecessor
to the exit block; it assumes that P is never initialised if the
block contains neither A nor B.

In some cases B may be made up of several individual insns.
create_pre_exit therefore records the range of modified registers
and stops on the first SET that doesn't assign to part of H.
It then asserts that it has found a SET or CLOBBER for the
whole of H (except in some corner cases).

This shows up a problem on SJLJ targets, where the call to the SJLJ
unregister function is inserted between A and B.  If B has been deleted,
the first SET we find may be setting up the argument to that call.
We then stop the search without having found either A or B, so the
assertion triggers.

At this point we have already searched past the call itself, and this
seems bogus.  We are supposed to enforce MODE_EXIT _after_ the last call.
I therefore just treated searches that reach a call in the same way as
searches that reach the beginning of a block.

Bootstrapped & regression-tested on x86_64-linux-gnu.  Also regression-tested
on sh-elf (multilibs {,-m2,-m3,-m4,-m4/-ml}) and sh-wrs-vxworks.  OK to
install?

Richard


gcc/
	* mode-switching.c (create_pre_exit): Don't search past calls.

Index: gcc/mode-switching.c
===================================================================
--- gcc/mode-switching.c	(revision 126715)
+++ gcc/mode-switching.c	(working copy)
@@ -246,6 +246,17 @@ create_pre_exit (int n_entities, int *en
 
 		if (INSN_P (return_copy))
 		  {
+		    /* When using SJLJ exceptions, the call to the
+		       unregister function is inserted between the
+		       clobber of the return value and the copy.
+		       We do not want to split the block before this
+		       or any other call; if we have not found the
+		       copy yet, the copy must have been deleted.  */
+		    if (CALL_P (return_copy))
+		      {
+			short_block = 1;
+			break;
+		      }
 		    return_copy_pat = PATTERN (return_copy);
 		    switch (GET_CODE (return_copy_pat))
 		      {

Re: Fix mode-switching problem for SH SJLJ exceptions

[Richard Sandiford]

Ping

Richard Sandiford <richard@codesourcery.com> writes:
> Functions that return a value usually end with the sequence:
>
>     A: (clobber (reg H))
>     ...
>     B: (set (reg H) (reg P))
>     ...
>     C: (use (reg H))
>
> where H is the hard return register and P is the pseudo register
> to which return values are assigned.
>
> mode-switching.c uses a specially-split exit block to enforce MODE_EXIT
> (the mode required on exit from a function).  The code that creates this
> block -- create_pre_exit -- tries to find insn B and starts the new block
> immediately before (or sometimes immediately after) it.
>
> However, B can be deleted if P is never initialised.  In this case,
> create_pre_exit splits the block immediately before A instead.
> The code only searches for A and B in the fallthrough predecessor
> to the exit block; it assumes that P is never initialised if the
> block contains neither A nor B.
>
> In some cases B may be made up of several individual insns.
> create_pre_exit therefore records the range of modified registers
> and stops on the first SET that doesn't assign to part of H.
> It then asserts that it has found a SET or CLOBBER for the
> whole of H (except in some corner cases).
>
> This shows up a problem on SJLJ targets, where the call to the SJLJ
> unregister function is inserted between A and B.  If B has been deleted,
> the first SET we find may be setting up the argument to that call.
> We then stop the search without having found either A or B, so the
> assertion triggers.
>
> At this point we have already searched past the call itself, and this
> seems bogus.  We are supposed to enforce MODE_EXIT _after_ the last call.
> I therefore just treated searches that reach a call in the same way as
> searches that reach the beginning of a block.
>
> Bootstrapped & regression-tested on x86_64-linux-gnu.  Also regression-tested
> on sh-elf (multilibs {,-m2,-m3,-m4,-m4/-ml}) and sh-wrs-vxworks.  OK to
> install?
>
> Richard
>
>
> gcc/
> 	* mode-switching.c (create_pre_exit): Don't search past calls.
>
> Index: gcc/mode-switching.c
> ===================================================================
> --- gcc/mode-switching.c	(revision 126715)
> +++ gcc/mode-switching.c	(working copy)
> @@ -246,6 +246,17 @@ create_pre_exit (int n_entities, int *en
>  
>  		if (INSN_P (return_copy))
>  		  {
> +		    /* When using SJLJ exceptions, the call to the
> +		       unregister function is inserted between the
> +		       clobber of the return value and the copy.
> +		       We do not want to split the block before this
> +		       or any other call; if we have not found the
> +		       copy yet, the copy must have been deleted.  */
> +		    if (CALL_P (return_copy))
> +		      {
> +			short_block = 1;
> +			break;
> +		      }
>  		    return_copy_pat = PATTERN (return_copy);
>  		    switch (GET_CODE (return_copy_pat))
>  		      {

Re: Fix mode-switching problem for SH SJLJ exceptions

[Ian Lance Taylor]

Richard Sandiford <richard@codesourcery.com> writes:

> gcc/
> 	* mode-switching.c (create_pre_exit): Don't search past calls.

This is OK.

Thanks.

Ian