[fortran,patch] Getting rid of an ICE on ISNAN

[fortran,patch] Getting rid of an ICE on ISNAN

[François-Xavier Coudert]

[-- Attachment #1: Type: text/plain, Size: 556 bytes --]

When generating code for ISNAN(real-constant), we call
build_call_expr() and get back an expression inside a NOP_EXPR, which
is not expected, and leads to an ICE further in the middle-end. Using
STRIP_TYPE_NOPS gets rid of the problem, even though it might be
better to fix it in the middle-end. Andrew, Richard, you had ideas
about this, do you want me to keep the PR open after this workaround
is committed, and change its component to "middle-end"?

Bootstrapped and regtested, with type checking, on x86_64-linux. OK to commit?
FX


:ADDPATCH fortran:

[-- Attachment #2: isnan.ChangeLog --]
[-- Type: application/octet-stream, Size: 302 bytes --]

2007-10-29  Francois-Xavier Coudert  <fxcoudert@gcc.gnu.org>

	PR fortran/33596
	* trans-intrinsic.c (gfc_conv_intrinsic_isnan): Strip NOP_EXPR
	from the result of build_call_expr.


2007-10-29  Francois-Xavier Coudert  <fxcoudert@gcc.gnu.org>

	PR fortran/33596
	* gfortran.dg/isnan_2.f90: New test.


[-- Attachment #3: isnan.diff --]
[-- Type: application/octet-stream, Size: 1171 bytes --]

Index: gcc/fortran/trans-intrinsic.c
===================================================================
--- gcc/fortran/trans-intrinsic.c	(revision 129647)
+++ gcc/fortran/trans-intrinsic.c	(working copy)
@@ -2757,6 +2757,7 @@ gfc_conv_intrinsic_isnan (gfc_se * se, g
 
   gfc_conv_intrinsic_function_args (se, expr, &arg, 1);
   se->expr = build_call_expr (built_in_decls[BUILT_IN_ISNAN], 1, arg);
+  STRIP_TYPE_NOPS (se->expr);
   se->expr = fold_convert (gfc_typenode_for_spec (&expr->ts), se->expr);
 }
 
Index: gcc/testsuite/gfortran.dg/isnan_2.f90
===================================================================
--- gcc/testsuite/gfortran.dg/isnan_2.f90	(revision 0)
+++ gcc/testsuite/gfortran.dg/isnan_2.f90	(revision 0)
@@ -0,0 +1,17 @@
+! Test for the ISNAN intrinsic on constants
+!
+! { dg-do run }
+! { dg-options "-fno-range-check" }
+! { dg-options "-fno-range-check -pedantic-errors -mieee" { target sh*-*-* } }
+!
+  implicit none
+  character(len=1) :: s
+  write(s,'(L1)') isnan(0.)
+  if (s /= 'F') call abort
+
+  write(s,'(L1)') isnan(exp(huge(0.)))
+  if (s /= 'F') call abort
+
+  write(s,'(L1)') isnan(0./0.)
+  if (s /= 'T') call abort
+end

Re: [fortran,patch] Getting rid of an ICE on ISNAN

[Richard Guenther]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 759 bytes --]

On Mon, 29 Oct 2007, François-Xavier Coudert wrote:

> When generating code for ISNAN(real-constant), we call
> build_call_expr() and get back an expression inside a NOP_EXPR, which
> is not expected, and leads to an ICE further in the middle-end. Using
> STRIP_TYPE_NOPS gets rid of the problem, even though it might be
> better to fix it in the middle-end. Andrew, Richard, you had ideas
> about this, do you want me to keep the PR open after this workaround
> is committed, and change its component to "middle-end"?
> 
> Bootstrapped and regtested, with type checking, on x86_64-linux. OK to commit?
> FX

I think this is ok, even from a middle-end point of view (that is, until
we disentangle middle-end fold from frontend fold).

Richard.

Re: [fortran,patch] Getting rid of an ICE on ISNAN

[FX Coudert]

> I think this is ok, even from a middle-end point of view (that is,  
> until
> we disentangle middle-end fold from frontend fold).

Given Richard's opinion and the fact that it's a one-liner, I've  
committed this patch as rev. 129782.

FX