epilogue count exceeds number of nested prologues

Hi all,

While trying to compile mozilla with libffi, the configure phase dies with a message:

Error: Epilogue count of 4294967296 exceeds number of nested prologues
(0)

I see that this message comes from a patch submitted by David in February. Is libffi broken? What is needed to fix it?

The offending assembly code is attached. it's the ".restore sp" near the end that is triggering the error.

thanks
randolph

	.file	"conftest.i"
	.pred.safe_across_calls p1-p5,p16-p63
	.text
	.align 16
	.global main#
	.proc main#
main:
	.prologue 2, 2
	.vframe r2
	mov r2 = r12
	.body
	;;

#APP
.data
.align 8
1: data8 @fptr(ffi_closure_UNIX)

     data8 1b
     data8 0
     data8 bar
     data8 @fptr(baz)
     data8 quux
     .text
     .global foo
     .proc foo
foo:
     addl r16 = @ltoff(1b), gp;;
     ld8  gp = [r16];;
     ld8  r17 = [gp];;
     ld8  r19 = [r17],8;
	 adds r18 = 16,gp;;
     ld8  r20 = [r17];;
     st8  [r18]=r20;
     mov b6 = r19
     br.cond.sptk.many b6
     .endp

#NO_APP

	;;
	mov r8 = r0
	.restore sp
	mov r12 = r2
	br.ret.sptk.many b0
	.endp main#
	.ident	"GCC: (GNU) 3.0.4"

Randolph> Hi all, While trying to compile mozilla with libffi, the Randolph> configure phase dies with a message:

Randolph> Error: Epilogue count of 4294967296 exceeds number of Randolph> nested prologues (0)

Randolph> I see that this message comes from a patch submitted by Randolph> David in February. Is libffi broken? What is needed to fix Randolph> it?

Randolph> The offending assembly code is attached. it's the Randolph> ".restore sp" near the end that is triggering the error.

This is a bug in the assembler that was fixed on Feb 22nd. I think it should be in binutils-2.12, though I didn't not verify.

	--david

Randolph> Hi all, While trying to compile mozilla with libffi, the Randolph> configure phase dies with a message:

Randolph> Error: Epilogue count of 4294967296 exceeds number of Randolph> nested prologues (0)

Randolph> I see that this message comes from a patch submitted by Randolph> David in February. Is libffi broken? What is needed to fix Randolph> it?

I now realize that you must have been using the latest binutils, as the old one silently generated bad unwind info.

As Richard pointed out in his mail, the offending code really is broken. I was planning to update the mozilla libffi anyhow, both to sync up with Hans's recent unwind fixes and to look into whether we could use the assembler directives added for C++ vtables to simplify the generation of static closures. Let me take a look at it.

	--david

Hi Randolph,

Randolph> Hi all,
Randolph> While trying to compile mozilla with libffi, the configure phase Randolph> dies with a message:

Randolph> Error: Epilogue count of 4294967296 exceeds number of nested prologues Randolph> (0)

Randolph> I see that this message comes from a patch submitted by David in Randolph> February. Is libffi broken? What is needed to fix it?

Randolph> The offending assembly code is attached. it's the ".restore sp" Randolph> near the end that is triggering the error.

I'm unable to reproduce this. If I'm understanding you right, the problem occurred during the configure phase of mozilla (not libffi). I just rebuilt Mozilla from the current CVS tree with gcc3.1 pre-release and the current CVS binutils. It built just fine and the resulting binary seems to work flawlessly.

From the assembly code you provided, it looks as if conftest.c tried to call FFI_STATIC_CLOSURE() from within a function, which can't possibly work.

Can you provide a more precise pointer where this conftest.c is being generated?

Thanks,

	--david

  Randolph> .file	"conftest.i"
  Randolph> .pred.safe_across_calls p1-p5,p16-p63
  Randolph> .text
  Randolph> .align 16
  Randolph> .global main#
  Randolph> .proc main#
  Randolph> main:
  Randolph> .prologue 2, 2
  Randolph> .vframe r2
  Randolph> mov r2 = r12
  Randolph> .body
  Randolph> ;;
  Randolph> #APP
  Randolph> .data
  Randolph> .align 8
  Randolph> 1:   data8 @fptr(ffi_closure_UNIX)
  Randolph> data8 1b
  Randolph> data8 0
  Randolph> data8 bar
  Randolph> data8 @fptr(baz)
  Randolph> data8 quux
  Randolph> .text
  Randolph> .global foo
  Randolph> .proc foo
  Randolph> foo:
  Randolph> addl r16 = @ltoff(1b), gp;;
  Randolph> ld8  gp = [r16];;
  Randolph> ld8  r17 = [gp];;
  Randolph> ld8  r19 = [r17],8;
  Randolph> adds r18 = 16,gp;;
  Randolph> ld8  r20 = [r17];;
  Randolph> st8  [r18]=r20;
  Randolph> mov b6 = r19
  Randolph> br.cond.sptk.many b6
  Randolph> .endp

Randolph> #NO_APP
Randolph> ;;
Randolph> mov r8 = r0
Randolph> .restore sp
Randolph> mov r12 = r2
Randolph> br.ret.sptk.many b0
Randolph> .endp main#
Randolph> .ident "GCC: (GNU) 3.0.4"

Randolph> _______________________________________________ Randolph> Linux-IA64 mailing list
Randolph> [email protected]
Randolph> http://lists.linuxia64.org/lists/listinfo/linux-ia64

Randolph> indeed, mozilla's configure.in says:

Randolph> if test "$LIBFFI_FOUND" = "yes"; then Randolph> AC_MSG_CHECKING("for FFI_STATIC_CLOSURE in libffi") Randolph> AC_TRY_COMPILE([#include "ffi.h"], Randolph> [ FFI_STATIC_CLOSURE(foo, bar, baz, quux); ], Randolph> [ LIBFFI_HAS_STATIC_CLOSURE=yes ]) Randolph> AC_MSG_RESULT("$result")
Randolph> fi

Randolph> which generates (from config.log):

Randolph> configure:13494: checking for FFI_STATIC_CLOSURE in libffi
Randolph> configure:13503: gcc-3.0 -c -I/home/randolph/debian/mozilla-0.9.9/build-tree/libffi-install/include -fshort-wchar -pthread -pipe -I/usr/X11R6/include conftest.c 1>&5
Randolph> {standard input}: Assembler messages:
Randolph> {standard input}:40: Error: Epilogue count of 4294967296 exceeds number of nested prologues (0) Randolph> configure: failed program was:
Randolph> #line 13496 "configure"
Randolph> #include "confdefs.h"
Randolph> #include "ffi.h"
Randolph> int main() {
Randolph> FFI_STATIC_CLOSURE(foo, bar, baz, quux);
Randolph> ; return 0; }

For some reason, I don't see this in my configure.in. Was this added by the Debian patch? In principle, it's a good idea to check for this macro, but I the check needs to be different. To get it to compile, FFI_STATIC_CLOSURE() would have to be declared at the outermost level. Alternatively, perhaps you can just grep for this macro in ffi.h (I think autoconf has support for doing this).

	--david

PS: I dropped binutils from the cc list, as it's clearly not a binutils
issue.

that's right

I'm building from Debian sources. It's possible that it's simply out of date (I'm building version 0.9.9).

indeed, mozilla's configure.in says:

if test "$LIBFFI_FOUND" = "yes"; then
AC_MSG_CHECKING("for FFI_STATIC_CLOSURE in libffi") AC_TRY_COMPILE([#include "ffi.h"],
[ FFI_STATIC_CLOSURE(foo, bar, baz, quux); ], [ LIBFFI_HAS_STATIC_CLOSURE=yes ])
AC_MSG_RESULT("$result")
fi

which generates (from config.log):

configure:13494: checking for FFI_STATIC_CLOSURE in libffi
configure:13503: gcc-3.0 -c -I/home/randolph/debian/mozilla-0.9.9/build-tree/libffi-install/include -fshort-wchar -pthread -pipe -I/usr/X11R6/include conftest.c 1>&5
{standard input}: Assembler messages:
{standard input}:40: Error: Epilogue count of 4294967296 exceeds number of nested prologues (0)
configure: failed program was:
#line 13496 "configure"
#include "confdefs.h"
#include "ffi.h"
int main() {
FFI_STATIC_CLOSURE(foo, bar, baz, quux);
; return 0; }

randolph

<sigh> you are right, i will go poke the Debian maintainer about this now. Sorry for the diversion.

randolph

I claim responsibility for this code. For some reason, it worked in older revisions.

But yes, I can see how that's incorrect, and should be changed. I can whip up a patch if no one else is tomorrow.