Re: target/10348: sparc64-sun-solaris2.7 testsuite failure in execute/20020720-1.c with -fpic/-fPIC

[Roger Sayle <roger@www.eyesopen.com>]

The following reply was made to PR target/10348; it has been noted by GNATS.

From: Roger Sayle <roger@www.eyesopen.com>
To: Kaveh Ghazi <ghazi@caip.rutgers.edu>,
   Eric Botcazou <ebotcazou@libertysurf.fr>
Cc: gcc-gnats@gcc.gnu.org, <gcc-bugs@gcc.gnu.org>
Subject: Re: target/10348: sparc64-sun-solaris2.7 testsuite failure in
 execute/20020720-1.c with -fpic/-fPIC
Date: Thu, 10 Apr 2003 07:11:31 -0600 (MDT)

 On Thu, 10 Apr 2003, Eric Botcazou wrote:
 > > I'm getting testsuite failures in execute/20020720-1.c on
 > > sparc64-sun-solaris2.7 at -O1 and above when adding -fpic or -fPIC to the
 > > testsuite pass.  The logfile shows that the function `link_error' is still
 > > called, i.e. the call is not being optimized away like it should.
 >
 > Note that this happens on sparc-sun-solaris2.9 too with -mcpu=v7|v8|v9.
 
 Unfortunately, there is no simple fix to the problem.  The 64-bit sparc
 PIC targets require three instructions to compare a floating point value
 against zero: the first to store the constant pool address+offset in a
 register, the second to load this into a FP register and a third to do
 the comparison.  Hence comparing "fabs" against zero requires four
 instructions.  This is one too many for combine, where this optimization
 normally gets applied, which only considers up to three instructions.
 
 I would recommend marking 20020720-1.x as XFAIL on sparc with a 64bit
 CPU type and -fpic or -fPIC, and lowering the GNATS PR priority to low.
 However, I believe the issue should be resolved shortly as soon as my
 "REG_EQUAL notes on cond_jumps" patch gets approved.  Its been a month!
 http://gcc.gnu.org/ml/gcc-patches/2003-03/msg00824.html
 
 This patch should allow GCSE to optimize "fabs(x) < 0.0" itself,
 independently of the floating point load and conditional branch
 instructions provided by the target's machine description.  I suspect
 this will fix 20020720-1.c on all GCC targets that don't require FP
 comparisons to be performed by libcalls (but that currently require
 four or more instructions).  I'm also investigating a libcall fix.
 
 Roger
 --