c/5315: Incorrect loads in FP compare code on i686

[pbarada at mail dot wm dot sps dot mot dot com]

>Number:         5315
>Category:       c
>Synopsis:       Incorrect loads in FP compare code on i686
>Confidential:   no
>Severity:       critical
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          wrong-code
>Submitter-Id:   net
>Arrival-Date:   Tue Jan 08 14:06:01 PST 2002
>Closed-Date:
>Last-Modified:
>Originator:     Peter Barada
>Release:        3.0.3 20011213 (prerelease)
>Organization:
Motorola, Inc.
>Environment:
System: Linux hyper.wm.sps.mot.com 2.4.2-2 #1 Sun Apr 8 20:41:30 EDT 2001 i686 unknown
Architecture: i686

	
host: i686-pc-linux-gnu
build: i686-pc-linux-gnu
target: i686-pc-linux-gnu
configured with: /home/pbarada/work/cvs-wavemark/cross-linux-tools/gcc-3.0.3-20011213/configure --target=i686-linux --prefix=/tmp/crap8 --enable-languages=c --with-local-prefix=/tmp/crap8/i686-linux --with-newlib --disable-shared : (reconfigured) /home/pbarada/work/cvs-wavemark/cross-linux-tools/gcc-3.0.3-20011213/configure --with-gcc-version-trigger=/home/pbarada/work/cvs-wavemark/cross-linux-tools/gcc-3.0.3-20011213/gcc/version.c --host=i686-pc-linux-gnu --target=i686-linux --prefix=/tmp/crap8 --enable-languages=c --with-local-prefix=/tmp/crap8/i686-linux --with-newlib --disable-shared
>Description:
Compile the test code with "-S -msoft-float -g -dp" options.

The code when compiled without optimization loads the same value(insn
32, insn 35) onto the stack twice before calling __gtsf2 which
prevents the else clause from ever being taken. code generated looks like:

interln:
.LFB1:
.LM1:
.LBB2:
.LBB3:
	pushl	%ebp	# 86	*pushsi2	[length = 1]
.LCFI0:
	movl	%esp, %ebp	# 88	*movsi_1/2	[length = 2]
.LCFI1:
	subl	$24, %esp	# 90	*pro_epilogue_adjust_stack_1/1	[length = 3]
.LCFI2:
.LM2:
	movl	12(%ebp), %eax	# 12	*movsf_1/4	[length = 3]
	movl	%eax, -8(%ebp)	# 14	*movsf_1/5	[length = 3]
	leal	-8(%ebp), %eax	# 76	*lea_1	[length = 2]
	andl	$2147483647, (%eax)	# 16	*andsi_1/1	[length = 6]
.LBE3:
.LBB4:
	movl	8(%ebp), %eax	# 23	*movsf_1/4	[length = 3]
	movl	%eax, -8(%ebp)	# 25	*movsf_1/5	[length = 3]
	leal	-8(%ebp), %eax	# 78	*lea_1	[length = 2]
	andl	$2147483647, (%eax)	# 27	*andsi_1/1	[length = 6]
.LBE4:
	movl	-8(%ebp), %eax	# 32	*movsf_1/4	[length = 3]
	movl	%eax, (%esp)	# 33	*movsf_1/5	[length = 3]
	movl	-8(%ebp), %eax	# 35	*movsf_1/4	[length = 3]
	movl	%eax, 4(%esp)	# 36	*movsf_1/5	[length = 3]
	call	__gtsf2	# 37	*call_value_0	[length = 6]

Note that insn 32 and 35 load the same value onto the stack.  Backing
up further, notice that insns 76/16 and insns 78/27 store the result of
the absolute value  for each side of the '>' expression into the same
location (-8(%ebp)).  If compiled with '-O', this problem does not
exhibit itself.

>How-To-Repeat:

#define FFABS(x) ({ \
  union {float f; unsigned int l;} z; \
  z.f = (float)(x); \
  z.l &= 0x7fffffff; \
  z.f; \
})

typedef float real;

typedef struct {
  real x,y;
} Point;

extern void foo(real, real);
extern void bar(real, real);

void interln (real bdx, real bdy)
{
  real c2;

  if (FFABS(bdy) > FFABS(bdx)) {
    foo(bdx, bdy);
  }
  else {
    bar(bdy, bdx);
  }
}

>Fix:
	
>Release-Note:
>Audit-Trail:
>Unformatted: