http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50485
Bug #: 50485
Summary: gcc.target/i386/sse4_1-blendps.c fails spuriously on
i686
Classification: Unclassified
Product: gcc
Version: 4.7.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: testsuite
AssignedTo: [email protected]
ReportedBy: [email protected]
gcc.target/i386/sse4_1-blendps.c fails spuriously for me on i686. Diego Novillo
mentions something similar in http://gcc.gnu.org/ml/gcc/2011-07/msg00296.html.
The test uses an uninitialized var src3. Initializing part of the var makes the
test fail reliably for me with the same failure mode:
...
Index: sse4_1-blendps.c
===================================================================
--- sse4_1-blendps.c (revision 178880)
+++ sse4_1-blendps.c (working copy)
@@ -64,6 +64,7 @@ TEST (void)
} src3;
int i;
+ src3.f[1] = __builtin_nansf ("");
init_blendps (src1.f, src2.f);
/* Check blendps imm8, m128, xmm */
...
The test aborts because the assignment 'tmp[1] = src2[1]' changes the Nan:
...
static int
check_blendps (__m128 *dst, float *src1, float *src2)
{
float tmp[4];
int j;
memcpy (&tmp[0], src1, sizeof (tmp));
for (j = 0; j < 4; j++)
if ((MASK & (1 << j)))
tmp[j] = src2[j];
return memcmp (dst, &tmp[0], sizeof (tmp));
}
...
The assignment is translated as a push/pop on the float stack:
...
#(insn 17 39 42 2 (set (reg:SF 8 st)
# (mem:SF (plus:SI (reg/v/f:SI 2 cx [orig:65 src2 ] [65])
# (const_int 4 [0x4])) [3 MEM[(float *)src2_10(D) + 4B]+0 S4
A32])) sse4_1-blendps.c:46 108 {*movsf_internal}
# (expr_list:REG_DEAD (reg/v/f:SI 2 cx [orig:65 src2 ] [65])
# (expr_list:REG_EQUIV (mem/s/c:SF (plus:SI (reg/f:SI 20 frame)
# (const_int -12 [0xfffffffffffffff4])) [3 tmp+4 S4 A32])
# (nil))))
flds 4(%ecx) # 17 *movsf_internal/1 [length = 3]
...
#(insn 18 15 21 2 (set (mem/s/c:SF (plus:SI (reg/f:SI 7 sp)
# (const_int 20 [0x14])) [3 tmp+4 S4 A32])
# (reg:SF 8 st)) sse4_1-blendps.c:46 108 {*movsf_internal}
# (expr_list:REG_DEAD (reg:SF 8 st)
# (nil)))
fstps 20(%esp) # 18 *movsf_internal/2 [length = 4]
...
By going through the float stack, the signalling Nan turns into a quiet Nan.
That seems to correspond with what is said at
http://stackoverflow.com/questions/2247447/usefulness-of-signaling-nan .
So after the push/pop tmp[1] contains a quiet Nan, while the corresponding part
of dst contains the bit representation of the signalling Nan, the memcmp
returns != 0 and the test aborts.