https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120910
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed| |2025-07-01
Keywords| |wrong-code
Status|UNCONFIRMED |NEW
Ever confirmed|0 |1
--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
testl $2139095040, %edx
je .L10
vmovss .LC6(%rip), %xmm0 //0x0.8p-24
vcvtps2ph $4, %xmm0, %xmm0
cmpl $-1043857408, %ebx
ja .L1
vmovd %ebx, %xmm3
vmovss .LC7(%rip), %xmm0
vcvtps2ph $4, %xmm0, %xmm0
vcomiss .LC2(%rip), %xmm3
ja .L1
ce3 does the move.
IF-CASE-1 found, start 6, then 7
rescanning insn with uid = 37.
changing bb of uid 39
changing bb of uid 107
from 7 to 6
changing bb of uid 42
from 7 to 6
deleting insn with uid = 91.
deleting block 7
Conversion succeeded on pass 1.
Confirmed.
I suspect the problem is
(insn 42 107 36 6 (set (reg:V8HI 20 xmm0 [117])
(vec_concat:V8HI (unspec:V4HI [
(reg:V4SF 20 xmm0 [118])
(const_int 4 [0x4])
] UNSPEC_VCVTPS2PH)
(const_vector:V4HI [
(const_int 0 [0]) repeated x4
]))) "t56.c":15:40 discrim 1 9621 {*vcvtps2ph}
(nil))
is not detected as being able to trap with -ftrapping-math.
Note with -fno-trapping-math, this transformation is correct as trapping in
this case deals with floating exceptions. So clang not producing any underflow
is just pure accident as their setting is off there. And since you are tracking
floating exceptions, you need -ftrapping-math.
