https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115161
--- Comment #23 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
(In reply to Sergei Trofimovich from comment #22)
> Here `pcmpeqd %xmm2,%xmm1` is a problematic instruction. Why does `gcc` use
> `%xmm2` (result of `cvttps2dq`) instead of, say `%xmm0` which contains
> `0xFFFFffff` pattern?
%xmm0 contains the first all-ones argument to assert_eq. The comparison needs
to happen against a register that contains 0x7fff_ffff (x4), and GCC "thinks"
that %xmm2 contains that value. At -O1, the optimization is done by the RTL CSE
pass (you can confirm by switching it off with -fdisable-rtl-cse1 along -O1),
it sees
(insn 6 5 7 2 (set (reg:V4SF 97)
(mem/u/c:V4SF (symbol_ref/u:DI ("*.LC1") [flags 0x2]) [0 S16 A128]))
"/usr/lib/gcc/x86_64-pc-linux-gnu/10.2.0/include/xmmintrin.h":386:19 1335
{movv4sf_internal}
(expr_list:REG_EQUAL (const_vector:V4SF [
(const_double:SF 4.294967296e+9 [0x0.8p+33]) repeated x4
])
(nil)))
(insn 10 9 11 2 (set (reg:V4SI 98)
(fix:V4SI (reg:V4SF 97)))
"/usr/lib/gcc/x86_64-pc-linux-gnu/10.2.0/include/emmintrin.h":832:19 2635
{fix_truncv4sfv4si2}
(expr_list:REG_EQUAL (const_vector:V4SI [
(const_int 2147483647 [0x7fffffff]) repeated x4
])
(expr_list:REG_DEAD (reg:V4SF 99)
(nil))))
Note the incorrect REG_EQUAL caused by inappropriate use of fix:V4SI where the
source code had cvttps2dq.
