https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80481
Uroš Bizjak <ubizjak at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Target Milestone|--- |10.0
--- Comment #10 from Uroš Bizjak <ubizjak at gmail dot com> ---
Actually, the issue is not fixed. As mentioned in the description, the issue is
best visible without -funrol-loops.
(GCC) 10.0.0 20190821 (experimental) compiles (-Ofast -fopenmp -march=knl) to:
.L31:
leaq (%rdx), %rsi
negq %rsi
vpermps (%r9,%rsi), %zmm8, %zmm0
vmovaps %zmm0, %zmm1
vmaxps (%r11,%rdx), %zmm3, %zmm0
vfnmadd132ps (%r14,%rdx), %zmm7, %zmm1
vmaxps %zmm1, %zmm0, %zmm0
vmovups %zmm0, 0(%r13,%rdx)
leaq 64(%rdx), %rdx
cmpq %r8, %rdx
jne .L31
As seen in the detailed dump,
#(insn:TI 856 852 1743 71 (set (reg:V16SF 20 xmm0 [orig:885 vect__72.36 ]
[885])
# (unspec:V16SF [
# (mem:V16SF (plus:DI (reg/f:DI 37 r9 [orig:198 vectp.34 ]
[198])
# (reg:DI 4 si [883])) [3 MEM[base: vectp.34_256, index:
_1006, offset: 0B]+0 S64 A32])
# (reg:V16SI 44 xmm8 [919])
# ] UNSPEC_VPERMVAR))
"/hdd/uros/git/gcc/gcc/testsuite/g++.dg/pr80481.C":59:28 4754
{avx512f_permvarv16sf}
# (expr_list:REG_DEAD (reg:DI 4 si [883])
# (nil)))
vpermps (%r9,%rsi), %zmm8, %zmm0 # 856 [c=68 l=7]
avx512f_permvarv16sf
#(insn:TI 1743 856 860 71 (set (reg:V16SF 21 xmm1 [orig:885 vect__72.36 ]
[885])
# (reg:V16SF 20 xmm0 [orig:885 vect__72.36 ] [885]))
"/hdd/uros/git/gcc/gcc/testsuite/g++.dg/pr80481.C":60:20 1255
{movv16sf_internal}
# (expr_list:REG_DEAD (reg:V16SF 20 xmm0 [orig:885 vect__72.36 ] [885])
# (nil)))
vmovaps %zmm0, %zmm1 # 1743 [c=4 l=6] movv16sf_internal/2
#(insn 860 1743 857 71 (set (reg:V16SF 20 xmm0 [orig:888 vect__13.45 ] [888])
# (smax:V16SF (reg:V16SF 23 xmm3 [890])
# (mem:V16SF (plus:DI (reg/f:DI 39 r11 [orig:187 vectp.43 ] [187])
# (reg:DI 1 dx [orig:478 ivtmp.111 ] [478])) [3 MEM[base:
vectp.43_238, index: ivtmp.111_997, offset: 0B]+0 S64 A32])))
"/hdd/uros/git/gcc/gcc/testsuite/g++.dg/pr80481.C":61:15 1631 {*smaxv16sf3}
# (nil))
vmaxps (%r11,%rdx), %zmm3, %zmm0 # 860 [c=68 l=6]
*smaxv16sf3/1
%zmm0 is killed in (insn 860) and thus dead in (insn 1743). (insn 856) could
simply use %zmm0 as its destination.
