https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107715
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Because store data races are allowed with -Ofast masked stores are not used so
we instead get
vect__ifc__80.24_114 = VEC_COND_EXPR <mask__58.15_104, vect__45.20_109,
vect__ifc__78.23_113>;
_ifc__80 = _58 ? _45 : _ifc__78;
MEM <vector(8) double> [(double *)vectp_c.25_116] = vect__ifc__80.24_114;
which somehow is later turned into masked stores? In fact we expand from
vect__43.18_107 = MEM <vector(8) double> [(double *)&a + ivtmp.75_134 * 1];
vect__ifc__78.23_113 = MEM <vector(8) double> [(double *)&c + 8B +
ivtmp.75_134 * 1];
_97 = .COND_FMA (mask__58.15_104, vect_pretmp_36.14_102,
vect_pretmp_36.14_102, vect__43.18_107, vect__ifc__78.23_113);
MEM <vector(8) double> [(double *)&c + 8B + ivtmp.75_134 * 1] = _97;
vect__38.29_121 = MEM <vector(8) double> [(double *)&c + ivtmp.75_134 * 1];
vect__39.32_124 = MEM <vector(8) double> [(double *)&e + ivtmp.75_134 * 1];
_98 = vect__35.11_99 >= { 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 };
_100 = .COND_FMA (_98, vect_pretmp_36.14_102, vect__39.32_124,
vect__38.29_121, vect__43.18_107);
MEM <vector(8) double> [(double *)&a + ivtmp.75_134 * 1] = _100;
the vectorizer has optimize_mask_stores () which is supposed to replace
.MASK_STORE with
if (mask != { 0, 0, 0 ... })
<code depending on the mask store>
and thus optimize the mask == 0 case. But that only triggers for .MASK_STORE.
You can see this when you force .MASK_STORE via -O3 -ffast-math (without
-fallow-store-data-races) you get this effect:
.L2:
vcmppd $13, %zmm1, %zmm0, %k1
kortestb %k1, %k1
jne .L33
.L3:
addq $64, %rax
cmpq $255936, %rax
je .L34
.L4:
vmovapd b(%rax), %zmm0
vmovapd d(%rax), %zmm2
vcmppd $1, %zmm1, %zmm0, %k1
kortestb %k1, %k1
je .L2
vmovapd %zmm2, %zmm3
vfmadd213pd a(%rax), %zmm2, %zmm3
vmovupd %zmm3, c+8(%rax){%k1}
vcmppd $13, %zmm1, %zmm0, %k1
kortestb %k1, %k1
je .L3
.p2align 4
.p2align 3
.L33:
vmovapd c(%rax), %zmm3
vfmadd132pd e(%rax), %zmm3, %zmm2
vmovapd %zmm2, a(%rax){%k1}
addq $64, %rax
cmpq $255936, %rax
jne .L4
.L34:
kortestb %k3, %k3
jne .L35
maybe you can benchmark with that? Still it shouldn't be 40 times slower,
but maybe that's the cache effect of using 4 arrays instead of 3? 4
arrays need 1MB while 3 fit into 750kB? L1 is exactly 1MB so we might
run into aliasing issues there with the 4 arrays.
