https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111793
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Blocks| |53947
Target| |x86_64-*-*
Keywords| |missed-optimization
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
#pragma omp declare simd simdlen(16) inbranch
__attribute__((noinline)) int
foo (int a, int b)
{
return a + b;
}
we present if-conversion with the following which I think is good:
<bb 3> [local count: 1073741824]:
# iter.49_5 = PHI <0(2), iter.49_6(7)>
# ivtmp_19 = PHI <16(2), ivtmp_18(7)>
b_2 = b.45[iter.49_5];
a_1 = a.44[iter.49_5];
_8 = mask.48_7(D) >> iter.49_5;
_9 = _8 & 1;
if (_9 == 0)
goto <bb 8>; [20.00%]
else
goto <bb 4>; [80.00%]
<bb 8> [local count: 214748360]:
goto <bb 5>; [100.00%]
<bb 4> [local count: 858993464]:
_3 = a_1 + b_2;
retval.43[iter.49_5] = _3;
<bb 5> [local count: 1073741824]:
iter.49_6 = iter.49_5 + 1;
ivtmp_18 = ivtmp_19 - 1;
if (ivtmp_18 != 0)
goto <bb 7>; [100.00%]
else
goto <bb 6>; [0.00%]
and we if-convert the body to
<bb 3> [local count: 1073741824]:
# iter.49_5 = PHI <iter.49_6(7), 0(15)>
# ivtmp_19 = PHI <ivtmp_18(7), 16(15)>
b_2 = b.45[iter.49_5];
a_1 = a.44[iter.49_5];
_8 = mask.48_7(D) >> iter.49_5;
_9 = _8 & 1;
_22 = _9 == 0;
_36 = (unsigned int) a_1;
_37 = (unsigned int) b_2;
_38 = _36 + _37;
_3 = (int) _38;
_39 = ~_22;
_40 = &retval.43[iter.49_5];
.MASK_STORE (_40, 32B, _39, _3);
iter.49_6 = iter.49_5 + 1;
ivtmp_18 = ivtmp_19 - 1;
if (ivtmp_18 != 0)
goto <bb 7>; [100.00%]
else
goto <bb 6>; [0.00%]
but rather than recovering the mask in its original form we vectorize this as
vect_cst__50 = {mask.48_7(D), mask.48_7(D), mask.48_7(D), mask.48_7(D),
mask.48_7(D), mask.48_7(D), mask.48_7(D), mask.48_7(D), mask.48_7(D),
mask.48_7(D), mask.48_7(D), mask.48_7(D), mask.48_7(D), mask.48_7(D),
mask.48_7(D), mask.48_7(D)};
<bb 3> [local count: 10631108]:
# vect_vec_iv_.125_42 = PHI <_43(7), { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
12, 13, 14, 15 }(2)>
# vectp_b.126_44 = PHI <vectp_b.126_45(7), &b.45(2)>
# vectp_a.129_47 = PHI <vectp_a.129_48(7), &a.44(2)>
# vectp_retval.140_61 = PHI <vectp_retval.140_62(7), &retval.43(2)>
# ivtmp_64 = PHI <ivtmp_65(7), 0(2)>
_43 = vect_vec_iv_.125_42 + { 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
16, 16, 16, 16 };
vect_b_2.128_46 = MEM <vector(16) int> [(int *)vectp_b.126_44];
vect_a_1.131_49 = MEM <vector(16) int> [(int *)vectp_a.129_47];
vect__8.132_51 = vect_cst__50 >> vect_vec_iv_.125_42;
vect__9.133_53 = vect__8.132_51 & { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1 };
mask__22.134_55 = vect__9.133_53 == { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0 };
vect__36.135_56 = VIEW_CONVERT_EXPR<vector(16) unsigned
int>(vect_a_1.131_49);
vect__37.136_57 = VIEW_CONVERT_EXPR<vector(16) unsigned
int>(vect_b_2.128_46);
vect__38.137_58 = vect__36.135_56 + vect__37.136_57;
vect__3.138_59 = VIEW_CONVERT_EXPR<vector(16) int>(vect__38.137_58);
mask__39.139_60 = ~mask__22.134_55;
if (mask__39.139_60 == { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 })
goto <bb 20>; [20.00%]
else
goto <bb 21>; [80.00%]
<bb 21> [local count: 8504886]:
.MASK_STORE (vectp_retval.140_61, 512B, mask__39.139_60, vect__3.138_59);
<bb 20> [local count: 10631108]:
vectp_b.126_45 = vectp_b.126_44 + 64;
vectp_a.129_48 = vectp_a.129_47 + 64;
vectp_retval.140_62 = vectp_retval.140_61 + 64;
ivtmp_65 = ivtmp_64 + 1;
if (ivtmp_65 < 1)
goto <bb 7>; [0.00%]
else
goto <bb 17>; [100.00%]
in the end leaving us with
<bb 2> [local count: 1073741824]:
vect_cst__50 = {mask.48_7(D), mask.48_7(D), mask.48_7(D), mask.48_7(D),
mask.48_7(D), mask.48_7(D), mask.48_7(D), mask.48_7(D), mask.48_7(D),
mask.48_7(D), mask.48_7(D), mask.48_7(D), mask.48_7(D), mask.48_7(D),
mask.48_7(D), mask.48_7(D)};
vect__8.132_51 = vect_cst__50 >> { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
13, 14, 15 };
vect__9.133_53 = vect__8.132_51 & { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1 };
vect__36.135_56 = VIEW_CONVERT_EXPR<vector(16) unsigned int>(simd.46_13(D));
vect__37.136_57 = VIEW_CONVERT_EXPR<vector(16) unsigned int>(simd.47_15(D));
vect__38.137_58 = vect__36.135_56 + vect__37.136_57;
vect__3.138_59 = VIEW_CONVERT_EXPR<vector(16) int>(vect__38.137_58);
mask__39.139_60 = vect__9.133_53 != { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0 };
if (mask__39.139_60 == { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 })
goto <bb 4>; [20.00%]
else
goto <bb 3>; [80.00%]
<bb 3> [local count: 858993419]:
.MASK_STORE (&retval.43, 512B, mask__39.139_60, vect__3.138_59);
<bb 4> [local count: 1073741824]:
_10 = VIEW_CONVERT_EXPR<vector(16) int>(retval.43);
return _10;
and
_ZGVeM16vv_foo:
.LFB4:
.cfi_startproc
movl $1, %eax
pushq %rbp
.cfi_def_cfa_offset 16
.cfi_offset 6, -16
vpbroadcastd %edi, %zmm2
vpaddd %zmm1, %zmm0, %zmm0
vpbroadcastd %eax, %zmm3
movq %rsp, %rbp
.cfi_def_cfa_register 6
andq $-64, %rsp
vpsrlvd .LC0(%rip), %zmm2, %zmm2
vpxor %xmm1, %xmm1, %xmm1
vpandd %zmm3, %zmm2, %zmm2
vpcmpd $4, %zmm1, %zmm2, %k1
kortestw %k1, %k1
je .L25
vmovdqa32 %zmm0, -64(%rsp){%k1}
.L25:
vmovdqa32 -64(%rsp), %zmm0
leave
.cfi_def_cfa 7, 8
ret
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations