https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68030
Bug ID: 68030
Summary: Redundant address calculations in vectorized loop
Product: gcc
Version: 6.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: kyukhin at gcc dot gnu.org
Target Milestone: ---
Created attachment 36548
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=36548&action=edit
Reproducer
Attached testcase.
Compiled w/: -S -Ofast -march=haswell 1.c
Code in main loop is:
.L7:
movq -176(%rbp), %rdx
vmovaps -272(%rbp), %ymm1
addl $1, %r9d
vmulps (%rdx,%rax), %ymm11, %ymm0
movq -184(%rbp), %rdx ; <- Fill
vfmadd231ps (%rdx,%rax), %ymm12, %ymm0
movq -192(%rbp), %rdx ; <- Fill
vmovaps %ymm0, %ymm15
vmulps (%rdx,%rax), %ymm13, %ymm0
movq -112(%rbp), %rdx; <- Fill
vfmadd231ps (%rdx,%rax), %ymm14, %ymm0
movq -160(%rbp), %rdx; <- Fill
vaddps %ymm0, %ymm15, %ymm0
vmulps (%rdx,%rax), %ymm9, %ymm15
movq -168(%rbp), %rdx
...
Those loads are related to the same array (global_Input in the source)
with fixed offsets from the base.
Unfortunately, this is not recognized by Gimple (optimized):
# ratio_mult_vf.11_1609 = PHI <504(8), 512(5)>
# bnd.10_1618 = PHI <63(8), 64(5)>
# niters.9_1610 = PHI <niters.9_339(8), 512(5)>
# prolog_loop_adjusted_niters.8_1611 = PHI
<prolog_loop_adjusted_niters.8_340(8), 0(5)>
# ivtmp_1615 = PHI <_401(8), 512(5)>
# ix_1592 = PHI <_456(8), 2(5)>
_999 = ivtmp.345_510 + prolog_loop_adjusted_niters.8_1611;
_998 = _999 * 4;
_995 = _998 + 18446744073709547488;
vectp.15_1 = pretmp_889 + _995; <-- addr is (base + _999*4 + OFFSET1)
_986 = *local_Filter_12;
vect_cst__984 = {_986, _986, _986, _986, _986, _986, _986, _986};
_975 = _998 + 18446744073709547492;
vectp.22_982 = pretmp_889 + _975; <-- addr is (base + _999*4 + OFFSET11)
_965 = MEM[(float *)local_Filter_12 + 4B];
vect_cst__964 = {_965, _965, _965, _965, _965, _965, _965, _965};
_956 = _998 + 18446744073709547496;
vectp.30_961 = pretmp_889 + _956; <-- addr is (base + _999*4 + OFFSET2)
_948 = MEM[(float *)local_Filter_12 + 8B];
vect_cst__947 = {_948, _948, _948, _948, _948, _948, _948, _948};
_940 = _998 + 18446744073709547500;
vectp.37_945 = pretmp_889 + _940; <-- addr is (base + _999*4 + OFFSET3)
_932 = MEM[(float *)local_Filter_12 + 12B];
vect_cst__931 = {_932, _932, _932, _932, _932, _932, _932, _932};
_924 = _998 + 18446744073709547504;
vectp.44_929 = pretmp_889 + _924; <-- addr is (base + _999*4 + OFFSET4)
_916 = MEM[(float *)local_Filter_12 + 16B];`
vect_cst__915 = {_916, _916, _916, _916, _916, _916, _916, _916};
_903 = _998 + 18446744073709549552;
vectp.53_911 = pretmp_889 + _903; <-- addr is (base + _999*4 + OFFSET5)
_895 = MEM[(float *)local_Filter_12 + 20B];
vect_cst__894 = {_895, _895, _895, _895, _895, _895, _895, _895};
_155 = _998 + 18446744073709549556;
vectp.60_892 = pretmp_889 + _155;<-- addr is (base + _999*4 + OFFSET6)
_500 = MEM[(float *)local_Filter_12 + 24B];
vect_cst__37 = {_500, _500, _500, _500, _500, _500, _500, _500};
_1070 = _998 + 18446744073709549560;
vectp.68_907 = pretmp_889 + _1070; <-- addr is (base + _999*4 + OFFSET7)
_1078 = MEM[(float *)local_Filter_12 + 28B];
vect_cst__1079 = {_1078, _1078, _1078, _1078, _1078, _1078, _1078, _1078};
_1087 = _998 + 18446744073709549564;
vectp.76_1082 = pretmp_889 + _1087; <-- addr is (base + _999*4 + OFFSET8)
_1095 = MEM[(float *)local_Filter_12 + 32B];
vect_cst__1096 = {_1095, _1095, _1095, _1095, _1095, _1095, _1095, _1095};
_1103 = _998 + 18446744073709549568; <-- addr is (base + _999*4 + OFFSET9)
vectp.83_1098 = pretmp_889 + _1103;
_1111 = MEM[(float *)local_Filter_12 + 36B];
...
<bb 10>:
# ivtmp.250_79 = PHI <ivtmp.250_56(10), 0(9)>
# ivtmp.253_329 = PHI <ivtmp.253_330(10), 0(9)>
vect__161.16_992 = MEM[base: vectp.15_1, index: ivtmp.253_329, offset: 0B];
// load @ (base + _999*4 + OFFSET1 + IV + 0)
vect__177.23_972 = MEM[base: vectp.22_982, index: ivtmp.253_329, offset: 0B];
// load @ (base + _999*4 + OFFSET11 + IV + 0)
vect__182.27_963 = vect_cst__964 * vect__177.23_972;
_1256 = vect_cst__984 * vect__161.16_992 + vect__182.27_963;
vect__193.31_953 = MEM[base: vectp.30_961, index: ivtmp.253_329, offset: 0B];
// load @ (base + _999*4 + OFFSET2 + IV + 0)
vect__209.38_937 = MEM[base: vectp.37_945, index: ivtmp.253_329, offset: 0B];
// load @ (base + _999*4 + OFFSET3 + IV + 0)
vect__214.42_930 = vect_cst__931 * vect__209.38_937;
_1235 = vect_cst__947 * vect__193.31_953 + vect__214.42_930;
_1307 = _1235 + _1256;
vect__225.45_921 = MEM[base: vectp.44_929, index: ivtmp.253_329, offset: 0B];
// load @ (base + _999*4 + OFFSET4 + IV + 0)
vect__247.54_900 = MEM[base: vectp.53_911, index: ivtmp.253_329, offset: 0B];
// load @ (base + _999*4 + OFFSET5 + IV + 0)
vect__252.58_893 = vect_cst__894 * vect__247.54_900;
_1291 = vect_cst__915 * vect__225.45_921 + vect__252.58_893;
_341 = _1291 + _1307;
vect__263.61_242 = MEM[base: vectp.60_892, index: ivtmp.253_329, offset: 0B];
// load @ (base + _999*4 + OFFSET6 + IV + 0)
vect__279.69_1073 = MEM[base: vectp.68_907, index: ivtmp.253_329, offset:
0B]; // load @ (base + _999*4 + OFFSET7 + IV + 0)
...
You can see all loads are differ in a constant.
All vector loads mentions may use common base for addres + constant offset.
Which pass is responsible for such optimization?
I'd like to see something like this:
_999 = ivtmp.345_510 + prolog_loop_adjusted_niters.8_1611;
_998 = _999 * 4;
vectp.15_1 = pretmp_889 + _998; <-- addr is (base + _999*4)
...
<bb 10>:
# ivtmp.250_79 = PHI <ivtmp.250_56(10), 0(9)>
# ivtmp.253_329 = PHI <ivtmp.253_330(10), 0(9)>
vect__161.16_992 = MEM[base: vectp.15_1, index: ivtmp.253_329, offset:
OFFSET1]; // load @ (base + _999*4 + IV + OFFSET1)
vect__177.23_972 = MEM[base: vectp.15_1, index: ivtmp.253_329, offset:
OFFSET11]; // load @ (base + _999*4 + IV + OFFSET11)
vect__182.27_963 = vect_cst__964 * vect__177.23_972;
_1256 = vect_cst__984 * vect__161.16_992 + vect__182.27_963;
vect__193.31_953 = MEM[base: vectp.15_1, index: ivtmp.253_329, offset:
OFFSET2]; // load @ (base + _999*4 + IV + OFFSET2)
vect__209.38_937 = MEM[base: vectp.15_1, index: ivtmp.253_329, offset:
OFFSET3]; // load @ (base + _999*4 + IV + OFFSET3)
vect__214.42_930 = vect_cst__931 * vect__209.38_937;
_1235 = vect_cst__947 * vect__193.31_953 + vect__214.42_930;
_1307 = _1235 + _1256;
vect__225.45_921 = MEM[base: vectp.15_1, index: ivtmp.253_329, offset:
OFFSET4]; // load @ (base + _999*4 + IV + OFFSET4)
vect__247.54_900 = MEM[base: vectp.15_1, index: ivtmp.253_329, offset:
OFFSET5]; // load @ (base + _999*4 + IV + OFFSET5)
vect__252.58_893 = vect_cst__894 * vect__247.54_900;
_1291 = vect_cst__915 * vect__225.45_921 + vect__252.58_893;
_341 = _1291 + _1307;
vect__263.61_242 = MEM[base: vectp.15_1, index: ivtmp.253_329, offset:
OFFSET6]; // load @ (base + _999*4 + IV + OFFSET6 )
vect__279.69_1073 = MEM[base: vectp.15_1, index: ivtmp.253_329, offset:
OFFSET7]; // load @ (base + _999*4 + IV + OFFSET7)
ICC exploit this successfully, it puts difference to offset:
402636: c4 01 64 59 a4 8a 20 vmulps 0x820(%r10,%r9,4),%ymm3,%ymm12
40263d: 08 00 00
402640: c4 01 14 59 b4 8a 30 vmulps 0x1030(%r10,%r9,4),%ymm13,%ymm14
402647: 10 00 00
40264a: c4 02 5d b8 a4 8a 18 vfmadd231ps
0x818(%r10,%r9,4),%ymm4,%ymm12
402651: 08 00 00
402654: c4 02 2d b8 b4 8a 28 vfmadd231ps
0x1028(%r10,%r9,4),%ymm10,%ymm14
40265b: 10 00 00
40265e: c4 41 24 57 db vxorps %ymm11,%ymm11,%ymm11
402663: c4 02 7d b8 9c 8a 2c vfmadd231ps
0x102c(%r10,%r9,4),%ymm0,%ymm11
40266a: 10 00 00
40266d: c4 02 75 b8 9c 8a 24 vfmadd231ps
0x1024(%r10,%r9,4),%ymm1,%ymm11
402674: 10 00 00
402677: c4 41 24 58 de vaddps %ymm14,%ymm11,%ymm11
40267c: c4 01 6c 59 b4 8a 20 vmulps 0x1020(%r10,%r9,4),%ymm2,%ymm14
This causes GCC to be ~50% slower than ICC on this kernel.
