https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99149
Bug ID: 99149
Summary: [11 Regression] ICE during vectorization when shared
trees contain different complex patterns
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: tnfchris at gcc dot gnu.org
Reporter: tnfchris at gcc dot gnu.org
Target Milestone: ---
Target: aarch64-*
The following testcase
class a {
float b;
float c;
public:
a(float d, float e) : b(d), c(e) {}
a operator+(a d) { return a(b + d.b, c + d.c); }
a operator*(a d) { return a(b * b - c * c, b * c + c * d.b); }
};
int f, g;
class {
a *h;
a *i;
public:
void j() {
a k = h[0], l = i[g], m = k * i[f];
i[g] = l + m;
i[f] = m;
}
} n;
main() { n.j(); }
crashes with aarch64-none-elf-g++ -w -march=armv8.3-a -O3 -S main.cpp
At -O3 there are two SLP trees created, but they share partially part of a
complex pattern:
Final SLP tree for instance 0x559a7e0:
node 0x56f5520 (max_nunits=2, refcnt=2)
op template: MEM <float> [(struct a *)_16] = _20;
stmt 0 MEM <float> [(struct a *)_16] = _20;
stmt 1 MEM <float> [(struct a *)_16 + 4B] = _23;
children 0x56f5058
node 0x56f5058 (max_nunits=2, refcnt=3)
op: VEC_PERM_EXPR
stmt 0 _20 = _18 - _19;
stmt 1 _23 = _21 + _22;
lane permutation { 0[0] 1[1] }
children 0x56f5410 0x56f5498
node 0x56f5410 (max_nunits=1, refcnt=1)
op template: _20 = _18 - _19;
{ }
children 0x56f50e0 0x56f5278
node 0x56f50e0 (max_nunits=2, refcnt=3)
op template: _18 = k$b_4 * k$b_4;
stmt 0 _18 = k$b_4 * k$b_4;
stmt 1 _21 = k$b_4 * k$c_5;
children 0x56f5168 0x56f51f0
node 0x56f5168 (max_nunits=2, refcnt=2)
op template: k$b_4 = MEM[(const struct a &)_3].b;
stmt 0 k$b_4 = MEM[(const struct a &)_3].b;
stmt 1 k$b_4 = MEM[(const struct a &)_3].b;
load permutation { 0 0 }
node 0x56f51f0 (max_nunits=2, refcnt=2)
op template: k$b_4 = MEM[(const struct a &)_3].b;
stmt 0 k$b_4 = MEM[(const struct a &)_3].b;
stmt 1 k$c_5 = MEM[(const struct a &)_3].c;
load permutation { 0 1 }
node 0x56f5278 (max_nunits=2, refcnt=3)
op template: _19 = k$c_5 * k$c_5;
stmt 0 _19 = k$c_5 * k$c_5;
stmt 1 _22 = k$c_5 * d$b_17;
children 0x56f5300 0x56f5388
node 0x56f5300 (max_nunits=2, refcnt=2)
op template: k$c_5 = MEM[(const struct a &)_3].c;
stmt 0 k$c_5 = MEM[(const struct a &)_3].c;
stmt 1 k$c_5 = MEM[(const struct a &)_3].c;
load permutation { 1 1 }
node (external) 0x56f5388 (max_nunits=1, refcnt=1)
{ k$c_5, d$b_17 }
node 0x56f5498 (max_nunits=1, refcnt=1)
op template: _23 = _21 + _22;
{ }
children 0x56f50e0 0x56f5278
These dump contains two roots, one at 0x56f5520 and one at 0x5763380:
Final SLP tree for instance 0x5763380:
node 0x56f4ec0 (max_nunits=2, refcnt=2)
op template: MEM <float> [(struct a *)_10] = _24;
stmt 0 MEM <float> [(struct a *)_10] = _24;
stmt 1 MEM <float> [(struct a *)_10 + 4B] = _25;
children 0x56f4f48
node 0x56f4f48 (max_nunits=2, refcnt=2)
op template: _24 = l$b_11 + _20;
stmt 0 _24 = l$b_11 + _20;
stmt 1 _25 = l$c_12 + _23;
children 0x56f4fd0 0x56f5058
node 0x56f4fd0 (max_nunits=2, refcnt=2)
op template: l$b_11 = MEM[(const struct a &)_10].b;
stmt 0 l$b_11 = MEM[(const struct a &)_10].b;
stmt 1 l$c_12 = MEM[(const struct a &)_10].c;
load permutation { 0 1 }
node 0x56f5058 (max_nunits=2, refcnt=2)
op: VEC_PERM_EXPR
stmt 0 _20 = _18 - _19;
stmt 1 _23 = _21 + _22;
lane permutation { 0[0] 1[1] }
children 0x56f5410 0x56f5498
node 0x56f5410 (max_nunits=1, refcnt=1)
op template: _20 = _18 - _19;
{ }
children 0x56f50e0 0x56f5278
node 0x56f50e0 (max_nunits=2, refcnt=3)
op template: _18 = k$b_4 * k$b_4;
stmt 0 _18 = k$b_4 * k$b_4;
stmt 1 _21 = k$b_4 * k$c_5;
children 0x56f5168 0x56f51f0
node 0x56f5168 (max_nunits=2, refcnt=2)
op template: k$b_4 = MEM[(const struct a &)_3].b;
stmt 0 k$b_4 = MEM[(const struct a &)_3].b;
stmt 1 k$b_4 = MEM[(const struct a &)_3].b;
load permutation { 0 0 }
node 0x56f51f0 (max_nunits=2, refcnt=2)
op template: k$b_4 = MEM[(const struct a &)_3].b;
stmt 0 k$b_4 = MEM[(const struct a &)_3].b;
stmt 1 k$c_5 = MEM[(const struct a &)_3].c;
load permutation { 0 1 }
node 0x56f5278 (max_nunits=2, refcnt=3)
op template: _19 = k$c_5 * k$c_5;
stmt 0 _19 = k$c_5 * k$c_5;
stmt 1 _22 = k$c_5 * d$b_17;
children 0x56f5300 0x56f5388
node 0x56f5300 (max_nunits=2, refcnt=2)
op template: k$c_5 = MEM[(const struct a &)_3].c;
stmt 0 k$c_5 = MEM[(const struct a &)_3].c;
stmt 1 k$c_5 = MEM[(const struct a &)_3].c;
load permutation { 1 1 }
node (external) 0x56f5388 (max_nunits=1, refcnt=1)
{ k$c_5, d$b_17 }
node 0x56f5498 (max_nunits=1, refcnt=1)
op template: _23 = _21 + _22;
{ }
children 0x56f50e0 0x56f5278
It's a bit had to see so let's render the graphs
https://sketchviz.com/@Mistuke/3f10160785531d484f9eba0de186bf65/5cedb98ad6ee1e083c3bf9574dba1c679851412e
The graph rooted to the store in *_10 matches COMPLEX_FMA but that rooted in
the store to *_16 only matches COMPLEX_MUL.
In this case the COMPLEX_FMA shouldn't be recognized as they both share the
VEC_PERM_EXPR.
Disabling the COMPLEX_FMA pattern shows that this does then generate the right
code:
main:
adrp x1, .LANCHOR0
add x2, x1, :lo12:.LANCHOR0
movi v0.2s, 0
mov w0, 0
ldr x4, [x1, #:lo12:.LANCHOR0]
ldrsw x3, [x2, 16]
ldr x1, [x2, 8]
ldr d1, [x4]
fcmla v0.2s, v1.2s, v1.2s, #0
fcmla v0.2s, v1.2s, v1.2s, #90
ldrsw x2, [x2, 20]
ldr d1, [x1, x3, lsl 3]
fadd v1.2s, v1.2s, v0.2s
str d1, [x1, x3, lsl 3]
str d0, [x1, x2, lsl 3]
ret
So the pattern matcher should probably check that all children it's going to
replace belong to the same instance. But I am not sure how.