Skip to content

Commit ce6d3a4

Browse files
alexey-bataevalexey-bataev
committed
[SLP][NFC]Add a test with incorrect cost model for short-circuit or/and, modeled via select
1 parent 8491df6 commit ce6d3a4

File tree

1 file changed

+157
-0
lines changed

1 file changed

+157
-0
lines changed

‎llvm/test/Transforms/SLPVectorizer/X86/select-logical-or-and-i1-vector.ll‎

Lines changed: 157 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,157 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
2+
; RUN: opt -S --passes=slp-vectorizer -mtriple=x86_64-pc-linux-gnu -mattr=+avx512vl < %s | FileCheck %s
3+
4+
; Verify that "select i1 %x, i1 true, i1 %y" (logical OR) and
5+
; "select i1 %x, i1 %y, i1 false" (logical AND) patterns are properly
6+
; costed as vector or/and, enabling SLP vectorization.
7+
; Reduced from a real-world molecular docking kernel where independent
8+
; scalar fcmps are combined with a loop-invariant scalar condition via
9+
; logical select, feeding into float selects stored to consecutive memory.
10+
11+
define void @select_logical_or_i1(ptr %dst,
12+
; CHECK-LABEL: define void @select_logical_or_i1(
13+
; CHECK-SAME: ptr [[DST:%.*]], float [[D0:%.*]], float [[D1:%.*]], float [[D2:%.*]], float [[D3:%.*]], float [[THRESHOLD:%.*]], float [[HPHB_VAL:%.*]], i1 [[SCALAR_COND:%.*]], float [[Y0:%.*]], float [[Y1:%.*]], float [[Y2:%.*]], float [[Y3:%.*]], float [[E0:%.*]], float [[E1:%.*]], float [[E2:%.*]], float [[E3:%.*]]) #[[ATTR0:[0-9]+]] {
14+
; CHECK-NEXT: [[ENTRY:.*:]]
15+
; CHECK-NEXT: [[CMP0:%.*]] = fcmp fast uge float [[D0]], [[THRESHOLD]]
16+
; CHECK-NEXT: [[CMP1:%.*]] = fcmp fast uge float [[D1]], [[THRESHOLD]]
17+
; CHECK-NEXT: [[CMP2:%.*]] = fcmp fast uge float [[D2]], [[THRESHOLD]]
18+
; CHECK-NEXT: [[CMP3:%.*]] = fcmp fast uge float [[D3]], [[THRESHOLD]]
19+
; CHECK-NEXT: [[OR3:%.*]] = select i1 [[CMP3]], i1 true, i1 [[SCALAR_COND]]
20+
; CHECK-NEXT: [[OR2:%.*]] = select i1 [[CMP2]], i1 true, i1 [[SCALAR_COND]]
21+
; CHECK-NEXT: [[OR1:%.*]] = select i1 [[CMP1]], i1 true, i1 [[SCALAR_COND]]
22+
; CHECK-NEXT: [[OR0:%.*]] = select i1 [[CMP0]], i1 true, i1 [[SCALAR_COND]]
23+
; CHECK-NEXT: [[TMP0:%.*]] = insertelement <4 x i1> poison, i1 [[OR0]], i32 0
24+
; CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x i1> [[TMP0]], i1 [[OR1]], i32 1
25+
; CHECK-NEXT: [[TMP2:%.*]] = insertelement <4 x i1> [[TMP1]], i1 [[OR2]], i32 2
26+
; CHECK-NEXT: [[TMP9:%.*]] = insertelement <4 x i1> [[TMP2]], i1 [[OR3]], i32 3
27+
; CHECK-NEXT: [[TMP10:%.*]] = insertelement <4 x float> poison, float [[HPHB_VAL]], i32 0
28+
; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <4 x float> [[TMP10]], <4 x float> poison, <4 x i32> zeroinitializer
29+
; CHECK-NEXT: [[TMP12:%.*]] = select <4 x i1> [[TMP9]], <4 x float> zeroinitializer, <4 x float> [[TMP11]]
30+
; CHECK-NEXT: [[TMP13:%.*]] = insertelement <4 x float> poison, float [[Y0]], i32 0
31+
; CHECK-NEXT: [[TMP14:%.*]] = insertelement <4 x float> [[TMP13]], float [[Y1]], i32 1
32+
; CHECK-NEXT: [[TMP15:%.*]] = insertelement <4 x float> [[TMP14]], float [[Y2]], i32 2
33+
; CHECK-NEXT: [[TMP16:%.*]] = insertelement <4 x float> [[TMP15]], float [[Y3]], i32 3
34+
; CHECK-NEXT: [[TMP17:%.*]] = fmul fast <4 x float> [[TMP12]], [[TMP16]]
35+
; CHECK-NEXT: [[TMP18:%.*]] = insertelement <4 x float> poison, float [[E0]], i32 0
36+
; CHECK-NEXT: [[TMP19:%.*]] = insertelement <4 x float> [[TMP18]], float [[E1]], i32 1
37+
; CHECK-NEXT: [[TMP20:%.*]] = insertelement <4 x float> [[TMP19]], float [[E2]], i32 2
38+
; CHECK-NEXT: [[TMP21:%.*]] = insertelement <4 x float> [[TMP20]], float [[E3]], i32 3
39+
; CHECK-NEXT: [[TMP22:%.*]] = fadd fast <4 x float> [[TMP21]], [[TMP17]]
40+
; CHECK-NEXT: store <4 x float> [[TMP22]], ptr [[DST]], align 4
41+
; CHECK-NEXT: ret void
42+
;
43+
float %d0, float %d1, float %d2, float %d3,
44+
float %threshold, float %hphb_val,
45+
i1 %scalar_cond,
46+
float %y0, float %y1, float %y2, float %y3,
47+
float %e0, float %e1, float %e2, float %e3) {
48+
entry:
49+
%cmp0 = fcmp fast uge float %d0, %threshold
50+
%cmp1 = fcmp fast uge float %d1, %threshold
51+
%cmp2 = fcmp fast uge float %d2, %threshold
52+
%cmp3 = fcmp fast uge float %d3, %threshold
53+
54+
; select i1 %cmp, i1 true, i1 %scalar_cond -> or i1 %cmp, %scalar_cond
55+
%or0 = select i1 %cmp0, i1 true, i1 %scalar_cond
56+
%or1 = select i1 %cmp1, i1 true, i1 %scalar_cond
57+
%or2 = select i1 %cmp2, i1 true, i1 %scalar_cond
58+
%or3 = select i1 %cmp3, i1 true, i1 %scalar_cond
59+
60+
%sel0 = select i1 %or0, float 0.000000e+00, float %hphb_val
61+
%sel1 = select i1 %or1, float 0.000000e+00, float %hphb_val
62+
%sel2 = select i1 %or2, float 0.000000e+00, float %hphb_val
63+
%sel3 = select i1 %or3, float 0.000000e+00, float %hphb_val
64+
65+
%mul0 = fmul fast float %sel0, %y0
66+
%mul1 = fmul fast float %sel1, %y1
67+
%mul2 = fmul fast float %sel2, %y2
68+
%mul3 = fmul fast float %sel3, %y3
69+
70+
%res0 = fadd fast float %e0, %mul0
71+
%res1 = fadd fast float %e1, %mul1
72+
%res2 = fadd fast float %e2, %mul2
73+
%res3 = fadd fast float %e3, %mul3
74+
75+
store float %res0, ptr %dst, align 4
76+
%p1 = getelementptr inbounds float, ptr %dst, i64 1
77+
store float %res1, ptr %p1, align 4
78+
%p2 = getelementptr inbounds float, ptr %dst, i64 2
79+
store float %res2, ptr %p2, align 4
80+
%p3 = getelementptr inbounds float, ptr %dst, i64 3
81+
store float %res3, ptr %p3, align 4
82+
ret void
83+
}
84+
85+
define void @select_logical_and_i1(ptr %dst,
86+
; CHECK-LABEL: define void @select_logical_and_i1(
87+
; CHECK-SAME: ptr [[DST:%.*]], float [[D0:%.*]], float [[D1:%.*]], float [[D2:%.*]], float [[D3:%.*]], float [[THRESHOLD:%.*]], float [[HPHB_VAL:%.*]], i1 [[SCALAR_COND:%.*]], float [[Y0:%.*]], float [[Y1:%.*]], float [[Y2:%.*]], float [[Y3:%.*]], float [[E0:%.*]], float [[E1:%.*]], float [[E2:%.*]], float [[E3:%.*]]) #[[ATTR0]] {
88+
; CHECK-NEXT: [[ENTRY:.*:]]
89+
; CHECK-NEXT: [[CMP0:%.*]] = fcmp fast uge float [[D0]], [[THRESHOLD]]
90+
; CHECK-NEXT: [[CMP1:%.*]] = fcmp fast uge float [[D1]], [[THRESHOLD]]
91+
; CHECK-NEXT: [[CMP2:%.*]] = fcmp fast uge float [[D2]], [[THRESHOLD]]
92+
; CHECK-NEXT: [[CMP3:%.*]] = fcmp fast uge float [[D3]], [[THRESHOLD]]
93+
; CHECK-NEXT: [[AND3:%.*]] = select i1 [[CMP3]], i1 [[SCALAR_COND]], i1 false
94+
; CHECK-NEXT: [[AND2:%.*]] = select i1 [[CMP2]], i1 [[SCALAR_COND]], i1 false
95+
; CHECK-NEXT: [[AND1:%.*]] = select i1 [[CMP1]], i1 [[SCALAR_COND]], i1 false
96+
; CHECK-NEXT: [[AND0:%.*]] = select i1 [[CMP0]], i1 [[SCALAR_COND]], i1 false
97+
; CHECK-NEXT: [[TMP0:%.*]] = insertelement <4 x i1> poison, i1 [[AND0]], i32 0
98+
; CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x i1> [[TMP0]], i1 [[AND1]], i32 1
99+
; CHECK-NEXT: [[TMP2:%.*]] = insertelement <4 x i1> [[TMP1]], i1 [[AND2]], i32 2
100+
; CHECK-NEXT: [[TMP9:%.*]] = insertelement <4 x i1> [[TMP2]], i1 [[AND3]], i32 3
101+
; CHECK-NEXT: [[TMP10:%.*]] = insertelement <4 x float> poison, float [[HPHB_VAL]], i32 0
102+
; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <4 x float> [[TMP10]], <4 x float> poison, <4 x i32> zeroinitializer
103+
; CHECK-NEXT: [[TMP12:%.*]] = select <4 x i1> [[TMP9]], <4 x float> zeroinitializer, <4 x float> [[TMP11]]
104+
; CHECK-NEXT: [[TMP13:%.*]] = insertelement <4 x float> poison, float [[Y0]], i32 0
105+
; CHECK-NEXT: [[TMP14:%.*]] = insertelement <4 x float> [[TMP13]], float [[Y1]], i32 1
106+
; CHECK-NEXT: [[TMP15:%.*]] = insertelement <4 x float> [[TMP14]], float [[Y2]], i32 2
107+
; CHECK-NEXT: [[TMP16:%.*]] = insertelement <4 x float> [[TMP15]], float [[Y3]], i32 3
108+
; CHECK-NEXT: [[TMP17:%.*]] = fmul fast <4 x float> [[TMP12]], [[TMP16]]
109+
; CHECK-NEXT: [[TMP18:%.*]] = insertelement <4 x float> poison, float [[E0]], i32 0
110+
; CHECK-NEXT: [[TMP19:%.*]] = insertelement <4 x float> [[TMP18]], float [[E1]], i32 1
111+
; CHECK-NEXT: [[TMP20:%.*]] = insertelement <4 x float> [[TMP19]], float [[E2]], i32 2
112+
; CHECK-NEXT: [[TMP21:%.*]] = insertelement <4 x float> [[TMP20]], float [[E3]], i32 3
113+
; CHECK-NEXT: [[TMP22:%.*]] = fadd fast <4 x float> [[TMP21]], [[TMP17]]
114+
; CHECK-NEXT: store <4 x float> [[TMP22]], ptr [[DST]], align 4
115+
; CHECK-NEXT: ret void
116+
;
117+
float %d0, float %d1, float %d2, float %d3,
118+
float %threshold, float %hphb_val,
119+
i1 %scalar_cond,
120+
float %y0, float %y1, float %y2, float %y3,
121+
float %e0, float %e1, float %e2, float %e3) {
122+
entry:
123+
%cmp0 = fcmp fast uge float %d0, %threshold
124+
%cmp1 = fcmp fast uge float %d1, %threshold
125+
%cmp2 = fcmp fast uge float %d2, %threshold
126+
%cmp3 = fcmp fast uge float %d3, %threshold
127+
128+
; select i1 %cmp, i1 %scalar_cond, i1 false -> and i1 %cmp, %scalar_cond
129+
%and0 = select i1 %cmp0, i1 %scalar_cond, i1 false
130+
%and1 = select i1 %cmp1, i1 %scalar_cond, i1 false
131+
%and2 = select i1 %cmp2, i1 %scalar_cond, i1 false
132+
%and3 = select i1 %cmp3, i1 %scalar_cond, i1 false
133+
134+
%sel0 = select i1 %and0, float 0.000000e+00, float %hphb_val
135+
%sel1 = select i1 %and1, float 0.000000e+00, float %hphb_val
136+
%sel2 = select i1 %and2, float 0.000000e+00, float %hphb_val
137+
%sel3 = select i1 %and3, float 0.000000e+00, float %hphb_val
138+
139+
%mul0 = fmul fast float %sel0, %y0
140+
%mul1 = fmul fast float %sel1, %y1
141+
%mul2 = fmul fast float %sel2, %y2
142+
%mul3 = fmul fast float %sel3, %y3
143+
144+
%res0 = fadd fast float %e0, %mul0
145+
%res1 = fadd fast float %e1, %mul1
146+
%res2 = fadd fast float %e2, %mul2
147+
%res3 = fadd fast float %e3, %mul3
148+
149+
store float %res0, ptr %dst, align 4
150+
%p1 = getelementptr inbounds float, ptr %dst, i64 1
151+
store float %res1, ptr %p1, align 4
152+
%p2 = getelementptr inbounds float, ptr %dst, i64 2
153+
store float %res2, ptr %p2, align 4
154+
%p3 = getelementptr inbounds float, ptr %dst, i64 3
155+
store float %res3, ptr %p3, align 4
156+
ret void
157+
}

0 commit comments

Comments
 (0)