xref: /llvm-project/llvm/test/Transforms/LoopVectorize/AArch64/predication_costs.ll (revision 2a859b20146108af84c741a509dc0e534e045768)
1; REQUIRES: asserts
2; RUN: opt < %s -force-vector-width=2 -passes=loop-vectorize -debug-only=loop-vectorize -disable-output 2>&1 | FileCheck %s
3
4target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"
5target triple = "aarch64--linux-gnu"
6
7; Check predication-related cost calculations, including scalarization overhead
8; and block probability scaling. Note that the functionality being tested is
9; not specific to AArch64. We specify a target to get actual values for the
10; instruction costs.
11
12; CHECK-LABEL: predicated_udiv
13;
14; This test checks that we correctly compute the cost of the predicated udiv
15; instruction. If we assume the block probability is 50%, we compute the cost
16; as:
17;
18; Cost of udiv:
19;   (udiv(2) + extractelement(8) + insertelement(4)) / 2 = 7
20;
21; CHECK: Scalarizing and predicating: %tmp4 = udiv i32 %tmp2, %tmp3
22; CHECK: Found an estimated cost of 7 for VF 2 For instruction: %tmp4 = udiv i32 %tmp2, %tmp3
23;
24define i32 @predicated_udiv(ptr %a, ptr %b, i1 %c, i64 %n) {
25entry:
26  br label %for.body
27
28for.body:
29  %i = phi i64 [ 0, %entry ], [ %i.next, %for.inc ]
30  %r = phi i32 [ 0, %entry ], [ %tmp6, %for.inc ]
31  %tmp0 = getelementptr inbounds i32, ptr %a, i64 %i
32  %tmp1 = getelementptr inbounds i32, ptr %b, i64 %i
33  %tmp2 = load i32, ptr %tmp0, align 4
34  %tmp3 = load i32, ptr %tmp1, align 4
35  br i1 %c, label %if.then, label %for.inc
36
37if.then:
38  %tmp4 = udiv i32 %tmp2, %tmp3
39  br label %for.inc
40
41for.inc:
42  %tmp5 = phi i32 [ %tmp3, %for.body ], [ %tmp4, %if.then]
43  %tmp6 = add i32 %r, %tmp5
44  %i.next = add nuw nsw i64 %i, 1
45  %cond = icmp slt i64 %i.next, %n
46  br i1 %cond, label %for.body, label %for.end
47
48for.end:
49  %tmp7 = phi i32 [ %tmp6, %for.inc ]
50  ret i32 %tmp7
51}
52
53; CHECK-LABEL: predicated_store
54;
55; This test checks that we correctly compute the cost of the predicated store
56; instruction. If we assume the block probability is 50%, we compute the cost
57; as:
58;
59; Cost of store:
60;   (store(4) + extractelement(4)) / 2 = 4
61;
62; CHECK: Scalarizing and predicating: store i32 %tmp2, ptr %tmp0, align 4
63; CHECK: Found an estimated cost of 4 for VF 2 For instruction: store i32 %tmp2, ptr %tmp0, align 4
64;
65define void @predicated_store(ptr %a, i1 %c, i32 %x, i64 %n) {
66entry:
67  br label %for.body
68
69for.body:
70  %i = phi i64 [ 0, %entry ], [ %i.next, %for.inc ]
71  %tmp0 = getelementptr inbounds i32, ptr %a, i64 %i
72  %tmp1 = load i32, ptr %tmp0, align 4
73  %tmp2 = add nsw i32 %tmp1, %x
74  br i1 %c, label %if.then, label %for.inc
75
76if.then:
77  store i32 %tmp2, ptr %tmp0, align 4
78  br label %for.inc
79
80for.inc:
81  %i.next = add nuw nsw i64 %i, 1
82  %cond = icmp slt i64 %i.next, %n
83  br i1 %cond, label %for.body, label %for.end
84
85for.end:
86  ret void
87}
88
89; CHECK-LABEL: predicated_store_phi
90;
91; Same as predicate_store except we use a pointer PHI to maintain the address
92;
93; CHECK: Found scalar instruction:   %addr = phi ptr [ %a, %entry ], [ %addr.next, %for.inc ]
94; CHECK: Found scalar instruction:   %addr.next = getelementptr inbounds i32, ptr %addr, i64 1
95; CHECK: Scalarizing and predicating: store i32 %tmp2, ptr %addr, align 4
96; CHECK: Found an estimated cost of 0 for VF 2 For instruction:   %addr = phi ptr [ %a, %entry ], [ %addr.next, %for.inc ]
97; CHECK: Found an estimated cost of 4 for VF 2 For instruction: store i32 %tmp2, ptr %addr, align 4
98;
99define void @predicated_store_phi(ptr %a, i1 %c, i32 %x, i64 %n) {
100entry:
101  br label %for.body
102
103for.body:
104  %i = phi i64 [ 0, %entry ], [ %i.next, %for.inc ]
105  %addr = phi ptr [ %a, %entry ], [ %addr.next, %for.inc ]
106  %tmp1 = load i32, ptr %addr, align 4
107  %tmp2 = add nsw i32 %tmp1, %x
108  br i1 %c, label %if.then, label %for.inc
109
110if.then:
111  store i32 %tmp2, ptr %addr, align 4
112  br label %for.inc
113
114for.inc:
115  %i.next = add nuw nsw i64 %i, 1
116  %cond = icmp slt i64 %i.next, %n
117  %addr.next = getelementptr inbounds i32, ptr %addr, i64 1
118  br i1 %cond, label %for.body, label %for.end
119
120for.end:
121  ret void
122}
123
124; CHECK-LABEL: predicated_udiv_scalarized_operand
125;
126; This test checks that we correctly compute the cost of the predicated udiv
127; instruction and the add instruction it uses. The add is scalarized and sunk
128; inside the predicated block.  If we assume the block probability is 50%, we
129; compute the cost as:
130;
131; Cost of add:
132;   (add(2) + extractelement(4)) / 2 = 3
133; Cost of udiv:
134;   (udiv(2) + extractelement(4) + insertelement(4)) / 2 = 5
135;
136; CHECK: Scalarizing: %tmp3 = add nsw i32 %tmp2, %x
137; CHECK: Scalarizing and predicating: %tmp4 = udiv i32 %tmp2, %tmp3
138; CHECK: Found an estimated cost of 3 for VF 2 For instruction: %tmp3 = add nsw i32 %tmp2, %x
139; CHECK: Found an estimated cost of 5 for VF 2 For instruction: %tmp4 = udiv i32 %tmp2, %tmp3
140;
141define i32 @predicated_udiv_scalarized_operand(ptr %a, i1 %c, i32 %x, i64 %n) {
142entry:
143  br label %for.body
144
145for.body:
146  %i = phi i64 [ 0, %entry ], [ %i.next, %for.inc ]
147  %r = phi i32 [ 0, %entry ], [ %tmp6, %for.inc ]
148  %tmp0 = getelementptr inbounds i32, ptr %a, i64 %i
149  %tmp2 = load i32, ptr %tmp0, align 4
150  br i1 %c, label %if.then, label %for.inc
151
152if.then:
153  %tmp3 = add nsw i32 %tmp2, %x
154  %tmp4 = udiv i32 %tmp2, %tmp3
155  br label %for.inc
156
157for.inc:
158  %tmp5 = phi i32 [ %tmp2, %for.body ], [ %tmp4, %if.then]
159  %tmp6 = add i32 %r, %tmp5
160  %i.next = add nuw nsw i64 %i, 1
161  %cond = icmp slt i64 %i.next, %n
162  br i1 %cond, label %for.body, label %for.end
163
164for.end:
165  %tmp7 = phi i32 [ %tmp6, %for.inc ]
166  ret i32 %tmp7
167}
168
169; CHECK-LABEL: predicated_store_scalarized_operand
170;
171; This test checks that we correctly compute the cost of the predicated store
172; instruction and the add instruction it uses. The add is scalarized and sunk
173; inside the predicated block.  If we assume the block probability is 50%, we
174; compute the cost as:
175;
176; Cost of add:
177;   (add(2) + extractelement(4)) / 2 = 3
178; Cost of store:
179;   store(4) / 2 = 2
180;
181; CHECK: Scalarizing: %tmp2 = add nsw i32 %tmp1, %x
182; CHECK: Scalarizing and predicating: store i32 %tmp2, ptr %tmp0, align 4
183; CHECK: Found an estimated cost of 3 for VF 2 For instruction: %tmp2 = add nsw i32 %tmp1, %x
184; CHECK: Found an estimated cost of 2 for VF 2 For instruction: store i32 %tmp2, ptr %tmp0, align 4
185;
186define void @predicated_store_scalarized_operand(ptr %a, i1 %c, i32 %x, i64 %n) {
187entry:
188  br label %for.body
189
190for.body:
191  %i = phi i64 [ 0, %entry ], [ %i.next, %for.inc ]
192  %tmp0 = getelementptr inbounds i32, ptr %a, i64 %i
193  %tmp1 = load i32, ptr %tmp0, align 4
194  br i1 %c, label %if.then, label %for.inc
195
196if.then:
197  %tmp2 = add nsw i32 %tmp1, %x
198  store i32 %tmp2, ptr %tmp0, align 4
199  br label %for.inc
200
201for.inc:
202  %i.next = add nuw nsw i64 %i, 1
203  %cond = icmp slt i64 %i.next, %n
204  br i1 %cond, label %for.body, label %for.end
205
206for.end:
207  ret void
208}
209
210; CHECK-LABEL: predication_multi_context
211;
212; This test checks that we correctly compute the cost of multiple predicated
213; instructions in the same block. The sdiv, udiv, and store must be scalarized
214; and predicated. The sub feeding the store is scalarized and sunk inside the
215; store's predicated block. However, the add feeding the sdiv and udiv cannot
216; be sunk and is not scalarized. If we assume the block probability is 50%, we
217; compute the cost as:
218;
219; Cost of add:
220;   add(1) = 1
221; Cost of sdiv:
222;   (sdiv(2) + extractelement(8) + insertelement(4)) / 2 = 7
223; Cost of udiv:
224;   (udiv(2) + extractelement(8) + insertelement(4)) / 2 = 7
225; Cost of sub:
226;   (sub(2) + extractelement(4)) / 2 = 3
227; Cost of store:
228;   store(4) / 2 = 2
229;
230; CHECK-NOT: Scalarizing: %tmp2 = add i32 %tmp1, %x
231; CHECK:     Scalarizing and predicating: %tmp3 = sdiv i32 %tmp1, %tmp2
232; CHECK:     Scalarizing and predicating: %tmp4 = udiv i32 %tmp3, %tmp2
233; CHECK:     Scalarizing: %tmp5 = sub i32 %tmp4, %x
234; CHECK:     Scalarizing and predicating: store i32 %tmp5, ptr %tmp0, align 4
235; CHECK:     Found an estimated cost of 1 for VF 2 For instruction: %tmp2 = add i32 %tmp1, %x
236; CHECK:     Found an estimated cost of 7 for VF 2 For instruction: %tmp3 = sdiv i32 %tmp1, %tmp2
237; CHECK:     Found an estimated cost of 7 for VF 2 For instruction: %tmp4 = udiv i32 %tmp3, %tmp2
238; CHECK:     Found an estimated cost of 3 for VF 2 For instruction: %tmp5 = sub i32 %tmp4, %x
239; CHECK:     Found an estimated cost of 2 for VF 2 For instruction: store i32 %tmp5, ptr %tmp0, align 4
240;
241define void @predication_multi_context(ptr %a, i1 %c, i32 %x, i64 %n) {
242entry:
243  br label %for.body
244
245for.body:
246  %i = phi i64 [ 0, %entry ], [ %i.next, %for.inc ]
247  %tmp0 = getelementptr inbounds i32, ptr %a, i64 %i
248  %tmp1 = load i32, ptr %tmp0, align 4
249  br i1 %c, label %if.then, label %for.inc
250
251if.then:
252  %tmp2 = add i32 %tmp1, %x
253  %tmp3 = sdiv i32 %tmp1, %tmp2
254  %tmp4 = udiv i32 %tmp3, %tmp2
255  %tmp5 = sub i32 %tmp4, %x
256  store i32 %tmp5, ptr %tmp0, align 4
257  br label %for.inc
258
259for.inc:
260  %i.next = add nuw nsw i64 %i, 1
261  %cond = icmp slt i64 %i.next, %n
262  br i1 %cond, label %for.body, label %for.end
263
264for.end:
265  ret void
266}
267