xref: /llvm-project/llvm/test/Transforms/LoopVectorize/AArch64/scalable-call.ll (revision 56c091ea7106507b36015297ee9005c9d5fab0bf) 
1; RUN: opt -S -passes=loop-vectorize,instcombine -force-vector-interleave=1 -mattr=+sve -mtriple aarch64-unknown-linux-gnu \
2; RUN:     -prefer-predicate-over-epilogue=scalar-epilogue -pass-remarks-missed=loop-vectorize < %s 2>%t | FileCheck %s
3; RUN: cat %t | FileCheck %s --check-prefix=CHECK-REMARKS
4; RUN: opt -S -passes=loop-vectorize,instcombine -force-vector-interleave=1 -force-target-instruction-cost=1 -mattr=+sve -mtriple aarch64-unknown-linux-gnu \
5; RUN:     -prefer-predicate-over-epilogue=scalar-epilogue -pass-remarks-missed=loop-vectorize < %s 2>%t | FileCheck %s
6; RUN: cat %t | FileCheck %s --check-prefix=CHECK-REMARKS
7
8define void @vec_load(i64 %N, ptr nocapture %a, ptr nocapture readonly %b) {
9; CHECK-LABEL: @vec_load
10; CHECK: vector.body:
11; CHECK: %[[LOAD:.*]] = load <vscale x 2 x double>, ptr
12; CHECK: call <vscale x 2 x double> @foo_vec(<vscale x 2 x double> %[[LOAD]])
13entry:
14  %cmp7 = icmp sgt i64 %N, 0
15  br i1 %cmp7, label %for.body, label %for.end
16
17for.body:                                         ; preds = %for.body.preheader, %for.body
18  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
19  %arrayidx = getelementptr inbounds double, ptr %b, i64 %iv
20  %0 = load double, ptr %arrayidx, align 8
21  %1 = call double @foo(double %0) #0
22  %add = fadd double %1, 1.000000e+00
23  %arrayidx2 = getelementptr inbounds double, ptr %a, i64 %iv
24  store double %add, ptr %arrayidx2, align 8
25  %iv.next = add nuw nsw i64 %iv, 1
26  %exitcond.not = icmp eq i64 %iv.next, %N
27  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !1
28
29for.end:                                 ; preds = %for.body, %entry
30  ret void
31}
32
33define void @vec_scalar(i64 %N, ptr nocapture %a) {
34; CHECK-LABEL: @vec_scalar
35; CHECK: vector.body:
36; CHECK: call <vscale x 2 x double> @foo_vec(<vscale x 2 x double> splat (double 1.000000e+01))
37entry:
38  %cmp7 = icmp sgt i64 %N, 0
39  br i1 %cmp7, label %for.body, label %for.end
40
41for.body:                                         ; preds = %for.body.preheader, %for.body
42  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
43  %0 = call double @foo(double 10.0) #0
44  %sub = fsub double %0, 1.000000e+00
45  %arrayidx = getelementptr inbounds double, ptr %a, i64 %iv
46  store double %sub, ptr %arrayidx, align 8
47  %iv.next = add nuw nsw i64 %iv, 1
48  %exitcond.not = icmp eq i64 %iv.next, %N
49  br i1 %exitcond.not, label %for.end, label %for.body, !llvm.loop !1
50
51for.end:                                 ; preds = %for.body, %entry
52  ret void
53}
54
55define void @vec_ptr(i64 %N, ptr noalias %a, ptr readnone %b) {
56; CHECK-LABEL: @vec_ptr
57; CHECK: for.body:
58; CHECK: %[[LOAD:.*]] = load ptr, ptr
59; CHECK: call i64 @bar(ptr %[[LOAD]])
60entry:
61  %cmp7 = icmp sgt i64 %N, 0
62  br i1 %cmp7, label %for.body, label %for.end
63
64for.body:
65  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
66  %gep = getelementptr ptr, ptr %b, i64 %iv
67  %load = load ptr, ptr %gep
68  %call = call i64 @bar(ptr %load) #1
69  %arrayidx = getelementptr inbounds i64, ptr %a, i64 %iv
70  store i64 %call, ptr %arrayidx
71  %iv.next = add nuw nsw i64 %iv, 1
72  %exitcond = icmp eq i64 %iv.next, 1024
73  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !1
74
75for.end:
76  ret void
77}
78
79define void @vec_intrinsic(i64 %N, ptr nocapture readonly %a) {
80; CHECK-LABEL: @vec_intrinsic
81; CHECK: vector.body:
82; CHECK: %[[LOAD:.*]] = load <vscale x 2 x double>, ptr
83; CHECK: call fast <vscale x 2 x double> @sin_vec_nxv2f64(<vscale x 2 x double> %[[LOAD]])
84entry:
85  %cmp7 = icmp sgt i64 %N, 0
86  br i1 %cmp7, label %for.body, label %for.end
87
88for.body:
89  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
90  %arrayidx = getelementptr inbounds double, ptr %a, i64 %iv
91  %0 = load double, ptr %arrayidx, align 8
92  %1 = call fast double @llvm.sin.f64(double %0) #2
93  %add = fadd fast double %1, 1.000000e+00
94  store double %add, ptr %arrayidx, align 8
95  %iv.next = add nuw nsw i64 %iv, 1
96  %exitcond = icmp eq i64 %iv.next, %N
97  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !1
98
99for.end:
100  ret void
101}
102
103; CHECK-REMARKS: UserVF ignored because of invalid costs.
104; CHECK-REMARKS-NEXT: t.c:3:10: Recipe with invalid costs prevented vectorization at VF=(vscale x 1): load
105; CHECK-REMARKS-NEXT: t.c:3:20: Recipe with invalid costs prevented vectorization at VF=(vscale x 1, vscale x 2): call to llvm.sin
106; CHECK-REMARKS-NEXT: t.c:3:30: Recipe with invalid costs prevented vectorization at VF=(vscale x 1): store
107define void @vec_sin_no_mapping(ptr noalias nocapture %dst, ptr noalias nocapture readonly %src, i64 %n) {
108; CHECK: @vec_sin_no_mapping
109; CHECK: call fast <2 x float> @llvm.sin.v2f32
110; CHECK-NOT: <vscale x
111entry:
112  br label %for.body
113
114for.body:                                         ; preds = %entry, %for.body
115  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
116  %arrayidx = getelementptr inbounds float, ptr %src, i64 %i.07
117  %0 = load float, ptr %arrayidx, align 4, !dbg !11
118  %1 = tail call fast float @llvm.sin.f32(float %0), !dbg !12
119  %arrayidx1 = getelementptr inbounds float, ptr %dst, i64 %i.07
120  store float %1, ptr %arrayidx1, align 4, !dbg !13
121  %inc = add nuw nsw i64 %i.07, 1
122  %exitcond.not = icmp eq i64 %inc, %n
123  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !1
124
125for.cond.cleanup:                                 ; preds = %for.body
126  ret void
127}
128
129; CHECK-REMARKS: UserVF ignored because of invalid costs.
130; CHECK-REMARKS-NEXT: t.c:3:10: Recipe with invalid costs prevented vectorization at VF=(vscale x 1): load
131; CHECK-REMARKS-NEXT: t.c:3:30: Recipe with invalid costs prevented vectorization at VF=(vscale x 1): fadd
132; CHECK-REMARKS-NEXT: t.c:3:30: Recipe with invalid costs prevented vectorization at VF=(vscale x 1, vscale x 2): call to llvm.sin
133; CHECK-REMARKS-NEXT: t.c:3:20: Recipe with invalid costs prevented vectorization at VF=(vscale x 1, vscale x 2): call to llvm.sin
134; CHECK-REMARKS-NEXT: t.c:3:40: Recipe with invalid costs prevented vectorization at VF=(vscale x 1): store
135define void @vec_sin_no_mapping_ite(ptr noalias nocapture %dst, ptr noalias nocapture readonly %src, i64 %n) {
136; CHECK: @vec_sin_no_mapping_ite
137; CHECK-NOT: <vscale x
138; CHECK: ret
139entry:
140  br label %for.body
141
142for.body:                                         ; preds = %entry, %if.end
143  %i.07 = phi i64 [ %inc, %if.end ], [ 0, %entry ]
144  %arrayidx = getelementptr inbounds float, ptr %src, i64 %i.07
145  %0 = load float, ptr %arrayidx, align 4, !dbg !11
146  %cmp = fcmp ugt float %0, 0.0000
147  br i1 %cmp, label %if.then, label %if.else
148if.then:
149  %1 = tail call fast float @llvm.sin.f32(float %0), !dbg !12
150  br label %if.end
151if.else:
152  %add = fadd float %0, 12.0, !dbg !13
153  %2 = tail call fast float @llvm.sin.f32(float %add), !dbg !13
154  br label %if.end
155if.end:
156  %3 = phi float [%1, %if.then], [%2, %if.else]
157  %arrayidx1 = getelementptr inbounds float, ptr %dst, i64 %i.07
158  store float %3, ptr %arrayidx1, align 4, !dbg !14
159  %inc = add nuw nsw i64 %i.07, 1
160  %exitcond.not = icmp eq i64 %inc, %n
161  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !1
162
163for.cond.cleanup:                                 ; preds = %for.body
164  ret void
165}
166
167; CHECK-REMARKS: UserVF ignored because of invalid costs.
168; CHECK-REMARKS-NEXT: t.c:3:10: Recipe with invalid costs prevented vectorization at VF=(vscale x 1): load
169; CHECK-REMARKS-NEXT: t.c:3:20: Recipe with invalid costs prevented vectorization at VF=(vscale x 1, vscale x 2): call to llvm.sin
170; CHECK-REMARKS-NEXT: t.c:3:30: Recipe with invalid costs prevented vectorization at VF=(vscale x 1): store
171define void @vec_sin_fixed_mapping(ptr noalias nocapture %dst, ptr noalias nocapture readonly %src, i64 %n) {
172; CHECK: @vec_sin_fixed_mapping
173; CHECK: call fast <2 x float> @llvm.sin.v2f32
174; CHECK-NOT: <vscale x
175entry:
176  br label %for.body
177
178for.body:                                         ; preds = %entry, %for.body
179  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
180  %arrayidx = getelementptr inbounds float, ptr %src, i64 %i.07
181  %0 = load float, ptr %arrayidx, align 4, !dbg !11
182  %1 = tail call fast float @llvm.sin.f32(float %0) #3, !dbg !12
183  %arrayidx1 = getelementptr inbounds float, ptr %dst, i64 %i.07
184  store float %1, ptr %arrayidx1, align 4, !dbg !13
185  %inc = add nuw nsw i64 %i.07, 1
186  %exitcond.not = icmp eq i64 %inc, %n
187  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !1
188
189for.cond.cleanup:                                 ; preds = %for.body
190  ret void
191}
192
193; Even though there are no function mappings attached to the call
194; in the loop below we can still vectorize the loop because SVE has
195; hardware support in the form of the 'fqsrt' instruction.
196define void @vec_sqrt_no_mapping(ptr noalias nocapture %dst, ptr noalias nocapture readonly %src, i64 %n) {
197; CHECK: @vec_sqrt_no_mapping
198; CHECK: call fast <vscale x 2 x float> @llvm.sqrt.nxv2f32
199entry:
200  br label %for.body
201
202for.body:                                         ; preds = %entry, %for.body
203  %i.07 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
204  %arrayidx = getelementptr inbounds float, ptr %src, i64 %i.07
205  %0 = load float, ptr %arrayidx, align 4
206  %1 = tail call fast float @llvm.sqrt.f32(float %0)
207  %arrayidx1 = getelementptr inbounds float, ptr %dst, i64 %i.07
208  store float %1, ptr %arrayidx1, align 4
209  %inc = add nuw nsw i64 %i.07, 1
210  %exitcond.not = icmp eq i64 %inc, %n
211  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !1
212
213for.cond.cleanup:                                 ; preds = %for.body
214  ret void
215}
216
217
218declare double @foo(double)
219declare i64 @bar(ptr)
220declare double @llvm.sin.f64(double)
221declare float @llvm.sin.f32(float)
222declare float @llvm.sqrt.f32(float)
223
224declare <vscale x 2 x double> @foo_vec(<vscale x 2 x double>)
225declare <vscale x 2 x i64> @bar_vec(<vscale x 2 x ptr>)
226declare <vscale x 2 x double> @sin_vec_nxv2f64(<vscale x 2 x double>)
227declare <2 x double> @sin_vec_v2f64(<2 x double>)
228
229attributes #0 = { "vector-function-abi-variant"="_ZGVsNxv_foo(foo_vec)" }
230attributes #1 = { "vector-function-abi-variant"="_ZGVsNxv_bar(bar_vec)" }
231attributes #2 = { "vector-function-abi-variant"="_ZGVsNxv_llvm.sin.f64(sin_vec_nxv2f64)" }
232attributes #3 = { "vector-function-abi-variant"="_ZGV_LLVM_N2v_llvm.sin.f64(sin_vec_v2f64)" }
233
234!1 = distinct !{!1, !2, !3}
235!2 = !{!"llvm.loop.vectorize.width", i32 2}
236!3 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
237
238!llvm.dbg.cu = !{!4}
239!llvm.module.flags = !{!7}
240!llvm.ident = !{!8}
241
242!4 = distinct !DICompileUnit(language: DW_LANG_C99, file: !5, producer: "clang", isOptimized: true, runtimeVersion: 0, emissionKind: NoDebug, enums: !6, splitDebugInlining: false, nameTableKind: None)
243!5 = !DIFile(filename: "t.c", directory: "somedir")
244!6 = !{}
245!7 = !{i32 2, !"Debug Info Version", i32 3}
246!8 = !{!"clang"}
247!9 = distinct !DISubprogram(name: "foo", scope: !5, file: !5, line: 2, type: !10, scopeLine: 2, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !4, retainedNodes: !6)
248!10 = !DISubroutineType(types: !6)
249!11 = !DILocation(line: 3, column: 10, scope: !9)
250!12 = !DILocation(line: 3, column: 20, scope: !9)
251!13 = !DILocation(line: 3, column: 30, scope: !9)
252!14 = !DILocation(line: 3, column: 40, scope: !9)
253