xref: /llvm-project/llvm/test/Transforms/LoopVectorize/X86/reduction-fastmath.ll (revision 0294dab79e24cc4fc41e2d9fc77ad02730e412bc)
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -S -passes=loop-vectorize < %s | FileCheck %s
3
4target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
5target triple = "x86_64-unknown-linux-gnu"
6
7define float @reduction_sum_float_ieee(i32 %n, ptr %array) {
8; CHECK-LABEL: @reduction_sum_float_ieee(
9; CHECK-NEXT:  entry:
10; CHECK-NEXT:    [[ENTRY_COND:%.*]] = icmp ne i32 0, 4096
11; CHECK-NEXT:    br i1 [[ENTRY_COND]], label [[LOOP_PREHEADER:%.*]], label [[LOOP_EXIT:%.*]]
12; CHECK:       loop.preheader:
13; CHECK-NEXT:    br label [[LOOP:%.*]]
14; CHECK:       loop:
15; CHECK-NEXT:    [[IDX:%.*]] = phi i32 [ [[IDX_INC:%.*]], [[LOOP]] ], [ 0, [[LOOP_PREHEADER]] ]
16; CHECK-NEXT:    [[SUM:%.*]] = phi float [ [[SUM_INC:%.*]], [[LOOP]] ], [ 0.000000e+00, [[LOOP_PREHEADER]] ]
17; CHECK-NEXT:    [[ADDRESS:%.*]] = getelementptr float, ptr [[ARRAY:%.*]], i32 [[IDX]]
18; CHECK-NEXT:    [[VALUE:%.*]] = load float, ptr [[ADDRESS]], align 4
19; CHECK-NEXT:    [[SUM_INC]] = fadd float [[SUM]], [[VALUE]]
20; CHECK-NEXT:    [[IDX_INC]] = add i32 [[IDX]], 1
21; CHECK-NEXT:    [[BE_COND:%.*]] = icmp ne i32 [[IDX_INC]], 4096
22; CHECK-NEXT:    br i1 [[BE_COND]], label [[LOOP]], label [[LOOP_EXIT_LOOPEXIT:%.*]]
23; CHECK:       loop.exit.loopexit:
24; CHECK-NEXT:    [[SUM_INC_LCSSA:%.*]] = phi float [ [[SUM_INC]], [[LOOP]] ]
25; CHECK-NEXT:    br label [[LOOP_EXIT]]
26; CHECK:       loop.exit:
27; CHECK-NEXT:    [[SUM_LCSSA:%.*]] = phi float [ 0.000000e+00, [[ENTRY:%.*]] ], [ [[SUM_INC_LCSSA]], [[LOOP_EXIT_LOOPEXIT]] ]
28; CHECK-NEXT:    ret float [[SUM_LCSSA]]
29;
30entry:
31  %entry.cond = icmp ne i32 0, 4096
32  br i1 %entry.cond, label %loop, label %loop.exit
33
34loop:
35  %idx = phi i32 [ 0, %entry ], [ %idx.inc, %loop ]
36  %sum = phi float [ 0.000000e+00, %entry ], [ %sum.inc, %loop ]
37  %address = getelementptr float, ptr %array, i32 %idx
38  %value = load float, ptr %address
39  %sum.inc = fadd float %sum, %value
40  %idx.inc = add i32 %idx, 1
41  %be.cond = icmp ne i32 %idx.inc, 4096
42  br i1 %be.cond, label %loop, label %loop.exit
43
44loop.exit:
45  %sum.lcssa = phi float [ %sum.inc, %loop ], [ 0.000000e+00, %entry ]
46  ret float %sum.lcssa
47}
48
49define float @reduction_sum_float_fastmath(i32 %n, ptr %array) {
50; CHECK-LABEL: @reduction_sum_float_fastmath(
51; CHECK-NEXT:  entry:
52; CHECK-NEXT:    [[ENTRY_COND:%.*]] = icmp ne i32 0, 4096
53; CHECK-NEXT:    br i1 [[ENTRY_COND]], label [[LOOP_PREHEADER:%.*]], label [[LOOP_EXIT:%.*]]
54; CHECK:       loop.preheader:
55; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
56; CHECK:       vector.ph:
57; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
58; CHECK:       vector.body:
59; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
60; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
61; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
62; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[INDEX]], 0
63; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr float, ptr [[ARRAY:%.*]], i32 [[TMP0]]
64; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr float, ptr [[TMP2]], i32 0
65; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr float, ptr [[TMP2]], i32 4
66; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP4]], align 4
67; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <4 x float>, ptr [[TMP5]], align 4
68; CHECK-NEXT:    [[TMP6]] = fadd fast <4 x float> [[VEC_PHI]], [[WIDE_LOAD]]
69; CHECK-NEXT:    [[TMP7]] = fadd fast <4 x float> [[VEC_PHI1]], [[WIDE_LOAD2]]
70; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
71; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], 4096
72; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
73; CHECK:       middle.block:
74; CHECK-NEXT:    [[BIN_RDX:%.*]] = fadd fast <4 x float> [[TMP7]], [[TMP6]]
75; CHECK-NEXT:    [[TMP9:%.*]] = call fast float @llvm.vector.reduce.fadd.v4f32(float 0.000000e+00, <4 x float> [[BIN_RDX]])
76; CHECK-NEXT:    br i1 true, label [[LOOP_EXIT_LOOPEXIT:%.*]], label [[SCALAR_PH]]
77; CHECK:       scalar.ph:
78; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[LOOP_PREHEADER]] ]
79; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ [[TMP9]], [[MIDDLE_BLOCK]] ], [ 0.000000e+00, [[LOOP_PREHEADER]] ]
80; CHECK-NEXT:    br label [[LOOP:%.*]]
81; CHECK:       loop:
82; CHECK-NEXT:    [[IDX:%.*]] = phi i32 [ [[IDX_INC:%.*]], [[LOOP]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
83; CHECK-NEXT:    [[SUM:%.*]] = phi float [ [[SUM_INC:%.*]], [[LOOP]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
84; CHECK-NEXT:    [[ADDRESS:%.*]] = getelementptr float, ptr [[ARRAY]], i32 [[IDX]]
85; CHECK-NEXT:    [[VALUE:%.*]] = load float, ptr [[ADDRESS]], align 4
86; CHECK-NEXT:    [[SUM_INC]] = fadd fast float [[SUM]], [[VALUE]]
87; CHECK-NEXT:    [[IDX_INC]] = add i32 [[IDX]], 1
88; CHECK-NEXT:    [[BE_COND:%.*]] = icmp ne i32 [[IDX_INC]], 4096
89; CHECK-NEXT:    br i1 [[BE_COND]], label [[LOOP]], label [[LOOP_EXIT_LOOPEXIT]], !llvm.loop [[LOOP3:![0-9]+]]
90; CHECK:       loop.exit.loopexit:
91; CHECK-NEXT:    [[SUM_INC_LCSSA:%.*]] = phi float [ [[SUM_INC]], [[LOOP]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
92; CHECK-NEXT:    br label [[LOOP_EXIT]]
93; CHECK:       loop.exit:
94; CHECK-NEXT:    [[SUM_LCSSA:%.*]] = phi float [ 0.000000e+00, [[ENTRY:%.*]] ], [ [[SUM_INC_LCSSA]], [[LOOP_EXIT_LOOPEXIT]] ]
95; CHECK-NEXT:    ret float [[SUM_LCSSA]]
96;
97entry:
98  %entry.cond = icmp ne i32 0, 4096
99  br i1 %entry.cond, label %loop, label %loop.exit
100
101loop:
102  %idx = phi i32 [ 0, %entry ], [ %idx.inc, %loop ]
103  %sum = phi float [ 0.000000e+00, %entry ], [ %sum.inc, %loop ]
104  %address = getelementptr float, ptr %array, i32 %idx
105  %value = load float, ptr %address
106  %sum.inc = fadd fast float %sum, %value
107  %idx.inc = add i32 %idx, 1
108  %be.cond = icmp ne i32 %idx.inc, 4096
109  br i1 %be.cond, label %loop, label %loop.exit
110
111loop.exit:
112  %sum.lcssa = phi float [ %sum.inc, %loop ], [ 0.000000e+00, %entry ]
113  ret float %sum.lcssa
114}
115
116define float @reduction_sum_float_only_reassoc(i32 %n, ptr %array) {
117; CHECK-LABEL: @reduction_sum_float_only_reassoc(
118; CHECK-NEXT:  entry:
119; CHECK-NEXT:    [[ENTRY_COND:%.*]] = icmp ne i32 0, 4096
120; CHECK-NEXT:    br i1 [[ENTRY_COND]], label [[LOOP_PREHEADER:%.*]], label [[LOOP_EXIT:%.*]]
121; CHECK:       loop.preheader:
122; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
123; CHECK:       vector.ph:
124; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
125; CHECK:       vector.body:
126; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
127; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ splat (float -0.000000e+00), [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
128; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ splat (float -0.000000e+00), [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
129; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[INDEX]], 0
130; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr float, ptr [[ARRAY:%.*]], i32 [[TMP0]]
131; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr float, ptr [[TMP2]], i32 0
132; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr float, ptr [[TMP2]], i32 4
133; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP4]], align 4
134; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <4 x float>, ptr [[TMP5]], align 4
135; CHECK-NEXT:    [[TMP6]] = fadd reassoc <4 x float> [[VEC_PHI]], [[WIDE_LOAD]]
136; CHECK-NEXT:    [[TMP7]] = fadd reassoc <4 x float> [[VEC_PHI1]], [[WIDE_LOAD2]]
137; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
138; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], 4096
139; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
140; CHECK:       middle.block:
141; CHECK-NEXT:    [[BIN_RDX:%.*]] = fadd reassoc <4 x float> [[TMP7]], [[TMP6]]
142; CHECK-NEXT:    [[TMP9:%.*]] = call reassoc float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float> [[BIN_RDX]])
143; CHECK-NEXT:    br i1 true, label [[LOOP_EXIT_LOOPEXIT:%.*]], label [[SCALAR_PH]]
144; CHECK:       scalar.ph:
145; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[LOOP_PREHEADER]] ]
146; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ [[TMP9]], [[MIDDLE_BLOCK]] ], [ -0.000000e+00, [[LOOP_PREHEADER]] ]
147; CHECK-NEXT:    br label [[LOOP:%.*]]
148; CHECK:       loop:
149; CHECK-NEXT:    [[IDX:%.*]] = phi i32 [ [[IDX_INC:%.*]], [[LOOP]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
150; CHECK-NEXT:    [[SUM:%.*]] = phi float [ [[SUM_INC:%.*]], [[LOOP]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
151; CHECK-NEXT:    [[ADDRESS:%.*]] = getelementptr float, ptr [[ARRAY]], i32 [[IDX]]
152; CHECK-NEXT:    [[VALUE:%.*]] = load float, ptr [[ADDRESS]], align 4
153; CHECK-NEXT:    [[SUM_INC]] = fadd reassoc float [[SUM]], [[VALUE]]
154; CHECK-NEXT:    [[IDX_INC]] = add i32 [[IDX]], 1
155; CHECK-NEXT:    [[BE_COND:%.*]] = icmp ne i32 [[IDX_INC]], 4096
156; CHECK-NEXT:    br i1 [[BE_COND]], label [[LOOP]], label [[LOOP_EXIT_LOOPEXIT]], !llvm.loop [[LOOP5:![0-9]+]]
157; CHECK:       loop.exit.loopexit:
158; CHECK-NEXT:    [[SUM_INC_LCSSA:%.*]] = phi float [ [[SUM_INC]], [[LOOP]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
159; CHECK-NEXT:    br label [[LOOP_EXIT]]
160; CHECK:       loop.exit:
161; CHECK-NEXT:    [[SUM_LCSSA:%.*]] = phi float [ -0.000000e+00, [[ENTRY:%.*]] ], [ [[SUM_INC_LCSSA]], [[LOOP_EXIT_LOOPEXIT]] ]
162; CHECK-NEXT:    ret float [[SUM_LCSSA]]
163;
164entry:
165  %entry.cond = icmp ne i32 0, 4096
166  br i1 %entry.cond, label %loop, label %loop.exit
167
168loop:
169  %idx = phi i32 [ 0, %entry ], [ %idx.inc, %loop ]
170  %sum = phi float [ -0.000000e+00, %entry ], [ %sum.inc, %loop ]
171  %address = getelementptr float, ptr %array, i32 %idx
172  %value = load float, ptr %address
173  %sum.inc = fadd reassoc float %sum, %value
174  %idx.inc = add i32 %idx, 1
175  %be.cond = icmp ne i32 %idx.inc, 4096
176  br i1 %be.cond, label %loop, label %loop.exit
177
178loop.exit:
179  %sum.lcssa = phi float [ %sum.inc, %loop ], [ -0.000000e+00, %entry ]
180  ret float %sum.lcssa
181}
182
183define float @reduction_sum_float_only_reassoc_and_contract(i32 %n, ptr %array) {
184; CHECK-LABEL: @reduction_sum_float_only_reassoc_and_contract(
185; CHECK-NEXT:  entry:
186; CHECK-NEXT:    [[ENTRY_COND:%.*]] = icmp ne i32 0, 4096
187; CHECK-NEXT:    br i1 [[ENTRY_COND]], label [[LOOP_PREHEADER:%.*]], label [[LOOP_EXIT:%.*]]
188; CHECK:       loop.preheader:
189; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
190; CHECK:       vector.ph:
191; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
192; CHECK:       vector.body:
193; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
194; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ splat (float -0.000000e+00), [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
195; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ splat (float -0.000000e+00), [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
196; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[INDEX]], 0
197; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr float, ptr [[ARRAY:%.*]], i32 [[TMP0]]
198; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr float, ptr [[TMP2]], i32 0
199; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr float, ptr [[TMP2]], i32 4
200; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP4]], align 4
201; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <4 x float>, ptr [[TMP5]], align 4
202; CHECK-NEXT:    [[TMP6]] = fadd reassoc contract <4 x float> [[VEC_PHI]], [[WIDE_LOAD]]
203; CHECK-NEXT:    [[TMP7]] = fadd reassoc contract <4 x float> [[VEC_PHI1]], [[WIDE_LOAD2]]
204; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
205; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], 4096
206; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
207; CHECK:       middle.block:
208; CHECK-NEXT:    [[BIN_RDX:%.*]] = fadd reassoc contract <4 x float> [[TMP7]], [[TMP6]]
209; CHECK-NEXT:    [[TMP9:%.*]] = call reassoc contract float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float> [[BIN_RDX]])
210; CHECK-NEXT:    br i1 true, label [[LOOP_EXIT_LOOPEXIT:%.*]], label [[SCALAR_PH]]
211; CHECK:       scalar.ph:
212; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[LOOP_PREHEADER]] ]
213; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ [[TMP9]], [[MIDDLE_BLOCK]] ], [ -0.000000e+00, [[LOOP_PREHEADER]] ]
214; CHECK-NEXT:    br label [[LOOP:%.*]]
215; CHECK:       loop:
216; CHECK-NEXT:    [[IDX:%.*]] = phi i32 [ [[IDX_INC:%.*]], [[LOOP]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
217; CHECK-NEXT:    [[SUM:%.*]] = phi float [ [[SUM_INC:%.*]], [[LOOP]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
218; CHECK-NEXT:    [[ADDRESS:%.*]] = getelementptr float, ptr [[ARRAY]], i32 [[IDX]]
219; CHECK-NEXT:    [[VALUE:%.*]] = load float, ptr [[ADDRESS]], align 4
220; CHECK-NEXT:    [[SUM_INC]] = fadd reassoc contract float [[SUM]], [[VALUE]]
221; CHECK-NEXT:    [[IDX_INC]] = add i32 [[IDX]], 1
222; CHECK-NEXT:    [[BE_COND:%.*]] = icmp ne i32 [[IDX_INC]], 4096
223; CHECK-NEXT:    br i1 [[BE_COND]], label [[LOOP]], label [[LOOP_EXIT_LOOPEXIT]], !llvm.loop [[LOOP7:![0-9]+]]
224; CHECK:       loop.exit.loopexit:
225; CHECK-NEXT:    [[SUM_INC_LCSSA:%.*]] = phi float [ [[SUM_INC]], [[LOOP]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
226; CHECK-NEXT:    br label [[LOOP_EXIT]]
227; CHECK:       loop.exit:
228; CHECK-NEXT:    [[SUM_LCSSA:%.*]] = phi float [ -0.000000e+00, [[ENTRY:%.*]] ], [ [[SUM_INC_LCSSA]], [[LOOP_EXIT_LOOPEXIT]] ]
229; CHECK-NEXT:    ret float [[SUM_LCSSA]]
230;
231entry:
232  %entry.cond = icmp ne i32 0, 4096
233  br i1 %entry.cond, label %loop, label %loop.exit
234
235loop:
236  %idx = phi i32 [ 0, %entry ], [ %idx.inc, %loop ]
237  %sum = phi float [ -0.000000e+00, %entry ], [ %sum.inc, %loop ]
238  %address = getelementptr float, ptr %array, i32 %idx
239  %value = load float, ptr %address
240  %sum.inc = fadd reassoc contract float %sum, %value
241  %idx.inc = add i32 %idx, 1
242  %be.cond = icmp ne i32 %idx.inc, 4096
243  br i1 %be.cond, label %loop, label %loop.exit
244
245loop.exit:
246  %sum.lcssa = phi float [ %sum.inc, %loop ], [ -0.000000e+00, %entry ]
247  ret float %sum.lcssa
248}
249
250; New instructions should have the same FMF as the original code.
251; Note that the select inherits FMF from its fcmp condition.
252
253define float @PR35538(ptr nocapture readonly %a, i32 %N) #0 {
254; CHECK-LABEL: @PR35538(
255; CHECK-NEXT:  entry:
256; CHECK-NEXT:    [[CMP12:%.*]] = icmp sgt i32 [[N:%.*]], 0
257; CHECK-NEXT:    br i1 [[CMP12]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_COND_CLEANUP:%.*]]
258; CHECK:       for.body.lr.ph:
259; CHECK-NEXT:    [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64
260; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 8
261; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
262; CHECK:       vector.ph:
263; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 8
264; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]
265; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
266; CHECK:       vector.body:
267; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
268; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ splat (float -1.000000e+00), [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
269; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ splat (float -1.000000e+00), [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
270; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
271; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds float, ptr [[A:%.*]], i64 [[TMP0]]
272; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 0
273; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 4
274; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP4]], align 4
275; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <4 x float>, ptr [[TMP5]], align 4
276; CHECK-NEXT:    [[TMP6:%.*]] = fcmp nnan ninf nsz oge <4 x float> [[WIDE_LOAD]], [[VEC_PHI]]
277; CHECK-NEXT:    [[TMP7:%.*]] = fcmp nnan ninf nsz oge <4 x float> [[WIDE_LOAD2]], [[VEC_PHI1]]
278; CHECK-NEXT:    [[TMP8]] = select <4 x i1> [[TMP6]], <4 x float> [[WIDE_LOAD]], <4 x float> [[VEC_PHI]]
279; CHECK-NEXT:    [[TMP9]] = select <4 x i1> [[TMP7]], <4 x float> [[WIDE_LOAD2]], <4 x float> [[VEC_PHI1]]
280; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
281; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
282; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
283; CHECK:       middle.block:
284; CHECK-NEXT:    [[RDX_MINMAX_CMP:%.*]] = fcmp nnan ninf nsz ogt <4 x float> [[TMP8]], [[TMP9]]
285; CHECK-NEXT:    [[RDX_MINMAX_SELECT:%.*]] = select nnan ninf nsz <4 x i1> [[RDX_MINMAX_CMP]], <4 x float> [[TMP8]], <4 x float> [[TMP9]]
286; CHECK-NEXT:    [[TMP11:%.*]] = call nnan ninf nsz float @llvm.vector.reduce.fmax.v4f32(<4 x float> [[RDX_MINMAX_SELECT]])
287; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
288; CHECK-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_CLEANUP_LOOPEXIT:%.*]], label [[SCALAR_PH]]
289; CHECK:       scalar.ph:
290; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_LR_PH]] ]
291; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ [[TMP11]], [[MIDDLE_BLOCK]] ], [ -1.000000e+00, [[FOR_BODY_LR_PH]] ]
292; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
293; CHECK:       for.cond.cleanup.loopexit:
294; CHECK-NEXT:    [[MAX_0__LCSSA:%.*]] = phi float [ [[MAX_0_:%.*]], [[FOR_BODY]] ], [ [[TMP11]], [[MIDDLE_BLOCK]] ]
295; CHECK-NEXT:    br label [[FOR_COND_CLEANUP]]
296; CHECK:       for.cond.cleanup:
297; CHECK-NEXT:    [[MAX_0_LCSSA:%.*]] = phi float [ -1.000000e+00, [[ENTRY:%.*]] ], [ [[MAX_0__LCSSA]], [[FOR_COND_CLEANUP_LOOPEXIT]] ]
298; CHECK-NEXT:    ret float [[MAX_0_LCSSA]]
299; CHECK:       for.body:
300; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
301; CHECK-NEXT:    [[MAX_013:%.*]] = phi float [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[MAX_0_]], [[FOR_BODY]] ]
302; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[INDVARS_IV]]
303; CHECK-NEXT:    [[TMP12:%.*]] = load float, ptr [[ARRAYIDX]], align 4
304; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp nnan ninf nsz oge float [[TMP12]], [[MAX_013]]
305; CHECK-NEXT:    [[MAX_0_]] = select i1 [[CMP1_INV]], float [[TMP12]], float [[MAX_013]]
306; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
307; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]
308; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP_LOOPEXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]
309;
310entry:
311  %cmp12 = icmp sgt i32 %N, 0
312  br i1 %cmp12, label %for.body.lr.ph, label %for.cond.cleanup
313
314for.body.lr.ph:
315  %wide.trip.count = zext i32 %N to i64
316  br label %for.body
317
318for.cond.cleanup:
319  %max.0.lcssa = phi float [ -1.000000e+00, %entry ], [ %max.0., %for.body ]
320  ret float %max.0.lcssa
321
322for.body:
323  %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
324  %max.013 = phi float [ -1.000000e+00, %for.body.lr.ph ], [ %max.0., %for.body ]
325  %arrayidx = getelementptr inbounds float, ptr %a, i64 %indvars.iv
326  %0 = load float, ptr %arrayidx, align 4
327  %cmp1.inv = fcmp nnan ninf nsz oge float %0, %max.013
328  %max.0. = select i1 %cmp1.inv, float %0, float %max.013
329  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
330  %exitcond = icmp eq i64 %indvars.iv.next, %wide.trip.count
331  br i1 %exitcond, label %for.cond.cleanup, label %for.body
332}
333
334; Same as above, but this time the select already has matching FMF with its condition.
335
336define float @PR35538_more_FMF(ptr nocapture readonly %a, i32 %N) #0 {
337; CHECK-LABEL: @PR35538_more_FMF(
338; CHECK-NEXT:  entry:
339; CHECK-NEXT:    [[CMP12:%.*]] = icmp sgt i32 [[N:%.*]], 0
340; CHECK-NEXT:    br i1 [[CMP12]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_COND_CLEANUP:%.*]]
341; CHECK:       for.body.lr.ph:
342; CHECK-NEXT:    [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64
343; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 8
344; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
345; CHECK:       vector.ph:
346; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 8
347; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]
348; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
349; CHECK:       vector.body:
350; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
351; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ splat (float -1.000000e+00), [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
352; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ splat (float -1.000000e+00), [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
353; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
354; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds float, ptr [[A:%.*]], i64 [[TMP0]]
355; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 0
356; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 4
357; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP4]], align 4
358; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <4 x float>, ptr [[TMP5]], align 4
359; CHECK-NEXT:    [[TMP6:%.*]] = fcmp nnan ninf oge <4 x float> [[WIDE_LOAD]], [[VEC_PHI]]
360; CHECK-NEXT:    [[TMP7:%.*]] = fcmp nnan ninf oge <4 x float> [[WIDE_LOAD2]], [[VEC_PHI1]]
361; CHECK-NEXT:    [[TMP8]] = select nnan ninf <4 x i1> [[TMP6]], <4 x float> [[WIDE_LOAD]], <4 x float> [[VEC_PHI]]
362; CHECK-NEXT:    [[TMP9]] = select nnan ninf <4 x i1> [[TMP7]], <4 x float> [[WIDE_LOAD2]], <4 x float> [[VEC_PHI1]]
363; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
364; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
365; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
366; CHECK:       middle.block:
367; CHECK-NEXT:    [[RDX_MINMAX_CMP:%.*]] = fcmp nnan ninf ogt <4 x float> [[TMP8]], [[TMP9]]
368; CHECK-NEXT:    [[RDX_MINMAX_SELECT:%.*]] = select nnan ninf <4 x i1> [[RDX_MINMAX_CMP]], <4 x float> [[TMP8]], <4 x float> [[TMP9]]
369; CHECK-NEXT:    [[TMP11:%.*]] = call nnan ninf float @llvm.vector.reduce.fmax.v4f32(<4 x float> [[RDX_MINMAX_SELECT]])
370; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
371; CHECK-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_CLEANUP_LOOPEXIT:%.*]], label [[SCALAR_PH]]
372; CHECK:       scalar.ph:
373; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_LR_PH]] ]
374; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ [[TMP11]], [[MIDDLE_BLOCK]] ], [ -1.000000e+00, [[FOR_BODY_LR_PH]] ]
375; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
376; CHECK:       for.cond.cleanup.loopexit:
377; CHECK-NEXT:    [[MAX_0__LCSSA:%.*]] = phi float [ [[MAX_0_:%.*]], [[FOR_BODY]] ], [ [[TMP11]], [[MIDDLE_BLOCK]] ]
378; CHECK-NEXT:    br label [[FOR_COND_CLEANUP]]
379; CHECK:       for.cond.cleanup:
380; CHECK-NEXT:    [[MAX_0_LCSSA:%.*]] = phi float [ -1.000000e+00, [[ENTRY:%.*]] ], [ [[MAX_0__LCSSA]], [[FOR_COND_CLEANUP_LOOPEXIT]] ]
381; CHECK-NEXT:    ret float [[MAX_0_LCSSA]]
382; CHECK:       for.body:
383; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
384; CHECK-NEXT:    [[MAX_013:%.*]] = phi float [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[MAX_0_]], [[FOR_BODY]] ]
385; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[INDVARS_IV]]
386; CHECK-NEXT:    [[TMP12:%.*]] = load float, ptr [[ARRAYIDX]], align 4
387; CHECK-NEXT:    [[CMP1_INV:%.*]] = fcmp nnan ninf oge float [[TMP12]], [[MAX_013]]
388; CHECK-NEXT:    [[MAX_0_]] = select nnan ninf i1 [[CMP1_INV]], float [[TMP12]], float [[MAX_013]]
389; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
390; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]
391; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP_LOOPEXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP11:![0-9]+]]
392;
393entry:
394  %cmp12 = icmp sgt i32 %N, 0
395  br i1 %cmp12, label %for.body.lr.ph, label %for.cond.cleanup
396
397for.body.lr.ph:
398  %wide.trip.count = zext i32 %N to i64
399  br label %for.body
400
401for.cond.cleanup:
402  %max.0.lcssa = phi float [ -1.000000e+00, %entry ], [ %max.0., %for.body ]
403  ret float %max.0.lcssa
404
405for.body:
406  %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
407  %max.013 = phi float [ -1.000000e+00, %for.body.lr.ph ], [ %max.0., %for.body ]
408  %arrayidx = getelementptr inbounds float, ptr %a, i64 %indvars.iv
409  %0 = load float, ptr %arrayidx, align 4
410  %cmp1.inv = fcmp nnan ninf oge float %0, %max.013
411  %max.0. = select nnan ninf i1 %cmp1.inv, float %0, float %max.013
412  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
413  %exitcond = icmp eq i64 %indvars.iv.next, %wide.trip.count
414  br i1 %exitcond, label %for.cond.cleanup, label %for.body
415}
416
417attributes #0 = { "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "unsafe-fp-math"="false" }
418