Skip to content

Commit 5db7aed

Browse files
artagnonartagnon
committed
[LV] More extensive operand-bundles testing
1 parent eac2a27 commit 5db7aed

File tree

2 files changed

+245
-63
lines changed

2 files changed

+245
-63
lines changed

‎llvm/test/Transforms/LoopVectorize/assume.ll‎

Lines changed: 10 additions & 63 deletions
Original file line numberDiff line numberDiff line change
@@ -34,9 +34,8 @@ define void @test1(ptr noalias nocapture %a, ptr noalias nocapture readonly %b)
3434
; CHECK-NEXT: [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1600
3535
; CHECK-NEXT: br i1 [[TMP12]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
3636
; CHECK: [[MIDDLE_BLOCK]]:
37-
; CHECK-NEXT: br label %[[FOR_END:.*]]
38-
; CHECK: [[FOR_END]]:
39-
; CHECK-NEXT: ret void
37+
; CHECK-NEXT: br [[FOR_END:label %.*]]
38+
; CHECK: [[SCALAR_PH:.*:]]
4039
;
4140
entry:
4241
br label %for.body
@@ -58,6 +57,10 @@ for.end: ; preds = %for.body
5857
ret void
5958
}
6059

60+
declare void @llvm.assume(i1) #0
61+
62+
attributes #0 = { nounwind willreturn }
63+
6164
define void @test2(ptr noalias %a, ptr noalias %b) {
6265
; CHECK-LABEL: define void @test2(
6366
; CHECK-SAME: ptr noalias [[A:%.*]], ptr noalias [[B:%.*]]) {
@@ -89,11 +92,10 @@ define void @test2(ptr noalias %a, ptr noalias %b) {
8992
; CHECK-NEXT: store <2 x float> [[TMP6]], ptr [[TMP8]], align 4
9093
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
9194
; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1600
92-
; CHECK-NEXT: br i1 [[TMP9]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
95+
; CHECK-NEXT: br i1 [[TMP9]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
9396
; CHECK: [[MIDDLE_BLOCK]]:
94-
; CHECK-NEXT: br label %[[FOR_END:.*]]
95-
; CHECK: [[FOR_END]]:
96-
; CHECK-NEXT: ret void
97+
; CHECK-NEXT: br [[FOR_END:label %.*]]
98+
; CHECK: [[SCALAR_PH:.*:]]
9799
;
98100
entry:
99101
%ptrint = ptrtoint ptr %a to i64
@@ -161,7 +163,7 @@ define void @predicated_assume(ptr noalias nocapture readonly %a, ptr noalias no
161163
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
162164
; CHECK-NEXT: [[VEC_IND_NEXT]] = add <2 x i64> [[STEP_ADD]], splat (i64 2)
163165
; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
164-
; CHECK-NEXT: br i1 [[TMP9]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
166+
; CHECK-NEXT: br i1 [[TMP9]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
165167
; CHECK: [[MIDDLE_BLOCK]]:
166168
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]]
167169
; CHECK-NEXT: br i1 [[CMP_N]], [[FOR_COND_CLEANUP_LOOPEXIT:label %.*]], label %[[SCALAR_PH]]
@@ -202,58 +204,3 @@ if.end5: ; preds = %for.body, %if.else
202204
%cmp = icmp eq i64 %indvars.iv.next, %0
203205
br i1 %cmp, label %for.cond.cleanup.loopexit, label %for.body
204206
}
205-
206-
define void @operand_bundle(ptr noalias %a, ptr noalias %b) {
207-
; CHECK-LABEL: define void @operand_bundle(
208-
; CHECK-SAME: ptr noalias [[A:%.*]], ptr noalias [[B:%.*]]) {
209-
; CHECK-NEXT: [[ENTRY:.*:]]
210-
; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
211-
; CHECK: [[VECTOR_PH]]:
212-
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
213-
; CHECK: [[VECTOR_BODY]]:
214-
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
215-
; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[INDEX]], 0
216-
; CHECK-NEXT: [[TMP1:%.*]] = add i64 [[INDEX]], 1
217-
; CHECK-NEXT: [[TMP2:%.*]] = add i64 [[INDEX]], 2
218-
; CHECK-NEXT: [[TMP3:%.*]] = add i64 [[INDEX]], 3
219-
; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds float, ptr [[B]], i64 [[TMP0]]
220-
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds float, ptr [[TMP4]], i32 2
221-
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x float>, ptr [[TMP4]], align 4
222-
; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <2 x float>, ptr [[TMP5]], align 4
223-
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP0]]) ]
224-
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP1]]) ]
225-
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP2]]) ]
226-
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP3]]) ]
227-
; CHECK-NEXT: [[TMP6:%.*]] = fadd <2 x float> [[WIDE_LOAD]], splat (float 1.000000e+00)
228-
; CHECK-NEXT: [[TMP7:%.*]] = fadd <2 x float> [[WIDE_LOAD1]], splat (float 1.000000e+00)
229-
; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[TMP0]]
230-
; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds float, ptr [[TMP8]], i32 2
231-
; CHECK-NEXT: store <2 x float> [[TMP6]], ptr [[TMP8]], align 4
232-
; CHECK-NEXT: store <2 x float> [[TMP7]], ptr [[TMP9]], align 4
233-
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
234-
; CHECK-NEXT: [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1600
235-
; CHECK-NEXT: br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
236-
; CHECK: [[MIDDLE_BLOCK]]:
237-
; CHECK-NEXT: br label %[[FOR_END:.*]]
238-
; CHECK: [[FOR_END]]:
239-
; CHECK-NEXT: ret void
240-
;
241-
entry:
242-
br label %for.body
243-
244-
for.body: ; preds = %for.body, %entry
245-
%iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
246-
%arrayidx = getelementptr inbounds float, ptr %b, i64 %iv
247-
%0 = load float, ptr %arrayidx, align 4
248-
%cmp1 = fcmp ogt float %0, 1.000000e+02
249-
tail call void @llvm.assume(i1 true) ["align"(ptr %a, i64 %iv)]
250-
%add = fadd float %0, 1.000000e+00
251-
%arrayidx5 = getelementptr inbounds float, ptr %a, i64 %iv
252-
store float %add, ptr %arrayidx5, align 4
253-
%iv.next = add nuw nsw i64 %iv, 1
254-
%exitcond = icmp eq i64 %iv, 1599
255-
br i1 %exitcond, label %for.end, label %for.body
256-
257-
for.end: ; preds = %for.body
258-
ret void
259-
}

‎llvm/test/Transforms/LoopVectorize/operand-bundles.ll‎

Lines changed: 235 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,235 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
2+
; RUN: opt -p loop-vectorize -force-vector-width=8 -S %s | FileCheck %s
3+
4+
define void @call_loop_invariant_operand_bundle(ptr %dst, {float, float} %sv) {
5+
; CHECK-LABEL: define void @call_loop_invariant_operand_bundle(
6+
; CHECK-SAME: ptr [[DST:%.*]], { float, float } [[SV:%.*]]) {
7+
; CHECK-NEXT: [[ENTRY:.*:]]
8+
; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
9+
; CHECK: [[VECTOR_PH]]:
10+
; CHECK-NEXT: [[TMP0:%.*]] = extractvalue { float, float } [[SV]], 0
11+
; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x float> poison, float [[TMP0]], i64 0
12+
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x float> [[BROADCAST_SPLATINSERT]], <8 x float> poison, <8 x i32> zeroinitializer
13+
; CHECK-NEXT: [[TMP1:%.*]] = extractvalue { float, float } [[SV]], 1
14+
; CHECK-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <8 x float> poison, float [[TMP1]], i64 0
15+
; CHECK-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <8 x float> [[BROADCAST_SPLATINSERT1]], <8 x float> poison, <8 x i32> zeroinitializer
16+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
17+
; CHECK: [[VECTOR_BODY]]:
18+
; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
19+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr float, ptr [[DST]], i32 [[INDEX]]
20+
; CHECK-NEXT: [[TMP3:%.*]] = call <8 x float> @llvm.pow.v8f32(<8 x float> [[BROADCAST_SPLAT]], <8 x float> [[BROADCAST_SPLAT2]]) [ "deopt"(float 1.000000e+01) ]
21+
; CHECK-NEXT: store <8 x float> [[TMP3]], ptr [[TMP2]], align 4
22+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
23+
; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1000
24+
; CHECK-NEXT: br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
25+
; CHECK: [[MIDDLE_BLOCK]]:
26+
; CHECK-NEXT: br label %[[EXIT:.*]]
27+
; CHECK: [[EXIT]]:
28+
; CHECK-NEXT: ret void
29+
;
30+
entry:
31+
br label %loop
32+
33+
loop:
34+
%iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
35+
%a = extractvalue { float, float } %sv, 0
36+
%b = extractvalue { float, float } %sv, 1
37+
%addr = getelementptr float, ptr %dst, i32 %iv
38+
%p = call float @llvm.pow.f32(float %a, float %b) [ "deopt"(float 10.0) ]
39+
store float %p, ptr %addr
40+
%iv.next = add nsw i32 %iv, 1
41+
%cond = icmp ne i32 %iv.next, 1000
42+
br i1 %cond, label %loop, label %exit
43+
44+
exit:
45+
ret void
46+
}
47+
48+
define void @call_unknown_operand_bundle(ptr %dst, {float, float} %sv) {
49+
; CHECK-LABEL: define void @call_unknown_operand_bundle(
50+
; CHECK-SAME: ptr [[DST:%.*]], { float, float } [[SV:%.*]]) {
51+
; CHECK-NEXT: [[ENTRY:.*:]]
52+
; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
53+
; CHECK: [[VECTOR_PH]]:
54+
; CHECK-NEXT: [[TMP0:%.*]] = extractvalue { float, float } [[SV]], 0
55+
; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x float> poison, float [[TMP0]], i64 0
56+
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x float> [[BROADCAST_SPLATINSERT]], <8 x float> poison, <8 x i32> zeroinitializer
57+
; CHECK-NEXT: [[TMP1:%.*]] = extractvalue { float, float } [[SV]], 1
58+
; CHECK-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <8 x float> poison, float [[TMP1]], i64 0
59+
; CHECK-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <8 x float> [[BROADCAST_SPLATINSERT1]], <8 x float> poison, <8 x i32> zeroinitializer
60+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
61+
; CHECK: [[VECTOR_BODY]]:
62+
; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
63+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr float, ptr [[DST]], i32 [[INDEX]]
64+
; CHECK-NEXT: [[TMP3:%.*]] = call <8 x float> @llvm.pow.v8f32(<8 x float> [[BROADCAST_SPLAT]], <8 x float> [[BROADCAST_SPLAT2]]) [ "unknown"(ptr null) ]
65+
; CHECK-NEXT: store <8 x float> [[TMP3]], ptr [[TMP2]], align 4
66+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
67+
; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1000
68+
; CHECK-NEXT: br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
69+
; CHECK: [[MIDDLE_BLOCK]]:
70+
; CHECK-NEXT: br label %[[EXIT:.*]]
71+
; CHECK: [[EXIT]]:
72+
; CHECK-NEXT: ret void
73+
;
74+
entry:
75+
br label %loop
76+
77+
loop:
78+
%iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
79+
%a = extractvalue { float, float } %sv, 0
80+
%b = extractvalue { float, float } %sv, 1
81+
%addr = getelementptr float, ptr %dst, i32 %iv
82+
%p = call float @llvm.pow.f32(float %a, float %b) [ "unknown"(ptr null) ]
83+
store float %p, ptr %addr
84+
%iv.next = add nsw i32 %iv, 1
85+
%cond = icmp ne i32 %iv.next, 1000
86+
br i1 %cond, label %loop, label %exit
87+
88+
exit:
89+
ret void
90+
}
91+
92+
define void @call_cold_operand_bundle(ptr %dst, {float, float} %sv) {
93+
; CHECK-LABEL: define void @call_cold_operand_bundle(
94+
; CHECK-SAME: ptr [[DST:%.*]], { float, float } [[SV:%.*]]) {
95+
; CHECK-NEXT: [[ENTRY:.*:]]
96+
; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
97+
; CHECK: [[VECTOR_PH]]:
98+
; CHECK-NEXT: [[TMP0:%.*]] = extractvalue { float, float } [[SV]], 0
99+
; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x float> poison, float [[TMP0]], i64 0
100+
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x float> [[BROADCAST_SPLATINSERT]], <8 x float> poison, <8 x i32> zeroinitializer
101+
; CHECK-NEXT: [[TMP1:%.*]] = extractvalue { float, float } [[SV]], 1
102+
; CHECK-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <8 x float> poison, float [[TMP1]], i64 0
103+
; CHECK-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <8 x float> [[BROADCAST_SPLATINSERT1]], <8 x float> poison, <8 x i32> zeroinitializer
104+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
105+
; CHECK: [[VECTOR_BODY]]:
106+
; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
107+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr float, ptr [[DST]], i32 [[INDEX]]
108+
; CHECK-NEXT: [[TMP3:%.*]] = call <8 x float> @llvm.pow.v8f32(<8 x float> [[BROADCAST_SPLAT]], <8 x float> [[BROADCAST_SPLAT2]]) [ "cold"() ]
109+
; CHECK-NEXT: store <8 x float> [[TMP3]], ptr [[TMP2]], align 4
110+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
111+
; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1000
112+
; CHECK-NEXT: br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
113+
; CHECK: [[MIDDLE_BLOCK]]:
114+
; CHECK-NEXT: br label %[[EXIT:.*]]
115+
; CHECK: [[EXIT]]:
116+
; CHECK-NEXT: ret void
117+
;
118+
entry:
119+
br label %loop
120+
121+
loop:
122+
%iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
123+
%a = extractvalue { float, float } %sv, 0
124+
%b = extractvalue { float, float } %sv, 1
125+
%addr = getelementptr float, ptr %dst, i32 %iv
126+
%p = call float @llvm.pow.f32(float %a, float %b) [ "cold"() ]
127+
store float %p, ptr %addr
128+
%iv.next = add nsw i32 %iv, 1
129+
%cond = icmp ne i32 %iv.next, 1000
130+
br i1 %cond, label %loop, label %exit
131+
132+
exit:
133+
ret void
134+
}
135+
136+
define void @assume_loop_variant_operand_bundle(ptr noalias %a, ptr noalias %b) {
137+
; CHECK-LABEL: define void @assume_loop_variant_operand_bundle(
138+
; CHECK-SAME: ptr noalias [[A:%.*]], ptr noalias [[B:%.*]]) {
139+
; CHECK-NEXT: [[ENTRY:.*:]]
140+
; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
141+
; CHECK: [[VECTOR_PH]]:
142+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
143+
; CHECK: [[VECTOR_BODY]]:
144+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
145+
; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[INDEX]], 0
146+
; CHECK-NEXT: [[TMP1:%.*]] = add i64 [[INDEX]], 1
147+
; CHECK-NEXT: [[TMP2:%.*]] = add i64 [[INDEX]], 2
148+
; CHECK-NEXT: [[TMP3:%.*]] = add i64 [[INDEX]], 3
149+
; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[INDEX]], 4
150+
; CHECK-NEXT: [[TMP5:%.*]] = add i64 [[INDEX]], 5
151+
; CHECK-NEXT: [[TMP6:%.*]] = add i64 [[INDEX]], 6
152+
; CHECK-NEXT: [[TMP7:%.*]] = add i64 [[INDEX]], 7
153+
; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds float, ptr [[B]], i64 [[TMP0]]
154+
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <8 x float>, ptr [[TMP8]], align 4
155+
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP0]]) ]
156+
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP1]]) ]
157+
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP2]]) ]
158+
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP3]]) ]
159+
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP4]]) ]
160+
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP5]]) ]
161+
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP6]]) ]
162+
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 [[TMP7]]) ]
163+
; CHECK-NEXT: [[TMP9:%.*]] = fadd <8 x float> [[WIDE_LOAD]], splat (float 1.000000e+00)
164+
; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[TMP0]]
165+
; CHECK-NEXT: store <8 x float> [[TMP9]], ptr [[TMP10]], align 4
166+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
167+
; CHECK-NEXT: [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1600
168+
; CHECK-NEXT: br i1 [[TMP11]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
169+
; CHECK: [[MIDDLE_BLOCK]]:
170+
; CHECK-NEXT: br label %[[EXIT:.*]]
171+
; CHECK: [[EXIT]]:
172+
; CHECK-NEXT: ret void
173+
;
174+
entry:
175+
br label %loop
176+
177+
loop:
178+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
179+
%arrayidx = getelementptr inbounds float, ptr %b, i64 %iv
180+
%0 = load float, ptr %arrayidx, align 4
181+
%cmp1 = fcmp ogt float %0, 1.000000e+02
182+
tail call void @llvm.assume(i1 true) [ "align"(ptr %a, i64 %iv) ]
183+
%add = fadd float %0, 1.000000e+00
184+
%arrayidx5 = getelementptr inbounds float, ptr %a, i64 %iv
185+
store float %add, ptr %arrayidx5, align 4
186+
%iv.next = add nuw nsw i64 %iv, 1
187+
%exitcond = icmp eq i64 %iv, 1599
188+
br i1 %exitcond, label %exit, label %loop
189+
190+
exit:
191+
ret void
192+
}
193+
194+
define void @assume_cold_operand_bundle(ptr noalias %a, ptr noalias %b) {
195+
; CHECK-LABEL: define void @assume_cold_operand_bundle(
196+
; CHECK-SAME: ptr noalias [[A:%.*]], ptr noalias [[B:%.*]]) {
197+
; CHECK-NEXT: [[ENTRY:.*:]]
198+
; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
199+
; CHECK: [[VECTOR_PH]]:
200+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
201+
; CHECK: [[VECTOR_BODY]]:
202+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
203+
; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds float, ptr [[B]], i64 [[INDEX]]
204+
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <8 x float>, ptr [[TMP0]], align 4
205+
; CHECK-NEXT: tail call void @llvm.assume(i1 true) [ "cold"() ]
206+
; CHECK-NEXT: [[TMP1:%.*]] = fadd <8 x float> [[WIDE_LOAD]], splat (float 1.000000e+00)
207+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[INDEX]]
208+
; CHECK-NEXT: store <8 x float> [[TMP1]], ptr [[TMP2]], align 4
209+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
210+
; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1600
211+
; CHECK-NEXT: br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
212+
; CHECK: [[MIDDLE_BLOCK]]:
213+
; CHECK-NEXT: br label %[[EXIT:.*]]
214+
; CHECK: [[EXIT]]:
215+
; CHECK-NEXT: ret void
216+
;
217+
entry:
218+
br label %loop
219+
220+
loop:
221+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
222+
%arrayidx = getelementptr inbounds float, ptr %b, i64 %iv
223+
%0 = load float, ptr %arrayidx, align 4
224+
%cmp1 = fcmp ogt float %0, 1.000000e+02
225+
tail call void @llvm.assume(i1 true) [ "cold"() ]
226+
%add = fadd float %0, 1.000000e+00
227+
%arrayidx5 = getelementptr inbounds float, ptr %a, i64 %iv
228+
store float %add, ptr %arrayidx5, align 4
229+
%iv.next = add nuw nsw i64 %iv, 1
230+
%exitcond = icmp eq i64 %iv, 1599
231+
br i1 %exitcond, label %exit, label %loop
232+
233+
exit:
234+
ret void
235+
}

0 commit comments

Comments
 (0)