Skip to content

Commit 7eea55f

Browse files
artagnonartagnon
authored
LoopLoadElim: re-org tests after invalid #96656 (#97598)
After pr96656.ll were added to LAA and LoopVersioning, it was decided that the bug is in a caller of LoopVersioning, not in LAA or LoopVersioning itself. The new candidate was LoopLoadElim, but #96656 has since been marked invalid. Hence, re-organize the added tests to avoid confusion, and the testcase from the investigation to LoopLoadElim.
1 parent 1e0d3c6 commit 7eea55f

File tree

4 files changed

+175
-78
lines changed

4 files changed

+175
-78
lines changed

‎llvm/test/Analysis/LoopAccessAnalysis/pr96656.ll

Lines changed: 0 additions & 49 deletions
This file was deleted.

‎llvm/test/Analysis/LoopAccessAnalysis/symbolic-stride.ll

Lines changed: 48 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -223,6 +223,54 @@ exit:
223223
ret void
224224
}
225225

226+
define double @single_iteration_unknown_stride(i32 %x, ptr %y, i1 %cond) {
227+
; CHECK-LABEL: 'single_iteration_unknown_stride'
228+
; CHECK-NEXT: loop.body:
229+
; CHECK-NEXT: Memory dependences are safe
230+
; CHECK-NEXT: Dependences:
231+
; CHECK-NEXT: Run-time memory checks:
232+
; CHECK-NEXT: Grouped accesses:
233+
; CHECK-EMPTY:
234+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
235+
; CHECK-NEXT: SCEV assumptions:
236+
; CHECK-NEXT: Equal predicate: %x == 1
237+
; CHECK-EMPTY:
238+
; CHECK-NEXT: Expressions re-written:
239+
; CHECK-NEXT: [PSE] %gep10 = getelementptr double, ptr %gep8, i64 %mul:
240+
; CHECK-NEXT: {(8 + %y),+,(8 * (sext i32 %x to i64))<nsw>}<%loop.body>
241+
; CHECK-NEXT: --> {(8 + %y),+,8}<%loop.body>
242+
;
243+
entry:
244+
br i1 %cond, label %noloop.exit, label %loop.ph
245+
246+
loop.ph: ; preds = %entry
247+
%sext7 = sext i32 %x to i64
248+
%gep8 = getelementptr i8, ptr %y, i64 8
249+
br label %loop.body
250+
251+
loop.body: ; preds = %loop.body, %loop.ph
252+
%iv = phi i64 [ 0, %loop.ph ], [ %iv.next, %loop.body ]
253+
%mul = mul i64 %iv, %sext7
254+
%gep10 = getelementptr double, ptr %gep8, i64 %mul
255+
%load11 = load double, ptr %gep10, align 8
256+
store double %load11, ptr %y, align 8
257+
%iv.next = add i64 %iv, 1
258+
%icmp = icmp eq i64 %iv, 0
259+
br i1 %icmp, label %loop.exit, label %loop.body
260+
261+
noloop.exit: ; preds = %entry
262+
%sext = sext i32 %x to i64
263+
%gep = getelementptr double, ptr %y, i64 %sext
264+
%load5 = load double, ptr %gep, align 8
265+
ret double %load5
266+
267+
loop.exit: ; preds = %loop.body
268+
%sext2 = sext i32 %x to i64
269+
%gep2 = getelementptr double, ptr %y, i64 %sext2
270+
%load6 = load double, ptr %gep2, align 8
271+
ret double %load6
272+
}
273+
226274
; A loop with two symbolic strides.
227275
define void @two_strides(ptr noalias %A, ptr noalias %B, i64 %N, i64 %stride.1, i64 %stride.2) {
228276
; CHECK-LABEL: 'two_strides'

‎llvm/test/Transforms/LoopLoadElim/unknown-stride-known-dep.ll

Lines changed: 91 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,91 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
2+
; RUN: opt -passes=loop-load-elim -S %s | FileCheck %s
3+
4+
; The test was originally written as part of the investigation of #96656.
5+
; The bug has now been marked as invalid, and we keep the test to show
6+
; LLE's operation on known dependence returned by LAA.
7+
8+
define void @unknown_stride_known_dependence(ptr %x, ptr %y, i1 %cond) {
9+
; CHECK-LABEL: define void @unknown_stride_known_dependence(
10+
; CHECK-SAME: ptr [[X:%.*]], ptr [[Y:%.*]], i1 [[COND:%.*]]) {
11+
; CHECK-NEXT: [[ENTRY:.*:]]
12+
; CHECK-NEXT: [[LOAD:%.*]] = load i32, ptr [[X]], align 4
13+
; CHECK-NEXT: br i1 [[COND]], label %[[NOLOOP_EXIT:.*]], label %[[LOOP_LVER_CHECK:.*]]
14+
; CHECK: [[LOOP_LVER_CHECK]]:
15+
; CHECK-NEXT: [[SEXT_X:%.*]] = sext i32 [[LOAD]] to i64
16+
; CHECK-NEXT: [[GEP_8:%.*]] = getelementptr i8, ptr [[Y]], i64 8
17+
; CHECK-NEXT: [[GEP_16:%.*]] = getelementptr i8, ptr [[Y]], i64 16
18+
; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[LOAD]], 1
19+
; CHECK-NEXT: br i1 [[IDENT_CHECK]], label %[[LOOP_PH_LVER_ORIG:.*]], label %[[LOOP_PH:.*]]
20+
; CHECK: [[LOOP_PH_LVER_ORIG]]:
21+
; CHECK-NEXT: br label %[[LOOP_LVER_ORIG:.*]]
22+
; CHECK: [[LOOP_LVER_ORIG]]:
23+
; CHECK-NEXT: [[IV_LVER_ORIG:%.*]] = phi i64 [ 0, %[[LOOP_PH_LVER_ORIG]] ], [ [[IV_NEXT_LVER_ORIG:%.*]], %[[LOOP_LVER_ORIG]] ]
24+
; CHECK-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i64 [[IV_LVER_ORIG]], [[SEXT_X]]
25+
; CHECK-NEXT: [[GEP_8_MUL_LVER_ORIG:%.*]] = getelementptr double, ptr [[GEP_8]], i64 [[MUL_LVER_ORIG]]
26+
; CHECK-NEXT: [[LOAD_8_LVER_ORIG:%.*]] = load double, ptr [[GEP_8_MUL_LVER_ORIG]], align 8
27+
; CHECK-NEXT: [[GEP_16_MUL_LVER_ORIG:%.*]] = getelementptr double, ptr [[GEP_16]], i64 [[MUL_LVER_ORIG]]
28+
; CHECK-NEXT: store double [[LOAD_8_LVER_ORIG]], ptr [[GEP_16_MUL_LVER_ORIG]], align 8
29+
; CHECK-NEXT: [[IV_NEXT_LVER_ORIG]] = add i64 [[IV_LVER_ORIG]], 1
30+
; CHECK-NEXT: [[ICMP_LVER_ORIG:%.*]] = icmp eq i64 [[IV_LVER_ORIG]], 1
31+
; CHECK-NEXT: br i1 [[ICMP_LVER_ORIG]], label %[[EXIT_LOOPEXIT_LOOPEXIT:.*]], label %[[LOOP_LVER_ORIG]]
32+
; CHECK: [[LOOP_PH]]:
33+
; CHECK-NEXT: [[LOAD_INITIAL:%.*]] = load double, ptr [[GEP_8]], align 8
34+
; CHECK-NEXT: br label %[[LOOP:.*]]
35+
; CHECK: [[LOOP]]:
36+
; CHECK-NEXT: [[STORE_FORWARDED:%.*]] = phi double [ [[LOAD_INITIAL]], %[[LOOP_PH]] ], [ [[STORE_FORWARDED]], %[[LOOP]] ]
37+
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[LOOP_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
38+
; CHECK-NEXT: [[MUL:%.*]] = mul i64 [[IV]], [[SEXT_X]]
39+
; CHECK-NEXT: [[GEP_8_MUL:%.*]] = getelementptr double, ptr [[GEP_8]], i64 [[MUL]]
40+
; CHECK-NEXT: [[LOAD_8:%.*]] = load double, ptr [[GEP_8_MUL]], align 8
41+
; CHECK-NEXT: [[GEP_16_MUL:%.*]] = getelementptr double, ptr [[GEP_16]], i64 [[MUL]]
42+
; CHECK-NEXT: store double [[STORE_FORWARDED]], ptr [[GEP_16_MUL]], align 8
43+
; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 1
44+
; CHECK-NEXT: [[ICMP:%.*]] = icmp eq i64 [[IV]], 1
45+
; CHECK-NEXT: br i1 [[ICMP]], label %[[EXIT_LOOPEXIT_LOOPEXIT1:.*]], label %[[LOOP]]
46+
; CHECK: [[NOLOOP_EXIT]]:
47+
; CHECK-NEXT: [[SEXT:%.*]] = sext i32 [[LOAD]] to i64
48+
; CHECK-NEXT: [[GEP_Y:%.*]] = getelementptr double, ptr [[Y]], i64 [[SEXT]]
49+
; CHECK-NEXT: [[LOAD_Y:%.*]] = load double, ptr [[GEP_Y]], align 8
50+
; CHECK-NEXT: store double [[LOAD_Y]], ptr [[X]], align 8
51+
; CHECK-NEXT: br label %[[EXIT:.*]]
52+
; CHECK: [[EXIT_LOOPEXIT_LOOPEXIT]]:
53+
; CHECK-NEXT: br label %[[EXIT_LOOPEXIT:.*]]
54+
; CHECK: [[EXIT_LOOPEXIT_LOOPEXIT1]]:
55+
; CHECK-NEXT: br label %[[EXIT_LOOPEXIT]]
56+
; CHECK: [[EXIT_LOOPEXIT]]:
57+
; CHECK-NEXT: br label %[[EXIT]]
58+
; CHECK: [[EXIT]]:
59+
; CHECK-NEXT: ret void
60+
;
61+
entry:
62+
%load = load i32, ptr %x, align 4
63+
br i1 %cond, label %noloop.exit, label %loop.ph
64+
65+
loop.ph: ; preds = %entry
66+
%sext.x = sext i32 %load to i64
67+
%gep.8 = getelementptr i8, ptr %y, i64 8
68+
%gep.16 = getelementptr i8, ptr %y, i64 16
69+
br label %loop
70+
71+
loop: ; preds = %loop, %loop.ph
72+
%iv = phi i64 [ 0, %loop.ph ], [ %iv.next, %loop ]
73+
%mul = mul i64 %iv, %sext.x
74+
%gep.8.mul = getelementptr double, ptr %gep.8, i64 %mul
75+
%load.8 = load double, ptr %gep.8.mul, align 8
76+
%gep.16.mul = getelementptr double, ptr %gep.16, i64 %mul
77+
store double %load.8, ptr %gep.16.mul
78+
%iv.next = add i64 %iv, 1
79+
%icmp = icmp eq i64 %iv, 1
80+
br i1 %icmp, label %exit, label %loop
81+
82+
noloop.exit: ; preds = %loop.ph
83+
%sext = sext i32 %load to i64
84+
%gep.y = getelementptr double, ptr %y, i64 %sext
85+
%load.y = load double, ptr %gep.y
86+
store double %load.y, ptr %x
87+
br label %exit
88+
89+
exit: ; preds = %loop.body
90+
ret void
91+
}

‎llvm/test/Transforms/LoopVersioning/pr96656.ll renamed to ‎llvm/test/Transforms/LoopVersioning/single-iteration.ll

Lines changed: 36 additions & 29 deletions
Original file line numberDiff line numberDiff line change
@@ -1,16 +1,20 @@
11
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
22
; RUN: opt -passes=loop-versioning -S %s | FileCheck %s
33

4-
define void @lver.check.unnecessary(ptr %arg, ptr %arg1, i1 %arg2) {
5-
; CHECK-LABEL: define void @lver.check.unnecessary(
6-
; CHECK-SAME: ptr [[ARG:%.*]], ptr [[ARG1:%.*]], i1 [[ARG2:%.*]]) {
4+
; Callers should not call LoopVersioning on single-iteration loops, as it
5+
; is very likely not profitable.
6+
; LoopVersioning faithfully versions single-iteration loops when the stride
7+
; is unknown.
8+
9+
define double @single_iteration_unknown_stride(i32 %x, ptr %y, i1 %cond) {
10+
; CHECK-LABEL: define double @single_iteration_unknown_stride(
11+
; CHECK-SAME: i32 [[X:%.*]], ptr [[Y:%.*]], i1 [[COND:%.*]]) {
712
; CHECK-NEXT: [[ENTRY:.*:]]
8-
; CHECK-NEXT: [[LOAD:%.*]] = load i32, ptr [[ARG]], align 4
9-
; CHECK-NEXT: br i1 [[ARG2]], label %[[NOLOOP_EXIT:.*]], label %[[LOOP_BODY_LVER_CHECK:.*]]
13+
; CHECK-NEXT: br i1 [[COND]], label %[[NOLOOP_EXIT:.*]], label %[[LOOP_BODY_LVER_CHECK:.*]]
1014
; CHECK: [[LOOP_BODY_LVER_CHECK]]:
11-
; CHECK-NEXT: [[SEXT7:%.*]] = sext i32 [[LOAD]] to i64
12-
; CHECK-NEXT: [[GEP8:%.*]] = getelementptr i8, ptr [[ARG1]], i64 8
13-
; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[LOAD]], 1
15+
; CHECK-NEXT: [[SEXT7:%.*]] = sext i32 [[X]] to i64
16+
; CHECK-NEXT: [[GEP8:%.*]] = getelementptr i8, ptr [[Y]], i64 8
17+
; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[X]], 1
1418
; CHECK-NEXT: br i1 [[IDENT_CHECK]], label %[[LOOP_BODY_PH_LVER_ORIG:.*]], label %[[LOOP_BODY_PH:.*]]
1519
; CHECK: [[LOOP_BODY_PH_LVER_ORIG]]:
1620
; CHECK-NEXT: br label %[[LOOP_BODY_LVER_ORIG:.*]]
@@ -19,7 +23,7 @@ define void @lver.check.unnecessary(ptr %arg, ptr %arg1, i1 %arg2) {
1923
; CHECK-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i64 [[PHI_LVER_ORIG]], [[SEXT7]]
2024
; CHECK-NEXT: [[GEP10_LVER_ORIG:%.*]] = getelementptr double, ptr [[GEP8]], i64 [[MUL_LVER_ORIG]]
2125
; CHECK-NEXT: [[LOAD11_LVER_ORIG:%.*]] = load double, ptr [[GEP10_LVER_ORIG]], align 8
22-
; CHECK-NEXT: store double [[LOAD11_LVER_ORIG]], ptr [[ARG1]], align 8
26+
; CHECK-NEXT: store double [[LOAD11_LVER_ORIG]], ptr [[Y]], align 8
2327
; CHECK-NEXT: [[ADD_LVER_ORIG]] = add i64 [[PHI_LVER_ORIG]], 1
2428
; CHECK-NEXT: [[ICMP_LVER_ORIG:%.*]] = icmp eq i64 [[PHI_LVER_ORIG]], 0
2529
; CHECK-NEXT: br i1 [[ICMP_LVER_ORIG]], label %[[LOOP_EXIT_LOOPEXIT:.*]], label %[[LOOP_BODY_LVER_ORIG]]
@@ -30,49 +34,52 @@ define void @lver.check.unnecessary(ptr %arg, ptr %arg1, i1 %arg2) {
3034
; CHECK-NEXT: [[MUL:%.*]] = mul i64 [[PHI]], [[SEXT7]]
3135
; CHECK-NEXT: [[GEP10:%.*]] = getelementptr double, ptr [[GEP8]], i64 [[MUL]]
3236
; CHECK-NEXT: [[LOAD11:%.*]] = load double, ptr [[GEP10]], align 8
33-
; CHECK-NEXT: store double [[LOAD11]], ptr [[ARG1]], align 8
37+
; CHECK-NEXT: store double [[LOAD11]], ptr [[Y]], align 8
3438
; CHECK-NEXT: [[ADD]] = add i64 [[PHI]], 1
3539
; CHECK-NEXT: [[ICMP:%.*]] = icmp eq i64 [[PHI]], 0
3640
; CHECK-NEXT: br i1 [[ICMP]], label %[[LOOP_EXIT_LOOPEXIT1:.*]], label %[[LOOP_BODY]]
3741
; CHECK: [[NOLOOP_EXIT]]:
38-
; CHECK-NEXT: [[SEXT:%.*]] = sext i32 [[LOAD]] to i64
39-
; CHECK-NEXT: [[GEP:%.*]] = getelementptr double, ptr [[ARG1]], i64 [[SEXT]]
42+
; CHECK-NEXT: [[SEXT:%.*]] = sext i32 [[X]] to i64
43+
; CHECK-NEXT: [[GEP:%.*]] = getelementptr double, ptr [[Y]], i64 [[SEXT]]
4044
; CHECK-NEXT: [[LOAD5:%.*]] = load double, ptr [[GEP]], align 8
41-
; CHECK-NEXT: store double [[LOAD5]], ptr [[ARG]], align 8
42-
; CHECK-NEXT: ret void
45+
; CHECK-NEXT: ret double [[LOAD5]]
4346
; CHECK: [[LOOP_EXIT_LOOPEXIT]]:
4447
; CHECK-NEXT: br label %[[LOOP_EXIT:.*]]
4548
; CHECK: [[LOOP_EXIT_LOOPEXIT1]]:
4649
; CHECK-NEXT: br label %[[LOOP_EXIT]]
4750
; CHECK: [[LOOP_EXIT]]:
48-
; CHECK-NEXT: ret void
51+
; CHECK-NEXT: [[SEXT2:%.*]] = sext i32 [[X]] to i64
52+
; CHECK-NEXT: [[GEP2:%.*]] = getelementptr double, ptr [[Y]], i64 [[SEXT2]]
53+
; CHECK-NEXT: [[LOAD6:%.*]] = load double, ptr [[GEP2]], align 8
54+
; CHECK-NEXT: ret double [[LOAD6]]
4955
;
5056
entry:
51-
%load = load i32, ptr %arg, align 4
52-
br i1 %arg2, label %noloop.exit, label %loop.ph
57+
br i1 %cond, label %noloop.exit, label %loop.ph
5358

5459
loop.ph: ; preds = %entry
55-
%sext7 = sext i32 %load to i64
56-
%gep8 = getelementptr i8, ptr %arg1, i64 8
60+
%sext7 = sext i32 %x to i64
61+
%gep8 = getelementptr i8, ptr %y, i64 8
5762
br label %loop.body
5863

5964
loop.body: ; preds = %loop.body, %loop.ph
60-
%phi = phi i64 [ 0, %loop.ph ], [ %add, %loop.body ]
61-
%mul = mul i64 %phi, %sext7
65+
%iv = phi i64 [ 0, %loop.ph ], [ %iv.next, %loop.body ]
66+
%mul = mul i64 %iv, %sext7
6267
%gep10 = getelementptr double, ptr %gep8, i64 %mul
6368
%load11 = load double, ptr %gep10, align 8
64-
store double %load11, ptr %arg1, align 8
65-
%add = add i64 %phi, 1
66-
%icmp = icmp eq i64 %phi, 0
69+
store double %load11, ptr %y, align 8
70+
%iv.next = add i64 %iv, 1
71+
%icmp = icmp eq i64 %iv, 0
6772
br i1 %icmp, label %loop.exit, label %loop.body
6873

6974
noloop.exit: ; preds = %entry
70-
%sext = sext i32 %load to i64
71-
%gep = getelementptr double, ptr %arg1, i64 %sext
75+
%sext = sext i32 %x to i64
76+
%gep = getelementptr double, ptr %y, i64 %sext
7277
%load5 = load double, ptr %gep, align 8
73-
store double %load5, ptr %arg, align 8
74-
ret void
78+
ret double %load5
7579

7680
loop.exit: ; preds = %loop.body
77-
ret void
81+
%sext2 = sext i32 %x to i64
82+
%gep2 = getelementptr double, ptr %y, i64 %sext2
83+
%load6 = load double, ptr %gep2, align 8
84+
ret double %load6
7885
}

0 commit comments

Comments
 (0)