Skip to content

Commit b6e89ad

Browse files
artagnonartagnon
committed
[LoopUtils] Minimize expanded RT-checks returned
Follow up on 528b13d ([SCEVExp] Add helper to clean up dead instructions after expansion.) to hoist the SCEVExapnder::eraseDeadInstructions call from LoopVectorize into the LoopUtils APIs add[Diff]RuntimeChecks, so that other callers (LoopDistribute and LoopVersioning) can benefit from the patch.
1 parent 96d5567 commit b6e89ad

File tree

6 files changed

+9
-19
lines changed

6 files changed

+9
-19
lines changed

‎llvm/lib/Transforms/Utils/LoopUtils.cpp‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2032,6 +2032,7 @@ Value *llvm::addRuntimeChecks(
20322032
MemoryRuntimeCheck = IsConflict;
20332033
}
20342034

2035+
Exp.eraseDeadInstructions(MemoryRuntimeCheck);
20352036
return MemoryRuntimeCheck;
20362037
}
20372038

@@ -2077,6 +2078,7 @@ Value *llvm::addDiffRuntimeChecks(
20772078
MemoryRuntimeCheck = IsConflict;
20782079
}
20792080

2081+
Expander.eraseDeadInstructions(MemoryRuntimeCheck);
20802082
return MemoryRuntimeCheck;
20812083
}
20822084

‎llvm/lib/Transforms/Utils/LoopVersioning.cpp‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -81,6 +81,8 @@ void LoopVersioning::versionLoop(
8181
} else
8282
RuntimeCheck = MemRuntimeCheck ? MemRuntimeCheck : SCEVRuntimeCheck;
8383

84+
Exp.eraseDeadInstructions(SCEVRuntimeCheck);
85+
8486
assert(RuntimeCheck && "called even though we don't need "
8587
"any runtime checks");
8688

‎llvm/lib/Transforms/Vectorize/LoopVectorize.cpp‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1849,7 +1849,6 @@ class GeneratedRTChecks {
18491849
}
18501850

18511851
SCEVExp.eraseDeadInstructions(SCEVCheckCond);
1852-
MemCheckExp.eraseDeadInstructions(MemRuntimeCheckCond);
18531852

18541853
if (!MemCheckBlock && !SCEVCheckBlock)
18551854
return;

‎llvm/test/Transforms/LoopDistribute/scev-inserted-runtime-check.ll‎

Lines changed: 0 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -12,7 +12,6 @@ define void @f(ptr noalias %a, ptr noalias %b, ptr noalias %c, ptr noalias %d, p
1212
; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[N:%.*]], -1
1313
; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32
1414
; CHECK-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]])
15-
; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0
1615
; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1
1716
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt i64 [[TMP0]], 4294967295
1817
; CHECK-NEXT: [[TMP3:%.*]] = or i1 [[MUL_OVERFLOW]], [[TMP1]]
@@ -142,7 +141,6 @@ define void @f_with_offset(ptr noalias %b, ptr noalias %c, ptr noalias %d, ptr n
142141
; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[N:%.*]], -1
143142
; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32
144143
; CHECK-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]])
145-
; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0
146144
; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1
147145
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt i64 [[TMP0]], 4294967295
148146
; CHECK-NEXT: [[TMP3:%.*]] = or i1 [[MUL_OVERFLOW]], [[TMP1]]

‎llvm/test/Transforms/LoopVersioning/incorrect-phi.ll‎

Lines changed: 5 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,4 @@
1-
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5 --check-globals none
22
; RUN: opt -passes=loop-versioning -S < %s | FileCheck %s
33

44
; Make sure all PHIs are properly updated in the exit block. Based on
@@ -9,10 +9,6 @@
99
define void @phi_with_poison() {
1010
; CHECK-LABEL: define void @phi_with_poison() {
1111
; CHECK-NEXT: [[BB6_LVER_CHECK:.*:]]
12-
; CHECK-NEXT: [[MUL:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 10, i64 0)
13-
; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i64, i1 } [[MUL]], 0
14-
; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i64, i1 } [[MUL]], 1
15-
; CHECK-NEXT: [[TMP0:%.*]] = sub i64 0, [[MUL_RESULT]]
1612
; CHECK-NEXT: br i1 poison, label %[[BB6_PH_LVER_ORIG:.*]], label %[[BB6_PH:.*]]
1713
; CHECK: [[BB6_PH_LVER_ORIG]]:
1814
; CHECK-NEXT: br label %[[BB6_LVER_ORIG:.*]]
@@ -32,8 +28,8 @@ define void @phi_with_poison() {
3228
; CHECK-NEXT: [[_TMP123:%.*]] = getelementptr [2 x [3 x [5 x i16]]], ptr @x, i16 0, i64 poison
3329
; CHECK-NEXT: [[_TMP126:%.*]] = getelementptr [3 x [5 x i16]], ptr [[_TMP123]], i16 0, i64 [[_TMP1423]]
3430
; CHECK-NEXT: [[_TMP129:%.*]] = getelementptr [5 x i16], ptr [[_TMP126]], i16 0, i64 poison
35-
; CHECK-NEXT: [[_TMP130:%.*]] = load i16, ptr [[_TMP129]], align 2
36-
; CHECK-NEXT: store i16 poison, ptr @x, align 2
31+
; CHECK-NEXT: [[_TMP130:%.*]] = load i16, ptr [[_TMP129]], align 2, !alias.scope [[META0:![0-9]+]]
32+
; CHECK-NEXT: store i16 poison, ptr @x, align 2, !alias.scope [[META3:![0-9]+]]
3733
; CHECK-NEXT: [[_TMP142]] = add i64 [[_TMP1423]], 1
3834
; CHECK-NEXT: br i1 false, label %[[BB6]], label %[[LOOP_EXIT_LOOPEXIT1:.*]]
3935
; CHECK: [[LOOP_EXIT_LOOPEXIT]]:
@@ -77,10 +73,6 @@ define void @phi_with_non_loop_defined_value() {
7773
; CHECK-LABEL: define void @phi_with_non_loop_defined_value() {
7874
; CHECK-NEXT: [[BB6_LVER_CHECK:.*:]]
7975
; CHECK-NEXT: [[T:%.*]] = add i16 1, 1
80-
; CHECK-NEXT: [[MUL:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 10, i64 0)
81-
; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i64, i1 } [[MUL]], 0
82-
; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i64, i1 } [[MUL]], 1
83-
; CHECK-NEXT: [[TMP0:%.*]] = sub i64 0, [[MUL_RESULT]]
8476
; CHECK-NEXT: br i1 poison, label %[[BB6_PH_LVER_ORIG:.*]], label %[[BB6_PH:.*]]
8577
; CHECK: [[BB6_PH_LVER_ORIG]]:
8678
; CHECK-NEXT: br label %[[BB6_LVER_ORIG:.*]]
@@ -100,8 +92,8 @@ define void @phi_with_non_loop_defined_value() {
10092
; CHECK-NEXT: [[_TMP123:%.*]] = getelementptr [2 x [3 x [5 x i16]]], ptr @x, i16 0, i64 poison
10193
; CHECK-NEXT: [[_TMP126:%.*]] = getelementptr [3 x [5 x i16]], ptr [[_TMP123]], i16 0, i64 [[_TMP1423]]
10294
; CHECK-NEXT: [[_TMP129:%.*]] = getelementptr [5 x i16], ptr [[_TMP126]], i16 0, i64 poison
103-
; CHECK-NEXT: [[_TMP130:%.*]] = load i16, ptr [[_TMP129]], align 2
104-
; CHECK-NEXT: store i16 poison, ptr @x, align 2
95+
; CHECK-NEXT: [[_TMP130:%.*]] = load i16, ptr [[_TMP129]], align 2, !alias.scope [[META5:![0-9]+]]
96+
; CHECK-NEXT: store i16 poison, ptr @x, align 2, !alias.scope [[META8:![0-9]+]]
10597
; CHECK-NEXT: [[_TMP142]] = add i64 [[_TMP1423]], 1
10698
; CHECK-NEXT: br i1 false, label %[[BB6]], label %[[LOOP_EXIT_LOOPEXIT1:.*]]
10799
; CHECK: [[LOOP_EXIT_LOOPEXIT]]:

‎llvm/test/Transforms/LoopVersioning/wrapping-pointer-versioning.ll‎

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -31,14 +31,12 @@ define void @f1(ptr noalias %a,
3131
; LV-NEXT: [[TMP0:%.*]] = add i64 [[N:%.*]], -1
3232
; LV-NEXT: [[TMP5:%.*]] = trunc i64 [[TMP0]] to i32
3333
; LV-NEXT: [[MUL2:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP5]])
34-
; LV-NEXT: [[MUL_RESULT1:%.*]] = extractvalue { i32, i1 } [[MUL2]], 0
3534
; LV-NEXT: [[MUL_OVERFLOW1:%.*]] = extractvalue { i32, i1 } [[MUL2]], 1
3635
; LV-NEXT: [[TMP1:%.*]] = icmp ugt i64 [[TMP0]], 4294967295
3736
; LV-NEXT: [[TMP8:%.*]] = or i1 [[MUL_OVERFLOW1]], [[TMP1]]
3837
; LV-NEXT: [[MUL1:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]])
3938
; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i64, i1 } [[MUL1]], 0
4039
; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i64, i1 } [[MUL1]], 1
41-
; LV-NEXT: [[TMP2:%.*]] = sub i64 0, [[MUL_RESULT]]
4240
; LV-NEXT: [[TMP3:%.*]] = getelementptr i8, ptr [[A:%.*]], i64 [[MUL_RESULT]]
4341
; LV-NEXT: [[TMP4:%.*]] = icmp ult ptr [[TMP3]], [[A]]
4442
; LV-NEXT: [[TMP6:%.*]] = or i1 [[TMP4]], [[MUL_OVERFLOW]]
@@ -264,7 +262,6 @@ define void @f3(ptr noalias %a,
264262
; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]])
265263
; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0
266264
; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1
267-
; LV-NEXT: [[TMP6:%.*]] = sub i64 0, [[MUL_RESULT3]]
268265
; LV-NEXT: [[TMP7:%.*]] = getelementptr i8, ptr [[A:%.*]], i64 [[MUL_RESULT3]]
269266
; LV-NEXT: [[TMP8:%.*]] = icmp ult ptr [[TMP7]], [[A]]
270267
; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP8]], [[MUL_OVERFLOW4]]

0 commit comments

Comments
 (0)