[llvm] 12aef5d - [LAA] Convert test to opaque pointers (NFC)
Nikita Popov via llvm-commits
llvm-commits at lists.llvm.org
Mon Dec 19 07:58:42 PST 2022
Author: Nikita Popov
Date: 2022-12-19T16:58:33+01:00
New Revision: 12aef5df0c7ee742a6c93215a189b4641748c8e7
URL: https://github.com/llvm/llvm-project/commit/12aef5df0c7ee742a6c93215a189b4641748c8e7
DIFF: https://github.com/llvm/llvm-project/commit/12aef5df0c7ee742a6c93215a189b4641748c8e7.diff
LOG: [LAA] Convert test to opaque pointers (NFC)
When converting this test to opaque pointers (and dropping bitcast),
we get improved memory checks. Per fhahn:
> It looks like the difference is due to the logic that determines
> pointer strides in LAA not handling bitcasts. Without the
> bitcasts, the logic now triggers successfully.
Differential Revision: https://reviews.llvm.org/D140204
Added:
Modified:
llvm/test/Analysis/LoopAccessAnalysis/memcheck-ni.ll
Removed:
################################################################################
diff --git a/llvm/test/Analysis/LoopAccessAnalysis/memcheck-ni.ll b/llvm/test/Analysis/LoopAccessAnalysis/memcheck-ni.ll
index 622fcc003f3f5..068673c7e0562 100644
--- a/llvm/test/Analysis/LoopAccessAnalysis/memcheck-ni.ll
+++ b/llvm/test/Analysis/LoopAccessAnalysis/memcheck-ni.ll
@@ -5,43 +5,32 @@
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128-ni:10:11:12:13"
%jl_value_t = type opaque
-%jl_array_t = type { i8 addrspace(13)*, i64, i16, i16, i32 }
+%jl_array_t = type { ptr addrspace(13), i64, i16, i16, i32 }
-define void @test(%jl_value_t addrspace(10)** %arg) {
+define void @test(ptr %arg) {
; CHECK-LABEL: @test(
; CHECK-NEXT: L74.lver.check:
; CHECK-NEXT: [[I:%.*]] = alloca [3 x i64], align 8
-; CHECK-NEXT: [[I1:%.*]] = load [[JL_VALUE_T:%.*]] addrspace(10)*, [[JL_VALUE_T]] addrspace(10)** [[ARG:%.*]], align 8
-; CHECK-NEXT: [[I2:%.*]] = getelementptr inbounds [[JL_VALUE_T]] addrspace(10)*, [[JL_VALUE_T]] addrspace(10)** [[ARG]], i64 1
-; CHECK-NEXT: [[I3:%.*]] = load [[JL_VALUE_T]] addrspace(10)*, [[JL_VALUE_T]] addrspace(10)** [[I2]], align 8
-; CHECK-NEXT: [[I4:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[I]], i64 0, i64 0
-; CHECK-NEXT: store i64 1, i64* [[I4]], align 8
-; CHECK-NEXT: [[I5:%.*]] = getelementptr inbounds [3 x i64], [3 x i64]* [[I]], i64 0, i64 1
-; CHECK-NEXT: [[I6:%.*]] = load i64, i64* inttoptr (i64 24 to i64*), align 8
-; CHECK-NEXT: [[I7:%.*]] = addrspacecast [[JL_VALUE_T]] addrspace(10)* [[I3]] to [[JL_VALUE_T]] addrspace(11)*
-; CHECK-NEXT: [[I8:%.*]] = bitcast [[JL_VALUE_T]] addrspace(11)* [[I7]] to double addrspace(13)* addrspace(11)*
-; CHECK-NEXT: [[I9:%.*]] = load double addrspace(13)*, double addrspace(13)* addrspace(11)* [[I8]], align 8
-; CHECK-NEXT: [[I10:%.*]] = addrspacecast [[JL_VALUE_T]] addrspace(10)* [[I1]] to [[JL_VALUE_T]] addrspace(11)*
-; CHECK-NEXT: [[I11:%.*]] = bitcast [[JL_VALUE_T]] addrspace(11)* [[I10]] to double addrspace(13)* addrspace(11)*
-; CHECK-NEXT: [[I12:%.*]] = load double addrspace(13)*, double addrspace(13)* addrspace(11)* [[I11]], align 8
-; CHECK-NEXT: [[I121:%.*]] = bitcast double addrspace(13)* [[I12]] to i8 addrspace(13)*
-; CHECK-NEXT: [[I13:%.*]] = load i64, i64* [[I5]], align 8
-; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr double, double addrspace(13)* [[I12]], i64 [[I6]]
-; CHECK-NEXT: [[SCEVGEP2:%.*]] = bitcast double addrspace(13)* [[SCEVGEP]] to i8 addrspace(13)*
-; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[I6]], -1
-; CHECK-NEXT: [[TMP1:%.*]] = mul i64 [[I13]], [[TMP0]]
-; CHECK-NEXT: [[SCEVGEP3:%.*]] = getelementptr double, double addrspace(13)* [[I9]], i64 [[TMP1]]
-; CHECK-NEXT: [[TMP2:%.*]] = icmp ult double addrspace(13)* [[I9]], [[SCEVGEP3]]
-; CHECK-NEXT: [[UMIN:%.*]] = select i1 [[TMP2]], double addrspace(13)* [[I9]], double addrspace(13)* [[SCEVGEP3]]
-; CHECK-NEXT: [[UMIN4:%.*]] = bitcast double addrspace(13)* [[UMIN]] to i8 addrspace(13)*
-; CHECK-NEXT: [[TMP3:%.*]] = icmp ugt double addrspace(13)* [[I9]], [[SCEVGEP3]]
-; CHECK-NEXT: [[UMAX:%.*]] = select i1 [[TMP3]], double addrspace(13)* [[I9]], double addrspace(13)* [[SCEVGEP3]]
-; CHECK-NEXT: [[SCEVGEP5:%.*]] = getelementptr double, double addrspace(13)* [[UMAX]], i64 1
-; CHECK-NEXT: [[SCEVGEP56:%.*]] = bitcast double addrspace(13)* [[SCEVGEP5]] to i8 addrspace(13)*
-; CHECK-NEXT: [[BOUND0:%.*]] = icmp ult i8 addrspace(13)* [[I121]], [[SCEVGEP56]]
-; CHECK-NEXT: [[BOUND1:%.*]] = icmp ult i8 addrspace(13)* [[UMIN4]], [[SCEVGEP2]]
+; CHECK-NEXT: [[I1:%.*]] = load ptr addrspace(10), ptr [[ARG:%.*]], align 8
+; CHECK-NEXT: [[I2:%.*]] = getelementptr inbounds ptr addrspace(10), ptr [[ARG]], i64 1
+; CHECK-NEXT: [[I3:%.*]] = load ptr addrspace(10), ptr [[I2]], align 8
+; CHECK-NEXT: store i64 1, ptr [[I]], align 8
+; CHECK-NEXT: [[I5:%.*]] = getelementptr inbounds [3 x i64], ptr [[I]], i64 0, i64 1
+; CHECK-NEXT: [[I6:%.*]] = load i64, ptr inttoptr (i64 24 to ptr), align 8
+; CHECK-NEXT: [[I7:%.*]] = addrspacecast ptr addrspace(10) [[I3]] to ptr addrspace(11)
+; CHECK-NEXT: [[I9:%.*]] = load ptr addrspace(13), ptr addrspace(11) [[I7]], align 8
+; CHECK-NEXT: [[I10:%.*]] = addrspacecast ptr addrspace(10) [[I1]] to ptr addrspace(11)
+; CHECK-NEXT: [[I12:%.*]] = load ptr addrspace(13), ptr addrspace(11) [[I10]], align 8
+; CHECK-NEXT: [[I13:%.*]] = load i64, ptr [[I5]], align 8
+; CHECK-NEXT: [[TMP0:%.*]] = shl i64 [[I6]], 3
+; CHECK-NEXT: [[UGLYGEP:%.*]] = getelementptr i8, ptr addrspace(13) [[I12]], i64 [[TMP0]]
+; CHECK-NEXT: [[UGLYGEP1:%.*]] = getelementptr i8, ptr addrspace(13) [[I9]], i64 [[TMP0]]
+; CHECK-NEXT: [[BOUND0:%.*]] = icmp ult ptr addrspace(13) [[I12]], [[UGLYGEP1]]
+; CHECK-NEXT: [[BOUND1:%.*]] = icmp ult ptr addrspace(13) [[I9]], [[UGLYGEP]]
; CHECK-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]
-; CHECK-NEXT: br i1 [[FOUND_CONFLICT]], label [[L74_PH_LVER_ORIG:%.*]], label [[L74_PH:%.*]]
+; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i64 [[I13]], 1
+; CHECK-NEXT: [[LVER_SAFE:%.*]] = or i1 [[FOUND_CONFLICT]], [[IDENT_CHECK]]
+; CHECK-NEXT: br i1 [[LVER_SAFE]], label [[L74_PH_LVER_ORIG:%.*]], label [[L74_PH:%.*]]
; CHECK: L74.ph.lver.orig:
; CHECK-NEXT: br label [[L74_LVER_ORIG:%.*]]
; CHECK: L74.lver.orig:
@@ -49,13 +38,11 @@ define void @test(%jl_value_t addrspace(10)** %arg) {
; CHECK-NEXT: [[VALUE_PHI21_LVER_ORIG:%.*]] = phi i64 [ 1, [[L74_PH_LVER_ORIG]] ], [ [[I22_LVER_ORIG:%.*]], [[L74_LVER_ORIG]] ]
; CHECK-NEXT: [[VALUE_PHI22_LVER_ORIG:%.*]] = phi i64 [ 1, [[L74_PH_LVER_ORIG]] ], [ [[I24_LVER_ORIG:%.*]], [[L74_LVER_ORIG]] ]
; CHECK-NEXT: [[I14_LVER_ORIG:%.*]] = add i64 [[VALUE_PHI21_LVER_ORIG]], -1
-; CHECK-NEXT: [[I15_LVER_ORIG:%.*]] = getelementptr inbounds double, double addrspace(13)* [[I9]], i64 [[I14_LVER_ORIG]]
-; CHECK-NEXT: [[I16_LVER_ORIG:%.*]] = bitcast double addrspace(13)* [[I15_LVER_ORIG]] to i64 addrspace(13)*
-; CHECK-NEXT: [[I17_LVER_ORIG:%.*]] = load i64, i64 addrspace(13)* [[I16_LVER_ORIG]], align 8
+; CHECK-NEXT: [[I15_LVER_ORIG:%.*]] = getelementptr inbounds double, ptr addrspace(13) [[I9]], i64 [[I14_LVER_ORIG]]
+; CHECK-NEXT: [[I17_LVER_ORIG:%.*]] = load i64, ptr addrspace(13) [[I15_LVER_ORIG]], align 8
; CHECK-NEXT: [[I18_LVER_ORIG:%.*]] = add i64 [[VALUE_PHI20_LVER_ORIG]], -1
-; CHECK-NEXT: [[I19_LVER_ORIG:%.*]] = getelementptr inbounds double, double addrspace(13)* [[I12]], i64 [[I18_LVER_ORIG]]
-; CHECK-NEXT: [[I20_LVER_ORIG:%.*]] = bitcast double addrspace(13)* [[I19_LVER_ORIG]] to i64 addrspace(13)*
-; CHECK-NEXT: store i64 [[I17_LVER_ORIG]], i64 addrspace(13)* [[I20_LVER_ORIG]], align 8
+; CHECK-NEXT: [[I19_LVER_ORIG:%.*]] = getelementptr inbounds double, ptr addrspace(13) [[I12]], i64 [[I18_LVER_ORIG]]
+; CHECK-NEXT: store i64 [[I17_LVER_ORIG]], ptr addrspace(13) [[I19_LVER_ORIG]], align 8
; CHECK-NEXT: [[I21_LVER_ORIG]] = add i64 [[VALUE_PHI20_LVER_ORIG]], 1
; CHECK-NEXT: [[I22_LVER_ORIG]] = add i64 [[I13]], [[VALUE_PHI21_LVER_ORIG]]
; CHECK-NEXT: [[I23_LVER_ORIG:%.*]] = icmp eq i64 [[VALUE_PHI22_LVER_ORIG]], [[I6]]
@@ -68,41 +55,36 @@ define void @test(%jl_value_t addrspace(10)** %arg) {
; CHECK-NEXT: [[VALUE_PHI21:%.*]] = phi i64 [ 1, [[L74_PH]] ], [ [[I22:%.*]], [[L74]] ]
; CHECK-NEXT: [[VALUE_PHI22:%.*]] = phi i64 [ 1, [[L74_PH]] ], [ [[I24:%.*]], [[L74]] ]
; CHECK-NEXT: [[I14:%.*]] = add i64 [[VALUE_PHI21]], -1
-; CHECK-NEXT: [[I15:%.*]] = getelementptr inbounds double, double addrspace(13)* [[I9]], i64 [[I14]]
-; CHECK-NEXT: [[I16:%.*]] = bitcast double addrspace(13)* [[I15]] to i64 addrspace(13)*
-; CHECK-NEXT: [[I17:%.*]] = load i64, i64 addrspace(13)* [[I16]], align 8, !alias.scope !0
+; CHECK-NEXT: [[I15:%.*]] = getelementptr inbounds double, ptr addrspace(13) [[I9]], i64 [[I14]]
+; CHECK-NEXT: [[I17:%.*]] = load i64, ptr addrspace(13) [[I15]], align 8, !alias.scope !0
; CHECK-NEXT: [[I18:%.*]] = add i64 [[VALUE_PHI20]], -1
-; CHECK-NEXT: [[I19:%.*]] = getelementptr inbounds double, double addrspace(13)* [[I12]], i64 [[I18]]
-; CHECK-NEXT: [[I20:%.*]] = bitcast double addrspace(13)* [[I19]] to i64 addrspace(13)*
-; CHECK-NEXT: store i64 [[I17]], i64 addrspace(13)* [[I20]], align 8, !alias.scope !3, !noalias !0
+; CHECK-NEXT: [[I19:%.*]] = getelementptr inbounds double, ptr addrspace(13) [[I12]], i64 [[I18]]
+; CHECK-NEXT: store i64 [[I17]], ptr addrspace(13) [[I19]], align 8, !alias.scope !3, !noalias !0
; CHECK-NEXT: [[I21]] = add i64 [[VALUE_PHI20]], 1
; CHECK-NEXT: [[I22]] = add i64 [[I13]], [[VALUE_PHI21]]
; CHECK-NEXT: [[I23:%.*]] = icmp eq i64 [[VALUE_PHI22]], [[I6]]
; CHECK-NEXT: [[I24]] = add i64 [[VALUE_PHI22]], 1
-; CHECK-NEXT: br i1 [[I23]], label [[L94_LOOPEXIT7:%.*]], label [[L74]]
+; CHECK-NEXT: br i1 [[I23]], label [[L94_LOOPEXIT2:%.*]], label [[L74]]
; CHECK: L94.loopexit:
; CHECK-NEXT: br label [[L94:%.*]]
-; CHECK: L94.loopexit7:
+; CHECK: L94.loopexit2:
; CHECK-NEXT: br label [[L94]]
; CHECK: L94:
; CHECK-NEXT: ret void
;
top:
%i = alloca [3 x i64], align 8
- %i1 = load %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %arg, align 8
- %i2 = getelementptr inbounds %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %arg, i64 1
- %i3 = load %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %i2, align 8
- %i4 = getelementptr inbounds [3 x i64], [3 x i64]* %i, i64 0, i64 0
- store i64 1, i64* %i4, align 8
- %i5 = getelementptr inbounds [3 x i64], [3 x i64]* %i, i64 0, i64 1
- %i6 = load i64, i64* inttoptr (i64 24 to i64*), align 8
- %i7 = addrspacecast %jl_value_t addrspace(10)* %i3 to %jl_value_t addrspace(11)*
- %i8 = bitcast %jl_value_t addrspace(11)* %i7 to double addrspace(13)* addrspace(11)*
- %i9 = load double addrspace(13)*, double addrspace(13)* addrspace(11)* %i8, align 8
- %i10 = addrspacecast %jl_value_t addrspace(10)* %i1 to %jl_value_t addrspace(11)*
- %i11 = bitcast %jl_value_t addrspace(11)* %i10 to double addrspace(13)* addrspace(11)*
- %i12 = load double addrspace(13)*, double addrspace(13)* addrspace(11)* %i11, align 8
- %i13 = load i64, i64* %i5, align 8
+ %i1 = load ptr addrspace(10), ptr %arg, align 8
+ %i2 = getelementptr inbounds ptr addrspace(10), ptr %arg, i64 1
+ %i3 = load ptr addrspace(10), ptr %i2, align 8
+ store i64 1, ptr %i, align 8
+ %i5 = getelementptr inbounds [3 x i64], ptr %i, i64 0, i64 1
+ %i6 = load i64, ptr inttoptr (i64 24 to ptr), align 8
+ %i7 = addrspacecast ptr addrspace(10) %i3 to ptr addrspace(11)
+ %i9 = load ptr addrspace(13), ptr addrspace(11) %i7, align 8
+ %i10 = addrspacecast ptr addrspace(10) %i1 to ptr addrspace(11)
+ %i12 = load ptr addrspace(13), ptr addrspace(11) %i10, align 8
+ %i13 = load i64, ptr %i5, align 8
br label %L74
L74: ; preds = %L74, %top
@@ -110,13 +92,11 @@ L74: ; preds = %L74, %top
%value_phi21 = phi i64 [ 1, %top ], [ %i22, %L74 ]
%value_phi22 = phi i64 [ 1, %top ], [ %i24, %L74 ]
%i14 = add i64 %value_phi21, -1
- %i15 = getelementptr inbounds double, double addrspace(13)* %i9, i64 %i14
- %i16 = bitcast double addrspace(13)* %i15 to i64 addrspace(13)*
- %i17 = load i64, i64 addrspace(13)* %i16, align 8
+ %i15 = getelementptr inbounds double, ptr addrspace(13) %i9, i64 %i14
+ %i17 = load i64, ptr addrspace(13) %i15, align 8
%i18 = add i64 %value_phi20, -1
- %i19 = getelementptr inbounds double, double addrspace(13)* %i12, i64 %i18
- %i20 = bitcast double addrspace(13)* %i19 to i64 addrspace(13)*
- store i64 %i17, i64 addrspace(13)* %i20, align 8
+ %i19 = getelementptr inbounds double, ptr addrspace(13) %i12, i64 %i18
+ store i64 %i17, ptr addrspace(13) %i19, align 8
%i21 = add i64 %value_phi20, 1
%i22 = add i64 %i13, %value_phi21
%i23 = icmp eq i64 %value_phi22, %i6
More information about the llvm-commits
mailing list