1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -memcpyopt -S -enable-memcpyopt-memoryssa=0 | FileCheck %s --check-prefix=NO_MSSA 3; RUN: opt < %s -memcpyopt -S -enable-memcpyopt-memoryssa=1 -verify-memoryssa | FileCheck %s --check-prefix=MSSA 4 5; Test whether memcpy-memcpy dependence is optimized across 6; basic blocks (conditional branches and invokes). 7; TODO: This is not supported yet. 8 9%struct.s = type { i32, i32 } 10 11@s_foo = private unnamed_addr constant %struct.s { i32 1, i32 2 }, align 4 12@s_baz = private unnamed_addr constant %struct.s { i32 1, i32 2 }, align 4 13@i = external constant i8* 14 15declare void @qux() 16declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture writeonly, i8* nocapture readonly, i64, i1) 17declare void @__cxa_throw(i8*, i8*, i8*) 18declare i32 @__gxx_personality_v0(...) 19declare i8* @__cxa_begin_catch(i8*) 20 21; A simple partial redundancy. Test that the second memcpy is optimized 22; to copy directly from the original source rather than from the temporary. 23 24define void @wobble(i8* noalias %dst, i8* %src, i1 %some_condition) { 25; NO_MSSA-LABEL: @wobble( 26; NO_MSSA-NEXT: bb: 27; NO_MSSA-NEXT: [[TEMP:%.*]] = alloca i8, i32 64, align 1 28; NO_MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TEMP]], i8* nonnull align 8 [[SRC:%.*]], i64 64, i1 false) 29; NO_MSSA-NEXT: br i1 [[SOME_CONDITION:%.*]], label [[MORE:%.*]], label [[OUT:%.*]] 30; NO_MSSA: out: 31; NO_MSSA-NEXT: call void @qux() 32; NO_MSSA-NEXT: unreachable 33; NO_MSSA: more: 34; NO_MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[DST:%.*]], i8* align 8 [[TEMP]], i64 64, i1 false) 35; NO_MSSA-NEXT: ret void 36; 37; MSSA-LABEL: @wobble( 38; MSSA-NEXT: bb: 39; MSSA-NEXT: [[TEMP:%.*]] = alloca i8, i32 64, align 1 40; MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TEMP]], i8* nonnull align 8 [[SRC:%.*]], i64 64, i1 false) 41; MSSA-NEXT: br i1 [[SOME_CONDITION:%.*]], label [[MORE:%.*]], label [[OUT:%.*]] 42; MSSA: out: 43; MSSA-NEXT: call void @qux() 44; MSSA-NEXT: unreachable 45; MSSA: more: 46; MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[DST:%.*]], i8* align 8 [[SRC]], i64 64, i1 false) 47; MSSA-NEXT: ret void 48; 49bb: 50 %temp = alloca i8, i32 64 51 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %temp, i8* nonnull align 8%src, i64 64, i1 false) 52 br i1 %some_condition, label %more, label %out 53 54out: 55 call void @qux() 56 unreachable 57 58more: 59 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %dst, i8* align 8 %temp, i64 64, i1 false) 60 ret void 61} 62 63; A CFG triangle with a partial redundancy targeting an alloca. Test that the 64; memcpy inside the triangle is optimized to copy directly from the original 65; source rather than from the temporary. 66 67define i32 @foo(i1 %t3) { 68; NO_MSSA-LABEL: @foo( 69; NO_MSSA-NEXT: bb: 70; NO_MSSA-NEXT: [[S:%.*]] = alloca [[STRUCT_S:%.*]], align 4 71; NO_MSSA-NEXT: [[T:%.*]] = alloca [[STRUCT_S]], align 4 72; NO_MSSA-NEXT: [[S1:%.*]] = bitcast %struct.s* [[S]] to i8* 73; NO_MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[S1]], i8* align 4 bitcast (%struct.s* @s_foo to i8*), i64 8, i1 false) 74; NO_MSSA-NEXT: br i1 [[T3:%.*]], label [[BB4:%.*]], label [[BB7:%.*]] 75; NO_MSSA: bb4: 76; NO_MSSA-NEXT: [[T5:%.*]] = bitcast %struct.s* [[T]] to i8* 77; NO_MSSA-NEXT: [[S6:%.*]] = bitcast %struct.s* [[S]] to i8* 78; NO_MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[T5]], i8* align 4 [[S6]], i64 8, i1 false) 79; NO_MSSA-NEXT: br label [[BB7]] 80; NO_MSSA: bb7: 81; NO_MSSA-NEXT: [[T8:%.*]] = getelementptr [[STRUCT_S]], %struct.s* [[T]], i32 0, i32 0 82; NO_MSSA-NEXT: [[T9:%.*]] = load i32, i32* [[T8]], align 4 83; NO_MSSA-NEXT: [[T10:%.*]] = getelementptr [[STRUCT_S]], %struct.s* [[T]], i32 0, i32 1 84; NO_MSSA-NEXT: [[T11:%.*]] = load i32, i32* [[T10]], align 4 85; NO_MSSA-NEXT: [[T12:%.*]] = add i32 [[T9]], [[T11]] 86; NO_MSSA-NEXT: ret i32 [[T12]] 87; 88; MSSA-LABEL: @foo( 89; MSSA-NEXT: bb: 90; MSSA-NEXT: [[S:%.*]] = alloca [[STRUCT_S:%.*]], align 4 91; MSSA-NEXT: [[T:%.*]] = alloca [[STRUCT_S]], align 4 92; MSSA-NEXT: [[S1:%.*]] = bitcast %struct.s* [[S]] to i8* 93; MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[S1]], i8* align 4 bitcast (%struct.s* @s_foo to i8*), i64 8, i1 false) 94; MSSA-NEXT: br i1 [[T3:%.*]], label [[BB4:%.*]], label [[BB7:%.*]] 95; MSSA: bb4: 96; MSSA-NEXT: [[T5:%.*]] = bitcast %struct.s* [[T]] to i8* 97; MSSA-NEXT: [[S6:%.*]] = bitcast %struct.s* [[S]] to i8* 98; MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[T5]], i8* align 4 bitcast (%struct.s* @s_foo to i8*), i64 8, i1 false) 99; MSSA-NEXT: br label [[BB7]] 100; MSSA: bb7: 101; MSSA-NEXT: [[T8:%.*]] = getelementptr [[STRUCT_S]], %struct.s* [[T]], i32 0, i32 0 102; MSSA-NEXT: [[T9:%.*]] = load i32, i32* [[T8]], align 4 103; MSSA-NEXT: [[T10:%.*]] = getelementptr [[STRUCT_S]], %struct.s* [[T]], i32 0, i32 1 104; MSSA-NEXT: [[T11:%.*]] = load i32, i32* [[T10]], align 4 105; MSSA-NEXT: [[T12:%.*]] = add i32 [[T9]], [[T11]] 106; MSSA-NEXT: ret i32 [[T12]] 107; 108bb: 109 %s = alloca %struct.s, align 4 110 %t = alloca %struct.s, align 4 111 %s1 = bitcast %struct.s* %s to i8* 112 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %s1, i8* align 4 bitcast (%struct.s* @s_foo to i8*), i64 8, i1 false) 113 br i1 %t3, label %bb4, label %bb7 114 115bb4: ; preds = %bb 116 %t5 = bitcast %struct.s* %t to i8* 117 %s6 = bitcast %struct.s* %s to i8* 118 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %t5, i8* align 4 %s6, i64 8, i1 false) 119 br label %bb7 120 121bb7: ; preds = %bb4, %bb 122 %t8 = getelementptr %struct.s, %struct.s* %t, i32 0, i32 0 123 %t9 = load i32, i32* %t8, align 4 124 %t10 = getelementptr %struct.s, %struct.s* %t, i32 0, i32 1 125 %t11 = load i32, i32* %t10, align 4 126 %t12 = add i32 %t9, %t11 127 ret i32 %t12 128} 129 130; A CFG diamond with an invoke on one side, and a partially redundant memcpy 131; into an alloca on the other. Test that the memcpy inside the diamond is 132; optimized to copy ; directly from the original source rather than from the 133; temporary. This more complex test represents a relatively common usage 134; pattern. 135 136define i32 @baz(i1 %t5) personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) { 137; NO_MSSA-LABEL: @baz( 138; NO_MSSA-NEXT: bb: 139; NO_MSSA-NEXT: [[S:%.*]] = alloca [[STRUCT_S:%.*]], align 4 140; NO_MSSA-NEXT: [[T:%.*]] = alloca [[STRUCT_S]], align 4 141; NO_MSSA-NEXT: [[S3:%.*]] = bitcast %struct.s* [[S]] to i8* 142; NO_MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[S3]], i8* align 4 bitcast (%struct.s* @s_baz to i8*), i64 8, i1 false) 143; NO_MSSA-NEXT: br i1 [[T5:%.*]], label [[BB6:%.*]], label [[BB22:%.*]] 144; NO_MSSA: bb6: 145; NO_MSSA-NEXT: invoke void @__cxa_throw(i8* null, i8* bitcast (i8** @i to i8*), i8* null) 146; NO_MSSA-NEXT: to label [[BB25:%.*]] unwind label [[BB9:%.*]] 147; NO_MSSA: bb9: 148; NO_MSSA-NEXT: [[T10:%.*]] = landingpad { i8*, i32 } 149; NO_MSSA-NEXT: catch i8* null 150; NO_MSSA-NEXT: br label [[BB13:%.*]] 151; NO_MSSA: bb13: 152; NO_MSSA-NEXT: [[T15:%.*]] = call i8* @__cxa_begin_catch(i8* null) 153; NO_MSSA-NEXT: br label [[BB23:%.*]] 154; NO_MSSA: bb22: 155; NO_MSSA-NEXT: [[T23:%.*]] = bitcast %struct.s* [[T]] to i8* 156; NO_MSSA-NEXT: [[S24:%.*]] = bitcast %struct.s* [[S]] to i8* 157; NO_MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[T23]], i8* align 4 [[S24]], i64 8, i1 false) 158; NO_MSSA-NEXT: br label [[BB23]] 159; NO_MSSA: bb23: 160; NO_MSSA-NEXT: [[T17:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[T]], i32 0, i32 0 161; NO_MSSA-NEXT: [[T18:%.*]] = load i32, i32* [[T17]], align 4 162; NO_MSSA-NEXT: [[T19:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[T]], i32 0, i32 1 163; NO_MSSA-NEXT: [[T20:%.*]] = load i32, i32* [[T19]], align 4 164; NO_MSSA-NEXT: [[T21:%.*]] = add nsw i32 [[T18]], [[T20]] 165; NO_MSSA-NEXT: ret i32 [[T21]] 166; NO_MSSA: bb25: 167; NO_MSSA-NEXT: unreachable 168; 169; MSSA-LABEL: @baz( 170; MSSA-NEXT: bb: 171; MSSA-NEXT: [[S:%.*]] = alloca [[STRUCT_S:%.*]], align 4 172; MSSA-NEXT: [[T:%.*]] = alloca [[STRUCT_S]], align 4 173; MSSA-NEXT: [[S3:%.*]] = bitcast %struct.s* [[S]] to i8* 174; MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[S3]], i8* align 4 bitcast (%struct.s* @s_baz to i8*), i64 8, i1 false) 175; MSSA-NEXT: br i1 [[T5:%.*]], label [[BB6:%.*]], label [[BB22:%.*]] 176; MSSA: bb6: 177; MSSA-NEXT: invoke void @__cxa_throw(i8* null, i8* bitcast (i8** @i to i8*), i8* null) 178; MSSA-NEXT: to label [[BB25:%.*]] unwind label [[BB9:%.*]] 179; MSSA: bb9: 180; MSSA-NEXT: [[T10:%.*]] = landingpad { i8*, i32 } 181; MSSA-NEXT: catch i8* null 182; MSSA-NEXT: br label [[BB13:%.*]] 183; MSSA: bb13: 184; MSSA-NEXT: [[T15:%.*]] = call i8* @__cxa_begin_catch(i8* null) 185; MSSA-NEXT: br label [[BB23:%.*]] 186; MSSA: bb22: 187; MSSA-NEXT: [[T23:%.*]] = bitcast %struct.s* [[T]] to i8* 188; MSSA-NEXT: [[S24:%.*]] = bitcast %struct.s* [[S]] to i8* 189; MSSA-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[T23]], i8* align 4 bitcast (%struct.s* @s_baz to i8*), i64 8, i1 false) 190; MSSA-NEXT: br label [[BB23]] 191; MSSA: bb23: 192; MSSA-NEXT: [[T17:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[T]], i32 0, i32 0 193; MSSA-NEXT: [[T18:%.*]] = load i32, i32* [[T17]], align 4 194; MSSA-NEXT: [[T19:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[T]], i32 0, i32 1 195; MSSA-NEXT: [[T20:%.*]] = load i32, i32* [[T19]], align 4 196; MSSA-NEXT: [[T21:%.*]] = add nsw i32 [[T18]], [[T20]] 197; MSSA-NEXT: ret i32 [[T21]] 198; MSSA: bb25: 199; MSSA-NEXT: unreachable 200; 201bb: 202 %s = alloca %struct.s, align 4 203 %t = alloca %struct.s, align 4 204 %s3 = bitcast %struct.s* %s to i8* 205 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %s3, i8* align 4 bitcast (%struct.s* @s_baz to i8*), i64 8, i1 false) 206 br i1 %t5, label %bb6, label %bb22 207 208bb6: ; preds = %bb 209 invoke void @__cxa_throw(i8* null, i8* bitcast (i8** @i to i8*), i8* null) 210 to label %bb25 unwind label %bb9 211 212bb9: ; preds = %bb6 213 %t10 = landingpad { i8*, i32 } 214 catch i8* null 215 br label %bb13 216 217bb13: ; preds = %bb9 218 %t15 = call i8* @__cxa_begin_catch(i8* null) 219 br label %bb23 220 221bb22: ; preds = %bb 222 %t23 = bitcast %struct.s* %t to i8* 223 %s24 = bitcast %struct.s* %s to i8* 224 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %t23, i8* align 4 %s24, i64 8, i1 false) 225 br label %bb23 226 227bb23: ; preds = %bb22, %bb13 228 %t17 = getelementptr inbounds %struct.s, %struct.s* %t, i32 0, i32 0 229 %t18 = load i32, i32* %t17, align 4 230 %t19 = getelementptr inbounds %struct.s, %struct.s* %t, i32 0, i32 1 231 %t20 = load i32, i32* %t19, align 4 232 %t21 = add nsw i32 %t18, %t20 233 ret i32 %t21 234 235bb25: ; preds = %bb6 236 unreachable 237} 238