1; RUN: opt < %s -basicaa -memcpyopt -dse -S | FileCheck -enable-var-scope %s 2 3target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128" 4target triple = "i686-apple-darwin9" 5 6%0 = type { x86_fp80, x86_fp80 } 7%1 = type { i32, i32 } 8 9define void @test1(%0* sret %agg.result, x86_fp80 %z.0, x86_fp80 %z.1) nounwind { 10entry: 11 %tmp2 = alloca %0 12 %memtmp = alloca %0, align 16 13 %tmp5 = fsub x86_fp80 0xK80000000000000000000, %z.1 14 call void @ccoshl(%0* sret %memtmp, x86_fp80 %tmp5, x86_fp80 %z.0) nounwind 15 %tmp219 = bitcast %0* %tmp2 to i8* 16 %memtmp20 = bitcast %0* %memtmp to i8* 17 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %tmp219, i8* align 16 %memtmp20, i32 32, i1 false) 18 %agg.result21 = bitcast %0* %agg.result to i8* 19 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %agg.result21, i8* align 16 %tmp219, i32 32, i1 false) 20 ret void 21 22; Check that one of the memcpy's are removed. 23;; FIXME: PR 8643 We should be able to eliminate the last memcpy here. 24 25; CHECK-LABEL: @test1( 26; CHECK: call void @ccoshl 27; CHECK: call void @llvm.memcpy 28; CHECK-NOT: llvm.memcpy 29; CHECK: ret void 30} 31 32declare void @ccoshl(%0* nocapture sret, x86_fp80, x86_fp80) nounwind 33 34 35; The intermediate alloca and one of the memcpy's should be eliminated, the 36; other should be related with a memmove. 37define void @test2(i8* %P, i8* %Q) nounwind { 38 %memtmp = alloca %0, align 16 39 %R = bitcast %0* %memtmp to i8* 40 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %R, i8* align 16 %P, i32 32, i1 false) 41 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %Q, i8* align 16 %R, i32 32, i1 false) 42 ret void 43 44; CHECK-LABEL: @test2( 45; CHECK-NEXT: call void @llvm.memmove{{.*}}(i8* align 16 %Q, i8* align 16 %P 46; CHECK-NEXT: ret void 47} 48 49; The intermediate alloca and one of the memcpy's should be eliminated, the 50; other should be related with a memcpy. 51define void @test2_memcpy(i8* noalias %P, i8* noalias %Q) nounwind { 52 %memtmp = alloca %0, align 16 53 %R = bitcast %0* %memtmp to i8* 54 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %R, i8* align 16 %P, i32 32, i1 false) 55 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %Q, i8* align 16 %R, i32 32, i1 false) 56 ret void 57 58; CHECK-LABEL: @test2_memcpy( 59; CHECK-NEXT: call void @llvm.memcpy{{.*}}(i8* align 16 %Q, i8* align 16 %P 60; CHECK-NEXT: ret void 61} 62 63 64 65 66@x = external global %0 67 68define void @test3(%0* noalias sret %agg.result) nounwind { 69 %x.0 = alloca %0 70 %x.01 = bitcast %0* %x.0 to i8* 71 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %x.01, i8* align 16 bitcast (%0* @x to i8*), i32 32, i1 false) 72 %agg.result2 = bitcast %0* %agg.result to i8* 73 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %agg.result2, i8* align 16 %x.01, i32 32, i1 false) 74 ret void 75; CHECK-LABEL: @test3( 76; CHECK-NEXT: %agg.result1 = bitcast 77; CHECK-NEXT: call void @llvm.memcpy 78; CHECK-NEXT: ret void 79} 80 81 82; PR8644 83define void @test4(i8 *%P) { 84 %A = alloca %1 85 %a = bitcast %1* %A to i8* 86 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %a, i8* align 4 %P, i64 8, i1 false) 87 call void @test4a(i8* align 1 byval %a) 88 ret void 89; CHECK-LABEL: @test4( 90; CHECK-NEXT: call void @test4a( 91} 92 93; Make sure we don't remove the memcpy if the source address space doesn't match the byval argument 94define void @test4_addrspace(i8 addrspace(1)* %P) { 95 %A = alloca %1 96 %a = bitcast %1* %A to i8* 97 call void @llvm.memcpy.p0i8.p1i8.i64(i8* align 4 %a, i8 addrspace(1)* align 4 %P, i64 8, i1 false) 98 call void @test4a(i8* align 1 byval %a) 99 ret void 100; CHECK-LABEL: @test4_addrspace( 101; CHECK: call void @llvm.memcpy.p0i8.p1i8.i64( 102; CHECK-NEXT: call void @test4a( 103} 104 105declare void @test4a(i8* align 1 byval) 106declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i1) nounwind 107declare void @llvm.memcpy.p0i8.p1i8.i64(i8* nocapture, i8 addrspace(1)* nocapture, i64, i1) nounwind 108declare void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* nocapture, i8 addrspace(1)* nocapture, i64, i1) nounwind 109 110%struct.S = type { i128, [4 x i8]} 111 112@sS = external global %struct.S, align 16 113 114declare void @test5a(%struct.S* align 16 byval) nounwind ssp 115 116 117; rdar://8713376 - This memcpy can't be eliminated. 118define i32 @test5(i32 %x) nounwind ssp { 119entry: 120 %y = alloca %struct.S, align 16 121 %tmp = bitcast %struct.S* %y to i8* 122 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 16 %tmp, i8* align 16 bitcast (%struct.S* @sS to i8*), i64 32, i1 false) 123 %a = getelementptr %struct.S, %struct.S* %y, i64 0, i32 1, i64 0 124 store i8 4, i8* %a 125 call void @test5a(%struct.S* align 16 byval %y) 126 ret i32 0 127 ; CHECK-LABEL: @test5( 128 ; CHECK: store i8 4 129 ; CHECK: call void @test5a(%struct.S* byval align 16 %y) 130} 131 132;; Noop memcpy should be zapped. 133define void @test6(i8 *%P) { 134 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %P, i8* align 4 %P, i64 8, i1 false) 135 ret void 136; CHECK-LABEL: @test6( 137; CHECK-NEXT: ret void 138} 139 140 141; PR9794 - Should forward memcpy into byval argument even though the memcpy 142; isn't itself 8 byte aligned. 143%struct.p = type { i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32 } 144 145define i32 @test7(%struct.p* nocapture align 8 byval %q) nounwind ssp { 146entry: 147 %agg.tmp = alloca %struct.p, align 4 148 %tmp = bitcast %struct.p* %agg.tmp to i8* 149 %tmp1 = bitcast %struct.p* %q to i8* 150 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %tmp, i8* align 4 %tmp1, i64 48, i1 false) 151 %call = call i32 @g(%struct.p* align 8 byval %agg.tmp) nounwind 152 ret i32 %call 153; CHECK-LABEL: @test7( 154; CHECK: call i32 @g(%struct.p* byval align 8 %q) [[$NUW:#[0-9]+]] 155} 156 157declare i32 @g(%struct.p* align 8 byval) 158 159declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i1) nounwind 160 161; PR11142 - When looking for a memcpy-memcpy dependency, don't get stuck on 162; instructions between the memcpy's that only affect the destination pointer. 163@test8.str = internal constant [7 x i8] c"ABCDEF\00" 164 165define void @test8() { 166; CHECK: test8 167; CHECK-NOT: memcpy 168 %A = tail call i8* @malloc(i32 10) 169 %B = getelementptr inbounds i8, i8* %A, i64 2 170 tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* %B, i8* getelementptr inbounds ([7 x i8], [7 x i8]* @test8.str, i64 0, i64 0), i32 7, i1 false) 171 %C = tail call i8* @malloc(i32 10) 172 %D = getelementptr inbounds i8, i8* %C, i64 2 173 tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* %D, i8* %B, i32 7, i1 false) 174 ret void 175; CHECK: ret void 176} 177 178declare noalias i8* @malloc(i32) 179 180; rdar://11341081 181%struct.big = type { [50 x i32] } 182 183define void @test9_addrspacecast() nounwind ssp uwtable { 184entry: 185; CHECK-LABEL: @test9_addrspacecast( 186; CHECK: f1 187; CHECK-NOT: memcpy 188; CHECK: f2 189 %b = alloca %struct.big, align 4 190 %tmp = alloca %struct.big, align 4 191 call void @f1(%struct.big* sret %tmp) 192 %0 = addrspacecast %struct.big* %b to i8 addrspace(1)* 193 %1 = addrspacecast %struct.big* %tmp to i8 addrspace(1)* 194 call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* align 4 %0, i8 addrspace(1)* align 4 %1, i64 200, i1 false) 195 call void @f2(%struct.big* %b) 196 ret void 197} 198 199define void @test9() nounwind ssp uwtable { 200entry: 201; CHECK: test9 202; CHECK: f1 203; CHECK-NOT: memcpy 204; CHECK: f2 205 %b = alloca %struct.big, align 4 206 %tmp = alloca %struct.big, align 4 207 call void @f1(%struct.big* sret %tmp) 208 %0 = bitcast %struct.big* %b to i8* 209 %1 = bitcast %struct.big* %tmp to i8* 210 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %0, i8* align 4 %1, i64 200, i1 false) 211 call void @f2(%struct.big* %b) 212 ret void 213} 214 215; rdar://14073661. 216; Test10 triggered assertion when the compiler try to get the size of the 217; opaque type of *x, where the x is the formal argument with attribute 'sret'. 218 219%opaque = type opaque 220declare void @foo(i32* noalias nocapture) 221 222define void @test10(%opaque* noalias nocapture sret %x, i32 %y) { 223 %a = alloca i32, align 4 224 store i32 %y, i32* %a 225 call void @foo(i32* noalias nocapture %a) 226 %c = load i32, i32* %a 227 %d = bitcast %opaque* %x to i32* 228 store i32 %c, i32* %d 229 ret void 230} 231 232; don't create new addressspacecasts when we don't know they're safe for the target 233define void @test11([20 x i32] addrspace(1)* nocapture dereferenceable(80) %P) { 234 %A = alloca [20 x i32], align 4 235 %a = bitcast [20 x i32]* %A to i8* 236 %b = bitcast [20 x i32] addrspace(1)* %P to i8 addrspace(1)* 237 call void @llvm.memset.p0i8.i64(i8* align 4 %a, i8 0, i64 80, i1 false) 238 call void @llvm.memcpy.p1i8.p0i8.i64(i8 addrspace(1)* align 4 %b, i8* align 4 %a, i64 80, i1 false) 239 ret void 240; CHECK-LABEL: @test11( 241; CHECK-NOT: addrspacecast 242} 243 244declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i1) nounwind 245declare void @llvm.memcpy.p1i8.p0i8.i64(i8 addrspace(1)* nocapture, i8* nocapture, i64, i1) nounwind 246 247declare void @f1(%struct.big* nocapture sret) 248declare void @f2(%struct.big*) 249 250; CHECK: attributes [[$NUW]] = { nounwind } 251; CHECK: attributes #1 = { argmemonly nounwind } 252; CHECK: attributes #2 = { nounwind ssp } 253; CHECK: attributes #3 = { nounwind ssp uwtable } 254