1; RUN: opt < %s -basicaa -memcpyopt -dse -S | FileCheck %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* %tmp219, i8* %memtmp20, i32 32, i32 16, i1 false) 18 %agg.result21 = bitcast %0* %agg.result to i8* 19 call void @llvm.memcpy.p0i8.p0i8.i32(i8* %agg.result21, i8* %tmp219, i32 32, i32 16, 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* 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* %R, i8* %P, i32 32, i32 16, i1 false) 41 call void @llvm.memcpy.p0i8.p0i8.i32(i8* %Q, i8* %R, i32 32, i32 16, i1 false) 42 ret void 43 44; CHECK-LABEL: @test2( 45; CHECK-NEXT: call void @llvm.memmove{{.*}}(i8* %Q, i8* %P 46; CHECK-NEXT: ret void 47} 48 49 50 51 52@x = external global %0 53 54define void @test3(%0* noalias sret %agg.result) nounwind { 55 %x.0 = alloca %0 56 %x.01 = bitcast %0* %x.0 to i8* 57 call void @llvm.memcpy.p0i8.p0i8.i32(i8* %x.01, i8* bitcast (%0* @x to i8*), i32 32, i32 16, i1 false) 58 %agg.result2 = bitcast %0* %agg.result to i8* 59 call void @llvm.memcpy.p0i8.p0i8.i32(i8* %agg.result2, i8* %x.01, i32 32, i32 16, i1 false) 60 ret void 61; CHECK-LABEL: @test3( 62; CHECK-NEXT: %agg.result1 = bitcast 63; CHECK-NEXT: call void @llvm.memcpy 64; CHECK-NEXT: ret void 65} 66 67 68; PR8644 69define void @test4(i8 *%P) { 70 %A = alloca %1 71 %a = bitcast %1* %A to i8* 72 call void @llvm.memcpy.p0i8.p0i8.i64(i8* %a, i8* %P, i64 8, i32 4, i1 false) 73 call void @test4a(i8* align 1 byval %a) 74 ret void 75; CHECK-LABEL: @test4( 76; CHECK-NEXT: call void @test4a( 77} 78 79declare void @test4a(i8* align 1 byval) 80declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i32, i1) nounwind 81 82%struct.S = type { i128, [4 x i8]} 83 84@sS = external global %struct.S, align 16 85 86declare void @test5a(%struct.S* align 16 byval) nounwind ssp 87 88 89; rdar://8713376 - This memcpy can't be eliminated. 90define i32 @test5(i32 %x) nounwind ssp { 91entry: 92 %y = alloca %struct.S, align 16 93 %tmp = bitcast %struct.S* %y to i8* 94 call void @llvm.memcpy.p0i8.p0i8.i64(i8* %tmp, i8* bitcast (%struct.S* @sS to i8*), i64 32, i32 16, i1 false) 95 %a = getelementptr %struct.S* %y, i64 0, i32 1, i64 0 96 store i8 4, i8* %a 97 call void @test5a(%struct.S* align 16 byval %y) 98 ret i32 0 99 ; CHECK-LABEL: @test5( 100 ; CHECK: store i8 4 101 ; CHECK: call void @test5a(%struct.S* byval align 16 %y) 102} 103 104;; Noop memcpy should be zapped. 105define void @test6(i8 *%P) { 106 call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %P, i64 8, i32 4, i1 false) 107 ret void 108; CHECK-LABEL: @test6( 109; CHECK-NEXT: ret void 110} 111 112 113; PR9794 - Should forward memcpy into byval argument even though the memcpy 114; isn't itself 8 byte aligned. 115%struct.p = type { i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32 } 116 117define i32 @test7(%struct.p* nocapture align 8 byval %q) nounwind ssp { 118entry: 119 %agg.tmp = alloca %struct.p, align 4 120 %tmp = bitcast %struct.p* %agg.tmp to i8* 121 %tmp1 = bitcast %struct.p* %q to i8* 122 call void @llvm.memcpy.p0i8.p0i8.i64(i8* %tmp, i8* %tmp1, i64 48, i32 4, i1 false) 123 %call = call i32 @g(%struct.p* align 8 byval %agg.tmp) nounwind 124 ret i32 %call 125; CHECK-LABEL: @test7( 126; CHECK: call i32 @g(%struct.p* byval align 8 %q) [[NUW:#[0-9]+]] 127} 128 129declare i32 @g(%struct.p* align 8 byval) 130 131declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind 132 133; PR11142 - When looking for a memcpy-memcpy dependency, don't get stuck on 134; instructions between the memcpy's that only affect the destination pointer. 135@test8.str = internal constant [7 x i8] c"ABCDEF\00" 136 137define void @test8() { 138; CHECK: test8 139; CHECK-NOT: memcpy 140 %A = tail call i8* @malloc(i32 10) 141 %B = getelementptr inbounds i8* %A, i64 2 142 tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* %B, i8* getelementptr inbounds ([7 x i8]* @test8.str, i64 0, i64 0), i32 7, i32 1, i1 false) 143 %C = tail call i8* @malloc(i32 10) 144 %D = getelementptr inbounds i8* %C, i64 2 145 tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* %D, i8* %B, i32 7, i32 1, i1 false) 146 ret void 147; CHECK: ret void 148} 149 150declare noalias i8* @malloc(i32) 151 152; rdar://11341081 153%struct.big = type { [50 x i32] } 154 155define void @test9() nounwind ssp uwtable { 156entry: 157; CHECK: test9 158; CHECK: f1 159; CHECK-NOT: memcpy 160; CHECK: f2 161 %b = alloca %struct.big, align 4 162 %tmp = alloca %struct.big, align 4 163 call void @f1(%struct.big* sret %tmp) 164 %0 = bitcast %struct.big* %b to i8* 165 %1 = bitcast %struct.big* %tmp to i8* 166 call void @llvm.memcpy.p0i8.p0i8.i64(i8* %0, i8* %1, i64 200, i32 4, i1 false) 167 call void @f2(%struct.big* %b) 168 ret void 169} 170 171; rdar://14073661. 172; Test10 triggered assertion when the compiler try to get the size of the 173; opaque type of *x, where the x is the formal argument with attribute 'sret'. 174 175%opaque = type opaque 176declare void @foo(i32* noalias nocapture) 177 178define void @test10(%opaque* noalias nocapture sret %x, i32 %y) { 179 %a = alloca i32, align 4 180 store i32 %y, i32* %a 181 call void @foo(i32* noalias nocapture %a) 182 %c = load i32* %a 183 %d = bitcast %opaque* %x to i32* 184 store i32 %c, i32* %d 185 ret void 186} 187 188declare void @f1(%struct.big* sret) 189declare void @f2(%struct.big*) 190 191; CHECK: attributes [[NUW]] = { nounwind } 192; CHECK: attributes #1 = { nounwind ssp } 193; CHECK: attributes #2 = { nounwind ssp uwtable } 194