| ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py | |
| ; RUN: opt -passes=memcpyopt -S < %s -verify-memoryssa | FileCheck %s | |
| target datalayout = "e-i64:64-f80:128-n8:16:32:64" | |
| target triple = "x86_64-unknown-linux-gnu" | |
| %S = type { ptr, i8, i32 } | |
| define void @copy(ptr %src, ptr %dst) { | |
| ; CHECK-LABEL: @copy( | |
| ; CHECK-NEXT: call void @llvm.memmove.p0.p0.i64(ptr align 8 [[DST:%.*]], ptr align 8 [[SRC:%.*]], i64 16, i1 false) | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %1 = load %S, ptr %src | |
| store %S %1, ptr %dst | |
| ret void | |
| } | |
| define void @noaliassrc(ptr noalias %src, ptr %dst) { | |
| ; CHECK-LABEL: @noaliassrc( | |
| ; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[DST:%.*]], ptr align 8 [[SRC:%.*]], i64 16, i1 false) | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %1 = load %S, ptr %src | |
| store %S %1, ptr %dst | |
| ret void | |
| } | |
| define void @noaliasdst(ptr %src, ptr noalias %dst) { | |
| ; CHECK-LABEL: @noaliasdst( | |
| ; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[DST:%.*]], ptr align 8 [[SRC:%.*]], i64 16, i1 false) | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %1 = load %S, ptr %src | |
| store %S %1, ptr %dst | |
| ret void | |
| } | |
| define void @destroysrc(ptr %src, ptr %dst) { | |
| ; CHECK-LABEL: @destroysrc( | |
| ; CHECK-NEXT: [[TMP1:%.*]] = load [[S:%.*]], ptr [[SRC:%.*]], align 8 | |
| ; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr align 8 [[SRC]], i8 0, i64 16, i1 false) | |
| ; CHECK-NEXT: store [[S]] [[TMP1]], ptr [[DST:%.*]], align 8 | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %1 = load %S, ptr %src | |
| store %S zeroinitializer, ptr %src | |
| store %S %1, ptr %dst | |
| ret void | |
| } | |
| define void @destroynoaliassrc(ptr noalias %src, ptr %dst) { | |
| ; CHECK-LABEL: @destroynoaliassrc( | |
| ; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[DST:%.*]], ptr align 8 [[SRC:%.*]], i64 16, i1 false) | |
| ; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr align 8 [[SRC]], i8 0, i64 16, i1 false) | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %1 = load %S, ptr %src | |
| store %S zeroinitializer, ptr %src | |
| store %S %1, ptr %dst | |
| ret void | |
| } | |
| define void @copyalias(ptr %src, ptr %dst) { | |
| ; CHECK-LABEL: @copyalias( | |
| ; CHECK-NEXT: [[TMP1:%.*]] = load [[S:%.*]], ptr [[SRC:%.*]], align 8 | |
| ; CHECK-NEXT: call void @llvm.memmove.p0.p0.i64(ptr align 8 [[DST:%.*]], ptr align 8 [[SRC]], i64 16, i1 false) | |
| ; CHECK-NEXT: store [[S]] [[TMP1]], ptr [[DST]], align 8 | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %1 = load %S, ptr %src | |
| %2 = load %S, ptr %src | |
| store %S %1, ptr %dst | |
| store %S %2, ptr %dst | |
| ret void | |
| } | |
| ; If the store address is computed in a complex manner, make | |
| ; sure we lift the computation as well if needed and possible. | |
| define void @addrproducer(ptr %src, ptr %dst) { | |
| ; CHECK-LABEL: @addrproducer( | |
| ; CHECK-NEXT: [[DST2:%.*]] = getelementptr [[S:%.*]], ptr [[DST:%.*]], i64 1 | |
| ; CHECK-NEXT: call void @llvm.memmove.p0.p0.i64(ptr align 8 [[DST2]], ptr align 8 [[SRC:%.*]], i64 16, i1 false) | |
| ; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr align 8 [[DST]], i8 undef, i64 16, i1 false) | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %1 = load %S, ptr %src | |
| store %S undef, ptr %dst | |
| %dst2 = getelementptr %S , ptr %dst, i64 1 | |
| store %S %1, ptr %dst2 | |
| ret void | |
| } | |
| define void @aliasaddrproducer(ptr %src, ptr %dst, ptr %dstidptr) { | |
| ; CHECK-LABEL: @aliasaddrproducer( | |
| ; CHECK-NEXT: [[TMP1:%.*]] = load [[S:%.*]], ptr [[SRC:%.*]], align 8 | |
| ; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr align 8 [[DST:%.*]], i8 undef, i64 16, i1 false) | |
| ; CHECK-NEXT: [[DSTINDEX:%.*]] = load i32, ptr [[DSTIDPTR:%.*]], align 4 | |
| ; CHECK-NEXT: [[DST2:%.*]] = getelementptr [[S]], ptr [[DST]], i32 [[DSTINDEX]] | |
| ; CHECK-NEXT: store [[S]] [[TMP1]], ptr [[DST2]], align 8 | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %1 = load %S, ptr %src | |
| store %S undef, ptr %dst | |
| %dstindex = load i32, ptr %dstidptr | |
| %dst2 = getelementptr %S , ptr %dst, i32 %dstindex | |
| store %S %1, ptr %dst2 | |
| ret void | |
| } | |
| define void @noaliasaddrproducer(ptr %src, ptr noalias %dst, ptr noalias %dstidptr) { | |
| ; CHECK-LABEL: @noaliasaddrproducer( | |
| ; CHECK-NEXT: [[TMP2:%.*]] = load i32, ptr [[DSTIDPTR:%.*]], align 4 | |
| ; CHECK-NEXT: [[DSTINDEX:%.*]] = or i32 [[TMP2]], 1 | |
| ; CHECK-NEXT: [[DST2:%.*]] = getelementptr [[S:%.*]], ptr [[DST:%.*]], i32 [[DSTINDEX]] | |
| ; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[DST2]], ptr align 8 [[SRC:%.*]], i64 16, i1 false) | |
| ; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr align 8 [[SRC]], i8 undef, i64 16, i1 false) | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %1 = load %S, ptr %src | |
| store %S undef, ptr %src | |
| %2 = load i32, ptr %dstidptr | |
| %dstindex = or i32 %2, 1 | |
| %dst2 = getelementptr %S , ptr %dst, i32 %dstindex | |
| store %S %1, ptr %dst2 | |
| ret void | |
| } | |
| define void @throwing_call(ptr noalias %src, ptr %dst) { | |
| ; CHECK-LABEL: @throwing_call( | |
| ; CHECK-NEXT: [[TMP1:%.*]] = load [[S:%.*]], ptr [[SRC:%.*]], align 8 | |
| ; CHECK-NEXT: call void @llvm.memset.p0.i64(ptr align 8 [[SRC]], i8 0, i64 16, i1 false) | |
| ; CHECK-NEXT: call void @call() #[[ATTR2:[0-9]+]] | |
| ; CHECK-NEXT: store [[S]] [[TMP1]], ptr [[DST:%.*]], align 8 | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %1 = load %S, ptr %src | |
| store %S zeroinitializer, ptr %src | |
| call void @call() readnone | |
| store %S %1, ptr %dst | |
| ret void | |
| } | |
| define void @loop_memoryphi(ptr %a, ptr %b) { | |
| ; CHECK-LABEL: @loop_memoryphi( | |
| ; CHECK-NEXT: br label [[LOOP:%.*]] | |
| ; CHECK: loop: | |
| ; CHECK-NEXT: call void @llvm.memmove.p0.p0.i64(ptr align 8 [[B:%.*]], ptr align 8 [[A:%.*]], i64 16, i1 false) | |
| ; CHECK-NEXT: br label [[LOOP]] | |
| ; | |
| br label %loop | |
| loop: | |
| %v = load { i64, i64 }, ptr %a | |
| store { i64, i64 } %v, ptr %b | |
| br label %loop | |
| } | |
| ; There are multiple instructions that can clobber the source memory here. | |
| ; We can move the dest write past the store to %ptr.24, but not the memcpy. | |
| ; Make sure we don't perform fca2memcpy conversion in this case. | |
| define void @multiple_clobbering(ptr %ptr, ptr %ptr.copy) { | |
| ; CHECK-LABEL: @multiple_clobbering( | |
| ; CHECK-NEXT: [[PTR_8:%.*]] = getelementptr inbounds nuw i8, ptr [[PTR:%.*]], i64 8 | |
| ; CHECK-NEXT: [[PTR_24:%.*]] = getelementptr inbounds nuw i8, ptr [[PTR]], i64 24 | |
| ; CHECK-NEXT: [[PTR_32:%.*]] = getelementptr inbounds nuw i8, ptr [[PTR]], i64 32 | |
| ; CHECK-NEXT: [[PTR_COPY_8:%.*]] = getelementptr inbounds nuw i8, ptr [[PTR_COPY:%.*]], i64 8 | |
| ; CHECK-NEXT: [[STRUCT:%.*]] = load { i32, i64 }, ptr [[PTR_COPY_8]], align 8 | |
| ; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr [[PTR_8]], ptr [[PTR_32]], i64 12, i1 false) | |
| ; CHECK-NEXT: store i64 1, ptr [[PTR_24]], align 8 | |
| ; CHECK-NEXT: store { i32, i64 } [[STRUCT]], ptr [[PTR_32]], align 8 | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %ptr.8 = getelementptr inbounds nuw i8, ptr %ptr, i64 8 | |
| %ptr.24 = getelementptr inbounds nuw i8, ptr %ptr, i64 24 | |
| %ptr.32 = getelementptr inbounds nuw i8, ptr %ptr, i64 32 | |
| %ptr.copy.8 = getelementptr inbounds nuw i8, ptr %ptr.copy, i64 8 | |
| %struct = load { i32, i64 }, ptr %ptr.copy.8, align 8 | |
| call void @llvm.memcpy.p0.p0.i64(ptr %ptr.8, ptr %ptr.32, i64 12, i1 false) | |
| store i64 1, ptr %ptr.24, align 8 | |
| store { i32, i64 } %struct, ptr %ptr.32, align 8 | |
| ret void | |
| } | |
| ; This is a regression test for miscompilation due to improper EEA invalidation. | |
| ; | |
| ; The pass first merges load %s -> %v and store %v -> dst into memcpy %s -> %dst. | |
| ; In order to do this before %s is overwritten, definition of %dst has to move | |
| ; above the call to @esc. | |
| ; | |
| ; Previously, with the invalid cache, EEA still considered the call to @esc as | |
| ; the earliest escape of %a making AA incorrectly conclude that | |
| ; memcpy %s -> %dst can't write to %a. | |
| ; | |
| ; That led to memcpy %a -> %out being replaced with memcpy %orig -> %out, which | |
| ; is a miscompilation. | |
| define void @addrproducer_captures(i64 %x, i64 %y, ptr %out) { | |
| ; CHECK-LABEL: @addrproducer_captures( | |
| ; CHECK-NEXT: [[A:%.*]] = alloca { i32, i32 }, align 8 | |
| ; CHECK-NEXT: [[ORIG:%.*]] = alloca i64, align 8 | |
| ; CHECK-NEXT: [[S:%.*]] = alloca i64, align 8 | |
| ; CHECK-NEXT: store i64 [[X:%.*]], ptr [[ORIG]], align 8 | |
| ; CHECK-NEXT: store i64 [[Y:%.*]], ptr [[S]], align 8 | |
| ; CHECK-NEXT: call void @call() | |
| ; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr [[A]], ptr [[ORIG]], i64 8, i1 false) | |
| ; CHECK-NEXT: [[DST:%.*]] = call ptr @opaque(ptr [[A]]) #[[ATTR3:[0-9]+]] | |
| ; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 4 [[DST]], ptr align 4 [[S]], i64 8, i1 false) | |
| ; CHECK-NEXT: store i32 0, ptr [[S]], align 4 | |
| ; CHECK-NEXT: call void @esc(ptr [[A]]) #[[ATTR4:[0-9]+]] | |
| ; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr [[OUT:%.*]], ptr [[A]], i64 8, i1 false) | |
| ; CHECK-NEXT: ret void | |
| ; | |
| %a = alloca { i32, i32 }, align 8 | |
| %orig = alloca i64, align 8 | |
| %s = alloca i64, align 8 | |
| store i64 %x, ptr %orig, align 8 | |
| store i64 %y, ptr %s, align 8 | |
| call void @call() | |
| call void @llvm.memcpy.p0.p0.i64(ptr %a, ptr %orig, i64 8, i1 false) | |
| %v = load { i32, i32 }, ptr %s, align 4 | |
| store i32 0, ptr %s, align 4 | |
| call void @esc(ptr %a) memory(inaccessiblemem: write) nounwind willreturn | |
| %dst = call ptr @opaque(ptr %a) memory(none) nounwind willreturn | |
| store { i32, i32 } %v, ptr %dst, align 4 | |
| ; this memcpy should remain %a -> %out | |
| call void @llvm.memcpy.p0.p0.i64(ptr %out, ptr %a, i64 8, i1 false) | |
| ret void | |
| } | |
| declare void @call() | |
| declare void @esc(ptr) | |
| declare ptr @opaque(ptr) |