100 lines · plain
1; RUN: llc -mtriple=i686-windows -mattr=+sse2 < %s | FileCheck %s2 3target datalayout = "e-m:w-p:32:32-i64:64-f80:32-n8:16:32-S32"4target triple = "i686-pc-windows-msvc"5 6; There is a conflict between lowering the X86 memory intrinsics and the "base"7; register used to address stack locals. See X86RegisterInfo::hasBaseRegister8; for when this is necessary. Typically, we chose ESI for the base register,9; which all of the X86 string instructions use.10 11declare void @escape_vla_and_icmp(ptr, i1 zeroext)12declare void @llvm.memcpy.p0.p0.i32(ptr nocapture, ptr nocapture readonly, i32, i1)13declare void @llvm.memset.p0.i32(ptr nocapture, i8, i32, i1)14 15define i32 @memcpy_novla_vector(ptr %vp0, ptr %a, ptr %b, i32 %n, i1 zeroext %cond) {16 %foo = alloca <4 x i32>, align 1617 call void @llvm.memcpy.p0.p0.i32(ptr align 4 %a, ptr align 4 %b, i32 128, i1 false)18 br i1 %cond, label %spill_vectors, label %no_vectors19 20no_vectors:21 ret i32 022 23spill_vectors:24 %vp1 = getelementptr <4 x i32>, ptr %vp0, i32 125 %v0 = load <4 x i32>, ptr %vp026 %v1 = load <4 x i32>, ptr %vp127 %vicmp = icmp slt <4 x i32> %v0, %v128 %icmp = extractelement <4 x i1> %vicmp, i32 029 call void @escape_vla_and_icmp(ptr null, i1 zeroext %icmp)30 %r = extractelement <4 x i32> %v0, i32 031 ret i32 %r32}33 34; CHECK-LABEL: _memcpy_novla_vector:35; CHECK: andl $-16, %esp36; CHECK-DAG: movl $32, %ecx37; CHECK-DAG: movl {{.*}}, %esi38; CHECK-DAG: movl {{.*}}, %edi39; CHECK: rep;movsl40 41define i32 @memcpy_vla_vector(ptr %vp0, ptr %a, ptr %b, i32 %n, i1 zeroext %cond) {42 %foo = alloca <4 x i32>, align 1643 call void @llvm.memcpy.p0.p0.i32(ptr align 4 %a, ptr align 4 %b, i32 128, i1 false)44 br i1 %cond, label %spill_vectors, label %no_vectors45 46no_vectors:47 ret i32 048 49spill_vectors:50 %vp1 = getelementptr <4 x i32>, ptr %vp0, i32 151 %v0 = load <4 x i32>, ptr %vp052 %v1 = load <4 x i32>, ptr %vp153 %vicmp = icmp slt <4 x i32> %v0, %v154 %icmp = extractelement <4 x i1> %vicmp, i32 055 %vla = alloca i8, i32 %n56 call void @escape_vla_and_icmp(ptr %vla, i1 zeroext %icmp)57 %r = extractelement <4 x i32> %v0, i32 058 ret i32 %r59}60 61; CHECK-LABEL: _memcpy_vla_vector:62; CHECK: andl $-16, %esp63; CHECK: movl %esp, %esi64; CHECK: pushl $12865; CHECK: calll _memcpy66; CHECK: calll __chkstk67 68; stosd doesn't clobber esi, so we can use it.69 70define i32 @memset_vla_vector(ptr %vp0, ptr %a, i32 %n, i1 zeroext %cond) {71 %foo = alloca <4 x i32>, align 1672 call void @llvm.memset.p0.i32(ptr align 4 %a, i8 42, i32 128, i1 false)73 br i1 %cond, label %spill_vectors, label %no_vectors74 75no_vectors:76 ret i32 077 78spill_vectors:79 %vp1 = getelementptr <4 x i32>, ptr %vp0, i32 180 %v0 = load <4 x i32>, ptr %vp081 %v1 = load <4 x i32>, ptr %vp182 %vicmp = icmp slt <4 x i32> %v0, %v183 %icmp = extractelement <4 x i1> %vicmp, i32 084 %vla = alloca i8, i32 %n85 call void @escape_vla_and_icmp(ptr %vla, i1 zeroext %icmp)86 %r = extractelement <4 x i32> %v0, i32 087 ret i32 %r88}89 90; CHECK-LABEL: _memset_vla_vector:91; CHECK: andl $-16, %esp92; CHECK: movl %esp, %esi93; CHECK-DAG: movl $707406378, %eax # imm = 0x2A2A2A2A94; CHECK-DAG: movl $32, %ecx95; CHECK-DAG: movl {{.*}}, %edi96; CHECK-NOT: movl {{.*}}, %esi97; CHECK: rep;stosl98 99; Add a test for memcmp if we ever add a special lowering for it.100