From cbc2c76b286c38a1fd006543b6b224c06cd96df1 Mon Sep 17 00:00:00 2001 From: Craig Topper Date: Sun, 8 Oct 2017 23:49:29 +0000 Subject: Merging r313366: ------------------------------------------------------------------------ r313366 | ctopper | 2017-09-15 10:09:03 -0700 (Fri, 15 Sep 2017) | 9 lines [X86] Don't create i64 constants on 32-bit targets when lowering v64i1 constant build vectors When handling a v64i1 build vector of constants on 32-bit targets we were creating an illegal i64 constant that we then bitcasted back to v64i1. We need to instead create two 32-bit constants, bitcast them to v32i1 and concat the result. We should also take care to handle the halves being all zeros/ones after the split. This patch splits the build vector and then recursively lowers the two pieces. This allows us to handle the all ones and all zeros cases with minimal effort. Ideally we'd just do the split and concat, and let lowering get called again on the new nodes, but getNode has special handling for CONCAT_VECTORS that reassembles the pieces back into a single BUILD_VECTOR. Hopefully the two temporary BUILD_VECTORS we had to create to do this that don't get returned don't cause any issues. Fixes PR34605. Differential Revision: https://reviews.llvm.org/D37858 ------------------------------------------------------------------------ git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_50@315198 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/X86/X86ISelLowering.cpp | 12 ++++++++ test/CodeGen/X86/pr34605.ll | 63 ++++++++++++++++++++++++++++++++++++++ 2 files changed, 75 insertions(+) create mode 100644 test/CodeGen/X86/pr34605.ll diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp index 957b46c40a6e..2255c5dcc4f3 100644 --- a/lib/Target/X86/X86ISelLowering.cpp +++ b/lib/Target/X86/X86ISelLowering.cpp @@ -7026,6 +7026,18 @@ X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const { return DAG.getTargetConstant(1, dl, VT); if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) { + if (VT == MVT::v64i1 && !Subtarget.is64Bit()) { + // Split the pieces. + SDValue Lower = + DAG.getBuildVector(MVT::v32i1, dl, Op.getNode()->ops().slice(0, 32)); + SDValue Upper = + DAG.getBuildVector(MVT::v32i1, dl, Op.getNode()->ops().slice(32, 32)); + // We have to manually lower both halves so getNode doesn't try to + // reassemble the build_vector. + Lower = LowerBUILD_VECTORvXi1(Lower, DAG); + Upper = LowerBUILD_VECTORvXi1(Upper, DAG); + return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lower, Upper); + } SDValue Imm = ConvertI1VectorToInteger(Op, DAG); if (Imm.getValueSizeInBits() == VT.getSizeInBits()) return DAG.getBitcast(VT, Imm); diff --git a/test/CodeGen/X86/pr34605.ll b/test/CodeGen/X86/pr34605.ll new file mode 100644 index 000000000000..8330511d6449 --- /dev/null +++ b/test/CodeGen/X86/pr34605.ll @@ -0,0 +1,63 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py +; RUN: llc < %s -mtriple=i386-unknown-linux-gnu -mattr=avx512bw,avx512vl,avx512dq | FileCheck %s + +define void @pr34605(i8* nocapture %s, i32 %p) { +; CHECK-LABEL: pr34605: +; CHECK: # BB#0: # %entry +; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax +; CHECK-NEXT: vpbroadcastd {{[0-9]+}}(%esp), %zmm0 +; CHECK-NEXT: vpcmpeqd {{\.LCPI.*}}, %zmm0, %k0 +; CHECK-NEXT: vpcmpeqd {{\.LCPI.*}}, %zmm0, %k1 +; CHECK-NEXT: kunpckwd %k0, %k1, %k0 +; CHECK-NEXT: vpcmpeqd {{\.LCPI.*}}, %zmm0, %k1 +; CHECK-NEXT: vpcmpeqd {{\.LCPI.*}}, %zmm0, %k2 +; CHECK-NEXT: kunpckwd %k1, %k2, %k1 +; CHECK-NEXT: kunpckdq %k0, %k1, %k0 +; CHECK-NEXT: kxord %k0, %k0, %k1 +; CHECK-NEXT: movl $1, %ecx +; CHECK-NEXT: kmovd %ecx, %k2 +; CHECK-NEXT: kunpckdq %k2, %k1, %k1 +; CHECK-NEXT: kandq %k1, %k0, %k1 +; CHECK-NEXT: vmovdqu8 {{\.LCPI.*}}, %zmm0 {%k1} {z} +; CHECK-NEXT: vmovdqu8 %zmm0, (%eax) +; CHECK-NEXT: vpxord %zmm0, %zmm0, %zmm0 +; CHECK-NEXT: vmovdqu32 %zmm0, 64(%eax) +; CHECK-NEXT: vmovdqu32 %zmm0, 128(%eax) +; CHECK-NEXT: vmovdqu32 %zmm0, 192(%eax) +; CHECK-NEXT: vmovdqu32 %zmm0, 256(%eax) +; CHECK-NEXT: vmovdqu32 %zmm0, 320(%eax) +; CHECK-NEXT: vmovdqu32 %zmm0, 384(%eax) +; CHECK-NEXT: vmovdqu32 %zmm0, 448(%eax) +; CHECK-NEXT: vzeroupper +; CHECK-NEXT: retl +entry: + %broadcast.splatinsert = insertelement <64 x i32> undef, i32 %p, i32 0 + %broadcast.splat = shufflevector <64 x i32> %broadcast.splatinsert, <64 x i32> undef, <64 x i32> zeroinitializer + %0 = icmp eq <64 x i32> %broadcast.splat, + %1 = and <64 x i1> %0, + %2 = zext <64 x i1> %1 to <64 x i8> + %3 = bitcast i8* %s to <64 x i8>* + store <64 x i8> %2, <64 x i8>* %3, align 1 + %4 = getelementptr inbounds i8, i8* %s, i32 64 + %5 = bitcast i8* %4 to <64 x i8>* + store <64 x i8> zeroinitializer, <64 x i8>* %5, align 1 + %6 = getelementptr inbounds i8, i8* %s, i32 128 + %7 = bitcast i8* %6 to <64 x i8>* + store <64 x i8> zeroinitializer, <64 x i8>* %7, align 1 + %8 = getelementptr inbounds i8, i8* %s, i32 192 + %9 = bitcast i8* %8 to <64 x i8>* + store <64 x i8> zeroinitializer, <64 x i8>* %9, align 1 + %10 = getelementptr inbounds i8, i8* %s, i32 256 + %11 = bitcast i8* %10 to <64 x i8>* + store <64 x i8> zeroinitializer, <64 x i8>* %11, align 1 + %12 = getelementptr inbounds i8, i8* %s, i32 320 + %13 = bitcast i8* %12 to <64 x i8>* + store <64 x i8> zeroinitializer, <64 x i8>* %13, align 1 + %14 = getelementptr inbounds i8, i8* %s, i32 384 + %15 = bitcast i8* %14 to <64 x i8>* + store <64 x i8> zeroinitializer, <64 x i8>* %15, align 1 + %16 = getelementptr inbounds i8, i8* %s, i32 448 + %17 = bitcast i8* %16 to <64 x i8>* + store <64 x i8> zeroinitializer, <64 x i8>* %17, align 1 + ret void +} -- cgit v1.2.3