]> granicus.if.org Git - llvm/commitdiff
[X86] Lower avx512f scatter intrinsics to X86MaskedScatterSDNode instead of going...
authorCraig Topper <craig.topper@intel.com>
Fri, 18 Jan 2019 20:14:46 +0000 (20:14 +0000)
committerCraig Topper <craig.topper@intel.com>
Fri, 18 Jan 2019 20:14:46 +0000 (20:14 +0000)
This sends these intrinsics through isel in a much more normal way. This should allow addressing mode matching in isel to make better use of the displacement field.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@351583 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/X86/X86ISelLowering.cpp
lib/Target/X86/X86IntrinsicsInfo.h

index d4179993f02c2ed92703c4184842407ad8b7c16c..5150dec943271c1cc02df0817e2eb8d79a6c32e4 100644 (file)
@@ -4822,6 +4822,17 @@ bool X86TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
     Info.flags |= MachineMemOperand::MOLoad;
     break;
   }
+  case SCATTER: {
+    Info.ptrVal = nullptr;
+    MVT DataVT = MVT::getVT(I.getArgOperand(3)->getType());
+    MVT IndexVT = MVT::getVT(I.getArgOperand(2)->getType());
+    unsigned NumElts = std::min(DataVT.getVectorNumElements(),
+                                IndexVT.getVectorNumElements());
+    Info.memVT = MVT::getVectorVT(DataVT.getVectorElementType(), NumElts);
+    Info.align = 1;
+    Info.flags |= MachineMemOperand::MOStore;
+    break;
+  }
   default:
     return false;
   }
@@ -22449,8 +22460,6 @@ static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
   if (!C)
     return SDValue();
   SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
-  SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
-  SDValue Segment = DAG.getRegister(0, MVT::i32);
   unsigned MinElts = std::min(Index.getSimpleValueType().getVectorNumElements(),
                               Src.getSimpleValueType().getVectorNumElements());
   MVT MaskVT = MVT::getVectorVT(MVT::i1, MinElts);
@@ -22460,10 +22469,13 @@ static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
   if (Mask.getValueType() != MaskVT)
     Mask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
 
+  MemIntrinsicSDNode *MemIntr = cast<MemIntrinsicSDNode>(Op);
+
   SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
-  SDValue Ops[] = {Base, Scale, Index, Disp, Segment, Mask, Src, Chain};
-  SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
-  return SDValue(Res, 1);
+  SDValue Ops[] = {Chain, Src, Mask, Base, Index, Scale};
+  SDValue Res = DAG.getTargetMemSDNode<X86MaskedScatterSDNode>(
+      VTs, Ops, dl, MemIntr->getMemoryVT(), MemIntr->getMemOperand());
+  return Res.getValue(1);
 }
 
 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
index 46f71609e82eef668efc552b6cc4086ed8e4dfe0..1604c690a19fd8660a71b9e46e9f0332edccb09c 100644 (file)
@@ -249,47 +249,47 @@ static const IntrinsicData IntrinsicsWithChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_pmovus_wb_mem_512, TRUNCATE_TO_MEM_VI8,
                      X86ISD::VTRUNCUS, 0),
 
-  X86_INTRINSIC_DATA(avx512_mask_scatter_dpd_512, SCATTER, X86::VSCATTERDPDZmr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatter_dpi_512, SCATTER, X86::VPSCATTERDDZmr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatter_dpq_512, SCATTER, X86::VPSCATTERDQZmr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatter_dps_512, SCATTER, X86::VSCATTERDPSZmr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatter_qpd_512, SCATTER, X86::VSCATTERQPDZmr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatter_qpi_512, SCATTER, X86::VPSCATTERQDZmr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatter_qpq_512, SCATTER, X86::VPSCATTERQQZmr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatter_qps_512, SCATTER, X86::VSCATTERQPSZmr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatterdiv2_df, SCATTER, X86::VSCATTERQPDZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatterdiv2_di, SCATTER, X86::VPSCATTERQQZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatterdiv4_df, SCATTER, X86::VSCATTERQPDZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatterdiv4_di, SCATTER, X86::VPSCATTERQQZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatterdiv4_sf, SCATTER, X86::VSCATTERQPSZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatterdiv4_si, SCATTER, X86::VPSCATTERQDZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatterdiv8_sf, SCATTER, X86::VSCATTERQPSZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scatterdiv8_si, SCATTER, X86::VPSCATTERQDZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scattersiv2_df, SCATTER, X86::VSCATTERDPDZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scattersiv2_di, SCATTER, X86::VPSCATTERDQZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scattersiv4_df, SCATTER, X86::VSCATTERDPDZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scattersiv4_di, SCATTER, X86::VPSCATTERDQZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scattersiv4_sf, SCATTER, X86::VSCATTERDPSZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scattersiv4_si, SCATTER, X86::VPSCATTERDDZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scattersiv8_sf, SCATTER, X86::VSCATTERDPSZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_mask_scattersiv8_si, SCATTER, X86::VPSCATTERDDZ256mr, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatter_dpd_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatter_dpi_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatter_dpq_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatter_dps_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatter_qpd_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatter_qpi_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatter_qpq_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatter_qps_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatterdiv2_df, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatterdiv2_di, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatterdiv4_df, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatterdiv4_di, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatterdiv4_sf, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatterdiv4_si, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatterdiv8_sf, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scatterdiv8_si, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scattersiv2_df, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scattersiv2_di, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scattersiv4_df, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scattersiv4_di, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scattersiv4_sf, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scattersiv4_si, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scattersiv8_sf, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_mask_scattersiv8_si, SCATTER, 0, 0),
 
-  X86_INTRINSIC_DATA(avx512_scatter_dpd_512, SCATTER, X86::VSCATTERDPDZmr, 0),
-  X86_INTRINSIC_DATA(avx512_scatter_dpi_512, SCATTER, X86::VPSCATTERDDZmr, 0),
-  X86_INTRINSIC_DATA(avx512_scatter_dpq_512, SCATTER, X86::VPSCATTERDQZmr, 0),
-  X86_INTRINSIC_DATA(avx512_scatter_dps_512, SCATTER, X86::VSCATTERDPSZmr, 0),
-  X86_INTRINSIC_DATA(avx512_scatter_qpd_512, SCATTER, X86::VSCATTERQPDZmr, 0),
-  X86_INTRINSIC_DATA(avx512_scatter_qpi_512, SCATTER, X86::VPSCATTERQDZmr, 0),
-  X86_INTRINSIC_DATA(avx512_scatter_qpq_512, SCATTER, X86::VPSCATTERQQZmr, 0),
-  X86_INTRINSIC_DATA(avx512_scatter_qps_512, SCATTER, X86::VSCATTERQPSZmr, 0),
-  X86_INTRINSIC_DATA(avx512_scatterdiv2_df, SCATTER, X86::VSCATTERQPDZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_scatterdiv2_di, SCATTER, X86::VPSCATTERQQZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_scatterdiv4_df, SCATTER, X86::VSCATTERQPDZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_scatterdiv4_di, SCATTER, X86::VPSCATTERQQZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_scatterdiv4_sf, SCATTER, X86::VSCATTERQPSZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_scatterdiv4_si, SCATTER, X86::VPSCATTERQDZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_scatterdiv8_sf, SCATTER, X86::VSCATTERQPSZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_scatterdiv8_si, SCATTER, X86::VPSCATTERQDZ256mr, 0),
+  X86_INTRINSIC_DATA(avx512_scatter_dpd_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatter_dpi_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatter_dpq_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatter_dps_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatter_qpd_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatter_qpi_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatter_qpq_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatter_qps_512, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatterdiv2_df, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatterdiv2_di, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatterdiv4_df, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatterdiv4_di, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatterdiv4_sf, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatterdiv4_si, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatterdiv8_sf, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scatterdiv8_si, SCATTER, 0, 0),
   X86_INTRINSIC_DATA(avx512_scatterpf_dpd_512, PREFETCH, X86::VSCATTERPF0DPDm,
                      X86::VSCATTERPF1DPDm),
   X86_INTRINSIC_DATA(avx512_scatterpf_dps_512, PREFETCH, X86::VSCATTERPF0DPSm,
@@ -298,14 +298,14 @@ static const IntrinsicData IntrinsicsWithChain[] = {
                      X86::VSCATTERPF1QPDm),
   X86_INTRINSIC_DATA(avx512_scatterpf_qps_512, PREFETCH, X86::VSCATTERPF0QPSm,
                      X86::VSCATTERPF1QPSm),
-  X86_INTRINSIC_DATA(avx512_scattersiv2_df, SCATTER, X86::VSCATTERDPDZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_scattersiv2_di, SCATTER, X86::VPSCATTERDQZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_scattersiv4_df, SCATTER, X86::VSCATTERDPDZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_scattersiv4_di, SCATTER, X86::VPSCATTERDQZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_scattersiv4_sf, SCATTER, X86::VSCATTERDPSZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_scattersiv4_si, SCATTER, X86::VPSCATTERDDZ128mr, 0),
-  X86_INTRINSIC_DATA(avx512_scattersiv8_sf, SCATTER, X86::VSCATTERDPSZ256mr, 0),
-  X86_INTRINSIC_DATA(avx512_scattersiv8_si, SCATTER, X86::VPSCATTERDDZ256mr, 0),
+  X86_INTRINSIC_DATA(avx512_scattersiv2_df, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scattersiv2_di, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scattersiv4_df, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scattersiv4_di, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scattersiv4_sf, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scattersiv4_si, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scattersiv8_sf, SCATTER, 0, 0),
+  X86_INTRINSIC_DATA(avx512_scattersiv8_si, SCATTER, 0, 0),
   X86_INTRINSIC_DATA(rdpmc,     RDPMC,  X86ISD::RDPMC_DAG, 0),
   X86_INTRINSIC_DATA(rdrand_16, RDRAND, X86ISD::RDRAND, 0),
   X86_INTRINSIC_DATA(rdrand_32, RDRAND, X86ISD::RDRAND, 0),