AMDGPU/GlobalISel: RegBankSelect for amdgcn.writelane

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@364808 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
index f6d8a1a20d77ef07c0fc29e0a13e64ea704d1a4c..c20b7976a4b154b37df1b15490dbff9262dba509 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ b/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -158,7 +158,25 @@ AMDGPURegisterBankInfo::getInstrAlternativeMappingsIntrinsic(
     const std::array<unsigned, 3> RegSrcOpIdx = { { 0, 2, 3 } };
     return addMappingFromTable<3>(MI, MRI, RegSrcOpIdx, makeArrayRef(Table));
   }
+  case Intrinsic::amdgcn_writelane: {
+    static const OpRegBankEntry<4> Table[4] = {
+      // Perfectly legal.
+      { { AMDGPU::VGPRRegBankID, AMDGPU::SGPRRegBankID, AMDGPU::SGPRRegBankID, AMDGPU::VGPRRegBankID }, 1 },
+
+      // Need readfirstlane of first op
+      { { AMDGPU::VGPRRegBankID, AMDGPU::VGPRRegBankID, AMDGPU::SGPRRegBankID, AMDGPU::VGPRRegBankID }, 2 },
+
+      // Need readfirstlane of second op
+      { { AMDGPU::VGPRRegBankID, AMDGPU::SGPRRegBankID, AMDGPU::VGPRRegBankID, AMDGPU::VGPRRegBankID }, 2 },
 
+      // Need readfirstlane of both ops
+      { { AMDGPU::VGPRRegBankID, AMDGPU::VGPRRegBankID, AMDGPU::VGPRRegBankID, AMDGPU::VGPRRegBankID }, 3 }
+    };
+
+    // rsrc, voffset, offset
+    const std::array<unsigned, 4> RegSrcOpIdx = { { 0, 2, 3, 4 } };
+    return addMappingFromTable<4>(MI, MRI, RegSrcOpIdx, makeArrayRef(Table));
+  }
   default:
     return RegisterBankInfo::getInstrAlternativeMappings(MI);
   }
@@ -764,6 +782,17 @@ void AMDGPURegisterBankInfo::constrainOpWithReadfirstlane(
   MI.getOperand(OpIdx).setReg(SGPR);
 }
 
+// For cases where only a single copy is inserted for matching register banks.
+// Replace the register in the instruction operand
+static void substituteSimpleCopyRegs(
+  const AMDGPURegisterBankInfo::OperandsMapper &OpdMapper, unsigned OpIdx) {
+  SmallVector<unsigned, 1> SrcReg(OpdMapper.getVRegs(OpIdx));
+  if (!SrcReg.empty()) {
+    assert(SrcReg.size() == 1);
+    OpdMapper.getMI().getOperand(OpIdx).setReg(SrcReg[0]);
+  }
+}
+
 void AMDGPURegisterBankInfo::applyMappingImpl(
     const OperandsMapper &OpdMapper) const {
   MachineInstr &MI = OpdMapper.getMI();
@@ -982,16 +1011,23 @@ void AMDGPURegisterBankInfo::applyMappingImpl(
       return;
     }
     case Intrinsic::amdgcn_readlane: {
-      SmallVector<unsigned, 1> SrcReg(OpdMapper.getVRegs(2));
+      substituteSimpleCopyRegs(OpdMapper, 2);
 
-      if (!SrcReg.empty()) {
-        assert(SrcReg.size() == 1);
-        MI.getOperand(2).setReg(SrcReg[0]);
-      }
+      assert(empty(OpdMapper.getVRegs(0)));
+      assert(empty(OpdMapper.getVRegs(3)));
 
+      // Make sure the index is an SGPR. It doesn't make sense to run this in a
+      // waterfall loop, so assume it's a uniform value.
+      constrainOpWithReadfirstlane(MI, MRI, 3); // Index
+      return;
+    }
+    case Intrinsic::amdgcn_writelane: {
       assert(empty(OpdMapper.getVRegs(0)));
+      assert(empty(OpdMapper.getVRegs(2)));
       assert(empty(OpdMapper.getVRegs(3)));
 
+      substituteSimpleCopyRegs(OpdMapper, 4); // VGPR input val
+      constrainOpWithReadfirstlane(MI, MRI, 2); // Source value
       constrainOpWithReadfirstlane(MI, MRI, 3); // Index
       return;
     }
@@ -1664,6 +1700,23 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
       OpdsMapping[2] = AMDGPU::getValueMapping(AMDGPU::VGPRRegBankID, SrcSize);
       break;
     }
+    case Intrinsic::amdgcn_writelane: {
+      unsigned DstSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
+      unsigned SrcReg = MI.getOperand(2).getReg();
+      unsigned SrcSize = MRI.getType(SrcReg).getSizeInBits();
+      unsigned SrcBank = getRegBankID(SrcReg, MRI, *TRI, AMDGPU::SGPRRegBankID);
+      unsigned IdxReg = MI.getOperand(3).getReg();
+      unsigned IdxSize = MRI.getType(IdxReg).getSizeInBits();
+      unsigned IdxBank = getRegBankID(IdxReg, MRI, *TRI, AMDGPU::SGPRRegBankID);
+      OpdsMapping[0] = AMDGPU::getValueMapping(AMDGPU::VGPRRegBankID, DstSize);
+
+      // These 2 must be SGPRs, but accept VGPRs. Readfirstlane will be inserted
+      // to legalize.
+      OpdsMapping[2] = AMDGPU::getValueMapping(SrcBank, SrcSize);
+      OpdsMapping[3] = AMDGPU::getValueMapping(IdxBank, IdxSize);
+      OpdsMapping[4] = AMDGPU::getValueMapping(AMDGPU::VGPRRegBankID, SrcSize);
+      break;
+    }
     }
     break;
   }

diff --git a/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn.writelane.mir b/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn.writelane.mir
new file mode 100644 (file)
index 0000000..669cfe8
--- /dev/null
+++ b/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn.writelane.mir
@@ -0,0 +1,98 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck %s
+# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck %s
+
+---
+name: writelane_sss
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $sgpr0, $sgpr1, $sgpr2
+    ; CHECK-LABEL: name: writelane_sss
+    ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
+    ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
+    ; CHECK: [[COPY2:%[0-9]+]]:sgpr(s32) = COPY $sgpr2
+    ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[COPY2]](s32)
+    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.writelane), [[COPY]](s32), [[COPY1]](s32), [[COPY3]](s32)
+    %0:_(s32) = COPY $sgpr0
+    %1:_(s32) = COPY $sgpr1
+    %2:_(s32) = COPY $sgpr2
+    %3:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.writelane), %0, %1, %2
+...
+
+---
+name: writelane_ssv
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $sgpr0, $sgpr1, $vgpr0
+    ; CHECK-LABEL: name: writelane_ssv
+    ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
+    ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr1
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.writelane), [[COPY]](s32), [[COPY1]](s32), [[COPY2]](s32)
+    %0:_(s32) = COPY $sgpr0
+    %1:_(s32) = COPY $sgpr1
+    %2:_(s32) = COPY $vgpr0
+    %3:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.writelane), %0, %1, %2
+...
+
+---
+name: writelane_vsv
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $sgpr0, $vgpr0, $vgpr1
+    ; CHECK-LABEL: name: writelane_vsv
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr_32(s32) = COPY $vgpr0
+    ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY]](s32), implicit $exec
+    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.writelane), [[V_READFIRSTLANE_B32_]], [[COPY1]](s32), [[COPY2]](s32)
+    %0:_(s32) = COPY $vgpr0
+    %1:_(s32) = COPY $sgpr0
+    %2:_(s32) = COPY $vgpr1
+    %3:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.writelane), %0, %1, %2
+...
+
+---
+name: writelane_vvv
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $vgpr0, $vgpr1
+    ; CHECK-LABEL: name: writelane_vvv
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr_32(s32) = COPY $vgpr0
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr_32(s32) = COPY $vgpr1
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr2
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY]](s32), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]](s32), implicit $exec
+    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.writelane), [[V_READFIRSTLANE_B32_]], [[V_READFIRSTLANE_B32_1]], [[COPY2]](s32)
+    %0:_(s32) = COPY $vgpr0
+    %1:_(s32) = COPY $vgpr1
+    %2:_(s32) = COPY $vgpr2
+    %3:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.writelane), %0, %1, %2
+...
+
+---
+name: writelane_svv
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $sgpr0, $vgpr0, $vgpr1
+    ; CHECK-LABEL: name: writelane_svv
+    ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr_32(s32) = COPY $vgpr0
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY $vgpr1
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]](s32), implicit $exec
+    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.writelane), [[COPY]](s32), [[V_READFIRSTLANE_B32_]], [[COPY2]](s32)
+    %0:_(s32) = COPY $sgpr0
+    %1:_(s32) = COPY $vgpr0
+    %2:_(s32) = COPY $vgpr1
+    %3:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.writelane), %0, %1, %2
+...