[GlobalISel][X86] Support vector type G_MERGE_VALUES selection.

Summary:
Support vector type G_MERGE_VALUES selection. For now G_MERGE_VALUES marked as legal for any type, so nothing to do in legalizer.
Split from https://reviews.llvm.org/D33665

Reviewers: qcolombet, t.p.northover, zvi, guyblank

Reviewed By: guyblank

Subscribers: rovka, kristof.beyls, guyblank, llvm-commits

Differential Revision: https://reviews.llvm.org/D33958

llvm-svn: 306665

3 files changed, 179 insertions, 0 deletions

diff --git a/llvm/lib/Target/X86/X86InstructionSelector.cpp b/llvm/lib/Target/X86/X86InstructionSelector.cpp
index f98c2a7e802..e34a90e975b 100644
--- a/llvm/lib/Target/X86/X86InstructionSelector.cpp
+++ b/llvm/lib/Target/X86/X86InstructionSelector.cpp
@@ -75,6 +75,8 @@ private:
   bool selectUadde(MachineInstr &I, MachineRegisterInfo &MRI,
                    MachineFunction &MF) const;
   bool selectCopy(MachineInstr &I, MachineRegisterInfo &MRI) const;
+  bool selectMergeValues(MachineInstr &I, MachineRegisterInfo &MRI,
+                         MachineFunction &MF) const;
   bool selectInsert(MachineInstr &I, MachineRegisterInfo &MRI,
                     MachineFunction &MF) const;
   bool selectExtract(MachineInstr &I, MachineRegisterInfo &MRI,
@@ -270,6 +272,8 @@ bool X86InstructionSelector::select(MachineInstr &I) const {
     return true;
   if (selectUadde(I, MRI, MF))
     return true;
+  if (selectMergeValues(I, MRI, MF))
+    return true;
   if (selectExtract(I, MRI, MF))
     return true;
   if (selectInsert(I, MRI, MF))
@@ -914,6 +918,55 @@ bool X86InstructionSelector::selectInsert(MachineInstr &I,
   return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
 }
 
+bool X86InstructionSelector::selectMergeValues(MachineInstr &I,
+                                               MachineRegisterInfo &MRI,
+                                               MachineFunction &MF) const {
+  if (I.getOpcode() != TargetOpcode::G_MERGE_VALUES)
+    return false;
+
+  // Split to inserts.
+  unsigned DstReg = I.getOperand(0).getReg();
+  unsigned SrcReg0 = I.getOperand(1).getReg();
+
+  const LLT DstTy = MRI.getType(DstReg);
+  const LLT SrcTy = MRI.getType(SrcReg0);
+  unsigned SrcSize = SrcTy.getSizeInBits();
+
+  const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI);
+
+  // For the first src use insertSubReg.
+  unsigned DefReg = MRI.createGenericVirtualRegister(DstTy);
+  MRI.setRegBank(DefReg, RegBank);
+  if (!emitInsertSubreg(DefReg, I.getOperand(1).getReg(), I, MRI, MF))
+    return false;
+
+  for (unsigned Idx = 2; Idx < I.getNumOperands(); ++Idx) {
+
+    unsigned Tmp = MRI.createGenericVirtualRegister(DstTy);
+    MRI.setRegBank(Tmp, RegBank);
+
+    MachineInstr &InsertInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
+                                        TII.get(TargetOpcode::G_INSERT), Tmp)
+                                    .addReg(DefReg)
+                                    .addReg(I.getOperand(Idx).getReg())
+                                    .addImm((Idx - 1) * SrcSize);
+
+    DefReg = Tmp;
+
+    if (!select(InsertInst))
+      return false;
+  }
+
+  MachineInstr &CopyInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
+                                    TII.get(TargetOpcode::COPY), DstReg)
+                                .addReg(DefReg);
+
+  if (!select(CopyInst))
+    return false;
+
+  I.eraseFromParent();
+  return true;
+}
 InstructionSelector *
 llvm::createX86InstructionSelector(const X86TargetMachine &TM,
                                    X86Subtarget &Subtarget,
diff --git a/llvm/test/CodeGen/X86/GlobalISel/select-merge-vec256.mir b/llvm/test/CodeGen/X86/GlobalISel/select-merge-vec256.mir
new file mode 100644
index 00000000000..8e31a904e36
--- /dev/null
+++ b/llvm/test/CodeGen/X86/GlobalISel/select-merge-vec256.mir
@@ -0,0 +1,52 @@
+# RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx               -global-isel -run-pass=instruction-select -verify-machineinstrs %s -o - | FileCheck %s --check-prefix=AVX
+# RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx512f,+avx512vl -global-isel -run-pass=instruction-select -verify-machineinstrs %s -o - | FileCheck %s --check-prefix=AVX512VL
+--- |
+  define void @test_merge() {
+    ret void
+  }
+...
+---
+name:            test_merge
+# AVX-LABEL: name:  test_merge
+#
+# AVX512VL-LABEL: name:  test_merge
+alignment:       4
+legalized:       true
+regBankSelected: true
+# AVX:           registers:
+# AVX-NEXT:        - { id: 0, class: vr128, preferred-register: '' }
+# AVX-NEXT:        - { id: 1, class: vr256, preferred-register: '' }
+# AVX-NEXT:        - { id: 2, class: vr256, preferred-register: '' }
+# AVX-NEXT:        - { id: 3, class: vr256, preferred-register: '' }
+#
+# AVX512VL:      registers:
+# AVX512VL-NEXT:   - { id: 0, class: vr128x, preferred-register: '' }
+# AVX512VL-NEXT:   - { id: 1, class: vr256x, preferred-register: '' }
+# AVX512VL-NEXT:   - { id: 2, class: vr256x, preferred-register: '' }
+# AVX512VL-NEXT:   - { id: 3, class: vr256x, preferred-register: '' }
+registers:
+  - { id: 0, class: vecr }
+  - { id: 1, class: vecr }
+# AVX:               %0 = IMPLICIT_DEF
+# AVX-NEXT:          undef %2.sub_xmm = COPY %0
+# AVX-NEXT:          %3 = VINSERTF128rr %2, %0, 1
+# AVX-NEXT:          %1 = COPY %3
+# AVX-NEXT:          %ymm0 = COPY %1
+# AVX-NEXT:          RET 0, implicit %ymm0
+#
+# AVX512VL:          %0 = IMPLICIT_DEF
+# AVX512VL-NEXT:     undef %2.sub_xmm = COPY %0
+# AVX512VL-NEXT:     %3 = VINSERTF32x4Z256rr %2, %0, 1
+# AVX512VL-NEXT:     %1 = COPY %3
+# AVX512VL-NEXT:     %ymm0 = COPY %1
+# AVX512VL-NEXT:     RET 0, implicit %ymm0
+body:             |
+  bb.1 (%ir-block.0):
+
+    %0(<4 x s32>) = IMPLICIT_DEF
+    %1(<8 x s32>) = G_MERGE_VALUES %0(<4 x s32>), %0(<4 x s32>)
+    %ymm0 = COPY %1(<8 x s32>)
+    RET 0, implicit %ymm0
+
+...
+
diff --git a/llvm/test/CodeGen/X86/GlobalISel/select-merge-vec512.mir b/llvm/test/CodeGen/X86/GlobalISel/select-merge-vec512.mir
new file mode 100644
index 00000000000..a072d582e50
--- /dev/null
+++ b/llvm/test/CodeGen/X86/GlobalISel/select-merge-vec512.mir
@@ -0,0 +1,74 @@
+# RUN: llc -mtriple=x86_64-linux-gnu -mattr=+avx512f -global-isel -run-pass=instruction-select -verify-machineinstrs %s -o - | FileCheck %s --check-prefix=ALL
+--- |
+  define void @test_merge_v128() {
+    ret void
+  }
+
+  define void @test_merge_v256() {
+    ret void
+  }
+
+...
+---
+name:            test_merge_v128
+# ALL-LABEL: name:  test_merge_v128
+alignment:       4
+legalized:       true
+regBankSelected: true
+# ALL:      registers:
+# ALL-NEXT:   - { id: 0, class: vr128x, preferred-register: '' }
+# ALL-NEXT:   - { id: 1, class: vr512, preferred-register: '' }
+# ALL-NEXT:   - { id: 2, class: vr512, preferred-register: '' }
+# ALL-NEXT:   - { id: 3, class: vr512, preferred-register: '' }
+# ALL-NEXT:   - { id: 4, class: vr512, preferred-register: '' }
+# ALL-NEXT:   - { id: 5, class: vr512, preferred-register: '' }
+registers:
+  - { id: 0, class: vecr }
+  - { id: 1, class: vecr }
+# ALL:          %0 = IMPLICIT_DEF
+# ALL-NEXT:     undef %2.sub_xmm = COPY %0
+# ALL-NEXT:     %3 = VINSERTF32x4Zrr %2, %0, 1
+# ALL-NEXT:     %4 = VINSERTF32x4Zrr %3, %0, 2
+# ALL-NEXT:     %5 = VINSERTF32x4Zrr %4, %0, 3
+# ALL-NEXT:     %1 = COPY %5
+# ALL-NEXT:     %zmm0 = COPY %1
+# ALL-NEXT:     RET 0, implicit %zmm0
+body:             |
+  bb.1 (%ir-block.0):
+
+    %0(<4 x s32>) = IMPLICIT_DEF
+    %1(<16 x s32>) = G_MERGE_VALUES %0(<4 x s32>), %0(<4 x s32>), %0(<4 x s32>), %0(<4 x s32>)
+    %zmm0 = COPY %1(<16 x s32>)
+    RET 0, implicit %zmm0
+
+...
+---
+name:            test_merge_v256
+# ALL-LABEL: name:  test_merge_v256
+alignment:       4
+legalized:       true
+regBankSelected: true
+# ALL:      registers:
+# ALL-NEXT:   - { id: 0, class: vr256x, preferred-register: '' }
+# ALL-NEXT:   - { id: 1, class: vr512, preferred-register: '' }
+# ALL-NEXT:   - { id: 2, class: vr512, preferred-register: '' }
+# ALL-NEXT:   - { id: 3, class: vr512, preferred-register: '' }
+registers:
+  - { id: 0, class: vecr }
+  - { id: 1, class: vecr }
+# ALL:          %0 = IMPLICIT_DEF
+# ALL-NEXT:     undef %2.sub_ymm = COPY %0
+# ALL-NEXT:     %3 = VINSERTF64x4Zrr %2, %0, 1
+# ALL-NEXT:     %1 = COPY %3
+# ALL-NEXT:     %zmm0 = COPY %1
+# ALL-NEXT:     RET 0, implicit %zmm0
+body:             |
+  bb.1 (%ir-block.0):
+
+    %0(<8 x s32>) = IMPLICIT_DEF
+    %1(<16 x s32>) = G_MERGE_VALUES %0(<8 x s32>), %0(<8 x s32>)
+    %zmm0 = COPY %1(<16 x s32>)
+    RET 0, implicit %zmm0
+
+...
+