[AIE2p] Use multi-slot pseudo for const COPY with unique def

Author: krishnamtibrewala

llvm/lib/Target/AIE/aie2p/AIE2PInstrInfo.cpp

@@ -557,10 +557,19 @@ void AIE2PInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
 
   if (AIE2P::mMvSclSrcRegClass.contains(SrcReg) &&
       AIE2P::mMvSclDstRegClass.contains(DstReg)) {
-    // Build MultiSlotPseudo in preference
-    const unsigned MOVSclOpcode = getScalarMovOpcode(DstReg, SrcReg);
-    BuildMI(MBB, MBBI, DL, get(MOVSclOpcode), DstReg)
-        .addReg(SrcReg, getKillRegState(KillSrc));
+    if (MachineInstr *MI = MRI.getUniqueVRegDef(SrcReg);
+        MI && MI->isMoveImmediate()) {
+      // Try modifying scalar move to pseudo immediate move.
+      const int64_t Imm = MI->getOperand(1).getImm();
+      APInt ImmVal = APInt(64, Imm);
+      auto OpCode = getConstantMovOpcode(MRI, DstReg, ImmVal);
+      BuildMI(MBB, MBBI, DL, get(OpCode), DstReg).addImm(Imm);
+    } else {
+      // Build MultiSlotPseudo in preference
+      const unsigned MOVSclOpcode = getScalarMovOpcode(DstReg, SrcReg);
+      BuildMI(MBB, MBBI, DL, get(MOVSclOpcode), DstReg)
+          .addReg(SrcReg, getKillRegState(KillSrc));
+    }
   } else if ((AIE2P::eLRegClass.contains(SrcReg)) &&
              (AIE2P::eLRegClass.contains(DstReg))) {
     BuildMI(MBB, MBBI, DL, get(AIE2P::MOV_alu_mv_mv_mv_scl),
@@ -1138,11 +1147,15 @@ unsigned AIE2PInstrInfo::getConstantMovOpcode(MachineRegisterInfo &MRI,
   unsigned int ImmSize = Val.getSignificantBits();
 
   const TargetRegisterClass *DstRegClass = nullptr;
-  const RegClassOrRegBank &RCB = MRI.getRegClassOrRegBank(Reg);
-  if (const RegisterBank *RB = RCB.dyn_cast<const RegisterBank *>())
-    DstRegClass = &TRI->getMinClassForRegBank(*RB, MRI.getType(Reg));
-  if (auto *TRC = RCB.dyn_cast<const TargetRegisterClass *>())
-    DstRegClass = TRC;
+  if (Register::isVirtualRegister(Reg)) {
+    const RegClassOrRegBank &RCB = MRI.getRegClassOrRegBank(Reg);
+    if (const RegisterBank *RB = RCB.dyn_cast<const RegisterBank *>())
+      DstRegClass = &TRI->getMinClassForRegBank(*RB, MRI.getType(Reg));
+    if (auto *TRC = RCB.dyn_cast<const TargetRegisterClass *>())
+      DstRegClass = TRC;
+  } else {
+    DstRegClass = TRI->getMinimalPhysRegClass(Reg);
+  }
   assert(DstRegClass != nullptr && "RC cannot be null");
   if (ImmSize <= 11) {
     if (regClassMatches(AIE2P::mAluCgRegClass, DstRegClass, Reg))

llvm/test/CodeGen/AIE/aie2p/end-to-end/conv2d_bfp16_convert.ll

@@ -23,9 +23,9 @@ define weak_odr dso_local void @convert_bf16_to_bfp16(ptr noalias %in, ptr noali
 ; CHECK-NEXT:    nop
 ; CHECK-NEXT:    nop
 ; CHECK-NEXT:    movx r24, #0
-; CHECK-NEXT:    mova dj0, #0; mov r26, r24
-; CHECK-NEXT:    vldb.fill.512 [p0, lf0, r24]; mov dj1, dj0
-; CHECK-NEXT:    movs dc1, dj0; vldb.pop.512 x0, [p0, lf0, r24]; mov dn1, dn0
+; CHECK-NEXT:    mov r26, r24
+; CHECK-NEXT:    mova dj0, #0; vldb.fill.512 [p0, lf0, r24]
+; CHECK-NEXT:    mova dc1, #0; vldb.pop.512 x0, [p0, lf0, r24]; movs dj1, dj0; mov dn1, dn0
 ; CHECK-NEXT:    vldb.pop.512.2d x2, [p0, lf0, r24, d1]
 ; CHECK-NEXT:    nop
 ; CHECK-NEXT:    vldb.fill.512 [p0, lf0, r24]
@@ -35,7 +35,7 @@ define weak_odr dso_local void @convert_bf16_to_bfp16(ptr noalias %in, ptr noali
 ; CHECK-NEXT:    nopa ; vldb.fill.512 [p0, lf0, r24]; nops ; nopxm ; nopv
 ; CHECK-NEXT:    nopa ; vldb.pop.512 x0, [p0, lf0, r24]; nops ; nopx ; vconv.fp32.bf16 cml0, x0; nopv
 ; CHECK-NEXT:    nopa ; vldb.pop.512.2d x2, [p0, lf0, r24, d1]; nops ; nopx ; vconv.fp32.bf16 cmh0, x2; nopv
-; CHECK-NEXT:    nopa ; nopb ; movs dc0, dj0; nopx ; mov p2, p1; nopv
+; CHECK-NEXT:    mova dc0, #0; nopb ; nops ; nopx ; mov p2, p1; nopv
 ; CHECK-NEXT:    // implicit-def: $sf
 ; CHECK-NEXT:    .p2align 4
 ; CHECK-NEXT:  .LBB0_1: // %for.body

llvm/test/CodeGen/AIE/aie2p/end-to-end/gelu-templated.ll

@@ -52,10 +52,10 @@ define void @gelu_fn(ptr noalias %ifm, ptr noalias %ofm, ptr nonnull align 64 de
 ; CHECK-NEXT:    nop
 ; CHECK-NEXT:    vconv.bf16.fp32 x5, cml1
 ; CHECK-NEXT:    vconv.bf16.fp32 x8, cml4; movxm ls, #.LBB0_1; vmul.f dm4, x10, x4, r2
-; CHECK-NEXT:    mova r3, #0; nopb ; vconv.bf16.fp32 x7, cml2; movxm le, #.L_LEnd0; vmul.f dm4, x5, x4, r2
-; CHECK-NEXT:    vconv.bf16.fp32 x5, cml3; mov s0, r3; vadd.f dm2, dm1, dm2, r0
+; CHECK-NEXT:    vconv.bf16.fp32 x7, cml2; movxm le, #.L_LEnd0; vmul.f dm4, x5, x4, r2
+; CHECK-NEXT:    vconv.bf16.fp32 x5, cml3; mov s0, #0; vadd.f dm2, dm1, dm2, r0
 ; CHECK-NEXT:    vmov cml2, cml0; vmul.f dm3, x7, x2, r2
-; CHECK-NEXT:    vlda.conv.fp32.bf16 cml1, [p0], #64; vfloor.s32.bf16 x1, wl8, s0; movx r4, #-5; vmul.f dm4, x5, x4, r2
+; CHECK-NEXT:    vlda.conv.fp32.bf16 cml1, [p0], #64; nopb ; vfloor.s32.bf16 x1, wl8, s0; movx r4, #-5; mov r3, #0; vmul.f dm4, x5, x4, r2
 ; CHECK-NEXT:    vfloor.s32.bf16 x3, wh8, s0; lshl r4, r1, r4; vbcst.16 x6, r3
 ; CHECK-NEXT:    mova r1, #2; vconv.bf16.fp32 x10, cml4; add.nc lc, r4, #-7
 ; CHECK-NEXT:    nopa ; nopb ; nops ; nopx ; vshuffle x1, x1, x3, r1; nopv

llvm/test/CodeGen/AIE/aie2p/postrapseudos/pseudomove.mir

@@ -47,6 +47,18 @@ body:             |
     $r1 = COPY $r0
 ...
 
+---
+name:            COPY_GPR_const
+alignment:       16
+body:             |
+  bb.0 (align 16):
+    ; CHECK-LABEL: name: COPY_GPR_const
+    ; CHECK: $r0 = MOV_RLC_imm11_pseudo 10
+    ; CHECK-NEXT: $r1 = MOV_RLC_imm11_pseudo 10
+    $r0 = MOV_RLC_imm11_pseudo 10
+    $r1 = COPY $r0
+...
+
 ---
 name:            COPY_non_GPR
 alignment:       16