[llvm] Reland "RegisterCoalescer: Add implicit-def of super register when coalescing SUBREG_TO_REG" (PR #168353)

[Sander de Smalen via llvm-commits]

================
@@ -1935,6 +1954,79 @@ void RegisterCoalescer::updateRegDefsUses(Register SrcReg, Register DstReg,
     if (DstInt && !Reads && SubIdx && !UseMI->isDebugInstr())
       Reads = DstInt->liveAt(LIS->getInstructionIndex(*UseMI));
 
+    bool RequiresImplicitRedef = false;
+    if (!SubregToRegSrcInsts.empty()) {
+      // We can only add an implicit-def and undef if the sub registers match,
+      // e.g.
+      //  %0:gr32      = INSTX
+      //  %0.sub8:gr32 = INSTY           // top 24 bits of %0 still defined
+      //  %1:gr64      = SUBREG_TO_REG 0, %0, %subreg.sub32
+      //
+      // This cannot be transformed into:
+      //  %1.sub32:gr64      = INSTX
+      //  undef %1.sub8:gr64 = INSTY , implicit-def %1
+      //
+      // Because that would thrash the top 24 bits of %1.sub32.
+      if (is_contained(SubregToRegSrcInsts, UseMI) &&
+          all_of(UseMI->defs(), [&SubIdx](const MachineOperand &MO) -> bool {
+            if (MO.isUndef())
+              return true;
+            return SubIdx && (!MO.getSubReg() || SubIdx == MO.getSubReg());
+          })) {
+        // Add implicit-def of super-register to express that the whole
+        // register is defined by the instruction.
+        MachineInstrBuilder MIB(*MF, UseMI);
+        MIB.addReg(DstReg, RegState::ImplicitDefine);
+        RequiresImplicitRedef = true;
+      }
+
+      // If the coalesed instruction doesn't fully define the register, we need
+      // to preserve the original super register liveness for SUBREG_TO_REG.
+      //
+      // We pretended SUBREG_TO_REG was a regular copy for coalescing purposes,
+      // but it introduces liveness for other subregisters. Downstream users may
+      // have been relying on those bits, so we need to ensure their liveness is
+      // captured with a def of other lanes.
+      if (DstInt && MRI->shouldTrackSubRegLiveness(DstReg)) {
+        // First check if there is sufficient granularity in terms of subranges.
+        LaneBitmask DstMask = MRI->getMaxLaneMaskForVReg(DstInt->reg());
+        LaneBitmask UsedLanes = TRI->getSubRegIndexLaneMask(SubIdx);
+        LaneBitmask UnusedLanes = DstMask & ~UsedLanes;
+        if ((UnusedLanes & ~DstIntLaneMask).any()) {
+          BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
+          DstInt->createSubRangeFrom(Allocator, UnusedLanes, *DstInt);
+          DstIntLaneMask |= UnusedLanes;
+        }
+
+        // After duplicating the live ranges for the low/hi bits, we
+        // need to update the subranges of the DstReg interval such that
+        // for a case like this:
+        //
+        //       entry:
+        //  16B    %1:gpr32 = INSTRUCTION    (<=> UseMI)
+        //            :
+        //       if.then:
+        //  32B    %1:gpr32 = MOVIMM32 ..
+        //  48B    %0:gpr64 = SUBREG_TO_REG 0, %1, sub32
+        //
+        //  Only the MOVIMM32 require a def of the top lanes and any intervals
+        //  for the top 32-bits of the def at 16B should be removed.
----------------
sdesmalen-arm wrote:

done.

https://github.com/llvm/llvm-project/pull/168353