[llvm] 14ea545 - [RISCV][InsertVSETVLI] Generalize scalar move rule for when AVL is unchanged

[Philip Reames via llvm-commits]

Author: Philip Reames
Date: 2022-12-07T10:28:31-08:00
New Revision: 14ea545a7d16dad1d26c8ba7f817a1b3d33d4132

URL: https://github.com/llvm/llvm-project/commit/14ea545a7d16dad1d26c8ba7f817a1b3d33d4132
DIFF: https://github.com/llvm/llvm-project/commit/14ea545a7d16dad1d26c8ba7f817a1b3d33d4132.diff

LOG: [RISCV][InsertVSETVLI] Generalize scalar move rule for when AVL is unchanged

By definition, the AVL of the scalar move is equally zero to the prior AVL if they are the same value.  This generalizes the existing code to the case where the scalar move has a register AVL which is unknown, but unchanged from the preceeding instruction.

This doesn't cause any interesting diffs on its own, but another patch makes this case much more common.  Split off to reduce a future diff.

Added: 
    

Modified: 
    llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp b/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
index ac26f1a43a70..aae570c872da 100644
--- a/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
@@ -304,11 +304,13 @@ class VSETVLIInfo {
     return false;
   }
 
+  bool hasEquallyZeroAVL(const VSETVLIInfo &Other) const {
+    if (hasSameAVL(Other))
+      return true;
+    return (hasNonZeroAVL() && Other.hasNonZeroAVL());
+  }
+
   bool hasSameAVL(const VSETVLIInfo &Other) const {
-    assert(isValid() && Other.isValid() &&
-           "Can't compare invalid VSETVLIInfos");
-    assert(!isUnknown() && !Other.isUnknown() &&
-           "Can't compare AVL in unknown state");
     if (hasAVLReg() && Other.hasAVLReg())
       return getAVLReg() == Other.getAVLReg();
 
@@ -771,12 +773,10 @@ bool RISCVInsertVSETVLI::needVSETVLI(const MachineInstr &MI,
     return true;
 
   // For vmv.s.x and vfmv.s.f, there is only two behaviors, VL = 0 and VL > 0.
-  // VL=0 is uninteresting (as it should have been deleted already), so it is
-  // compatible if we can prove both are non-zero.  Additionally, if writing
-  // to an implicit_def operand, we don't need to preserve any other bits and
-  // are thus compatible with any larger etype, and can disregard policy bits.
-  if (isScalarMoveInstr(MI) &&
-      CurInfo.hasNonZeroAVL() && Require.hasNonZeroAVL()) {
+  // Additionally, if writing to an implicit_def operand, we don't need to
+  // preserve any other bits and are thus compatible with any larger etype,
+  // and can disregard policy bits.
+  if (isScalarMoveInstr(MI) && CurInfo.hasEquallyZeroAVL(Require)) {
     auto *VRegDef = MRI->getVRegDef(MI.getOperand(1).getReg());
     if (VRegDef && VRegDef->isImplicitDef() &&
         CurInfo.getSEW() >= Require.getSEW())
@@ -830,7 +830,7 @@ void RISCVInsertVSETVLI::transferBefore(VSETVLIInfo &Info, const MachineInstr &M
   // prevent extending live range of an avl register operand.
   // TODO: We can probably relax this for immediates.
   if (isScalarMoveInstr(MI) && PrevInfo.isValid() &&
-      PrevInfo.hasNonZeroAVL() && Info.hasNonZeroAVL() &&
+      PrevInfo.hasEquallyZeroAVL(Info) &&
       Info.hasSameVLMAX(PrevInfo)) {
     if (PrevInfo.hasAVLImm())
       Info.setAVLImm(PrevInfo.getAVLImm());