[llvm] [RISCV] Minimally modify incoming state in transferBefore (PR #72352)

[Philip Reames via llvm-commits]

================
@@ -1013,73 +1015,76 @@ bool RISCVInsertVSETVLI::needVSETVLI(const MachineInstr &MI,
   return true;
 }
 
-// Given an incoming state reaching MI, modifies that state so that it is minimally
-// compatible with MI.  The resulting state is guaranteed to be semantically legal
-// for MI, but may not be the state requested by MI.
+// Given an incoming state reaching MI, minimally modifies that state so that it
+// is compatible with MI. The resulting state is guaranteed to be semantically
+// legal for MI, but may not be the state requested by MI.
 void RISCVInsertVSETVLI::transferBefore(VSETVLIInfo &Info,
                                         const MachineInstr &MI) const {
   uint64_t TSFlags = MI.getDesc().TSFlags;
   if (!RISCVII::hasSEWOp(TSFlags))
     return;
 
-  const VSETVLIInfo NewInfo = computeInfoForInstr(MI, TSFlags, MRI);
+  VSETVLIInfo NewInfo = computeInfoForInstr(MI, TSFlags, MRI);
   if (Info.isValid() && !needVSETVLI(MI, NewInfo, Info))
     return;
 
-  const VSETVLIInfo PrevInfo = Info;
-  Info = NewInfo;
+  if (Info.hasSEWLMULRatioOnly() || !Info.isValid() || Info.isUnknown())
+    Info = NewInfo;
 
-  if (!RISCVII::hasVLOp(TSFlags))
-    return;
+  DemandedFields Demanded = getDemanded(MI, MRI, ST);
 
   // If we don't use LMUL or the SEW/LMUL ratio, then adjust LMUL so that we
   // maintain the SEW/LMUL ratio. This allows us to eliminate VL toggles in more
   // places.
-  DemandedFields Demanded = getDemanded(MI, MRI, ST);
   if (!Demanded.LMUL && !Demanded.SEWLMULRatio && Info.isValid() &&
-      PrevInfo.isValid() && !Info.isUnknown() && !PrevInfo.isUnknown()) {
-    if (auto NewVLMul = RISCVVType::getSameRatioLMUL(
-            PrevInfo.getSEW(), PrevInfo.getVLMUL(), Info.getSEW()))
-      Info.setVLMul(*NewVLMul);
-  }
-
-  // For vmv.s.x and vfmv.s.f, there are only two behaviors, VL = 0 and
-  // VL > 0. We can discard the user requested AVL and just use the last
-  // one if we can prove it equally zero.  This removes a vsetvli entirely
-  // if the types match or allows use of cheaper avl preserving variant
-  // if VLMAX doesn't change.  If VLMAX might change, we couldn't use
-  // the 'vsetvli x0, x0, vtype" variant, so we avoid the transform to
-  // prevent extending live range of an avl register operand.
-  // TODO: We can probably relax this for immediates.
-  if (isScalarInsertInstr(MI) && PrevInfo.isValid() &&
----------------
preames wrote:

A good first step (before this change), might be to rework this code in terms of the demanded fields without otherwise changing structure.  (i.e. this can be driven off only demanding zeroness of VL, not isScalarInsertInst directly.)

https://github.com/llvm/llvm-project/pull/72352