AMDGPU: Move insertion into V2SCopies map (#130776)

Insert the start instruction directly into the map before the uses. This
prevents improperly re-visting sgpr->vgpr phi inputs multiple times
which
would trigger a use after free.

I don't particularly trust the iteration scheme here. This is also
unnecessarily revisting transitive users of a phi or reg_sequence for
every
input operand, but I will address that separately.

Fixes #130646. I also believe it fixes #130119, although that test fails
less consistently for me.

Author: arsenm

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

@@ -691,6 +691,9 @@ bool SIFixSGPRCopies::run(MachineFunction &MF) {
                             TII->get(AMDGPU::COPY), NewDst)
                         .addReg(MO.getReg());
                 MO.setReg(NewDst);
+
+                // FIXME: We are transitively revisiting users of this
+                // instruction for every input.
                 analyzeVGPRToSGPRCopy(NewCopy);
               }
             }
@@ -928,6 +931,8 @@ void SIFixSGPRCopies::analyzeVGPRToSGPRCopy(MachineInstr* MI) {
 
   V2SCopyInfo Info(getNextVGPRToSGPRCopyId(), MI,
                       TRI->getRegSizeInBits(*DstRC));
+  V2SCopies[Info.ID] = Info;
+
   SmallVector<MachineInstr *, 8> AnalysisWorklist;
   // Needed because the SSA is not a tree but a graph and may have
   // forks and joins. We should not then go same way twice.
@@ -976,7 +981,6 @@ void SIFixSGPRCopies::analyzeVGPRToSGPRCopy(MachineInstr* MI) {
       AnalysisWorklist.push_back(U);
     }
   }
-  V2SCopies[Info.ID] = Info;
 }
 
 // The main function that computes the VGPR to SGPR copy score