[llvm] [X86] Set up the framework for optimization of CCMP/CTEST (PR #84603)

[Shengchen Kan via llvm-commits]

================
@@ -3377,7 +3377,9 @@ X86TargetLowering::getJumpConditionMergingParams(Instruction::BinaryOps Opc,
                                                  const Value *Lhs,
                                                  const Value *Rhs) const {
   using namespace llvm::PatternMatch;
-  int BaseCost = BrMergingBaseCostThresh.getValue();
+  // Disable condition merging when CCMP is available b/c we can eliminate
+  // branches in a more efficient way.
+  int BaseCost = Subtarget.hasCCMP() ? -1 : BrMergingBaseCostThresh.getValue();
----------------
KanRobert wrote:

@goldsteinn We can't lower `(and/or setcc, setcc)` as `ccmp` b/c CCMP has a CMP semantic. Instead, what we need to need to do is sth like

1.  Combine the DAG
```
  // sub(and(setcc(cc1, ...), setcc(cc2, sub (X, Y))), 1)
  // brcond ne
  //
  //  ->
  //
  //  ccmp(X, Y, cc1, cc2)
  //  brcond cc2
  //
  //  if only flag has users.
```
2. Define fragment for `ccmp`
3. add pattern for CCMP using the fragment

I drafted a patch locally and found that it's more complex than the MIR pass.

IIUC, some checks in `canConvert` are conservative and not real restrictions, it's borrowed from AArch64ConditionalCompares.cpp, which was designed to run
immediately before the early if-conversion pass.




https://github.com/llvm/llvm-project/pull/84603