[NewPM] Run the Speculative Execution Pass only if the target has divergent branches

[leonardchan]

Summary

The SpeculativeExecutionPass was running under the new PM for non-GPU targets. It appears that the pass would run regardless as long as speed optimizations were available. Run the Speculative Execution Pass only if the target has divergent branches.

[arsenm]

Commit message should say only if?

[leonardchan]

In D76389#1951135, @arsenm wrote:

Commit message should say only if?

Updated

[arsenm]

clang/test/CodeGen/thinlto-distributed-newpm.ll

110 ↗

(On Diff #251182)

The pass name shouldn't change here?

llvm/lib/Passes/PassBuilder.cpp

443

I don't expect this to be a separate pass, just added based on a divergent target

[leonardchan]

llvm/lib/Passes/PassBuilder.cpp

443

Updated such that the current pass is used, although it's a bit more difficult to test since the same pass name will appear when dumping the passes.

[arsenm]

llvm/include/llvm/Transforms/Scalar/SpeculativeExecution.h

97 ↗

(On Diff #255474)

Still define a second pass

llvm/lib/Passes/PassBuilder.cpp

443

There's a check for divergent branches in the builder so you don't need to run the pass at all

[arsenm]

This seems like it covers a different case than D82735?

[leonardchan]

In D76389#2120868, @arsenm wrote:

This seems like it covers a different case than D82735?

I haven't rebased in a while, but D82735 covers this. Abandoning since this is addressed in that patch.

llvm/lib/Passes/PassBuilder.cpp

if (Level.getSpeedupLevel() > 1) {
     if (EnableGVNSink) {
       FPM.addPass(GVNSinkPass());
       FPM.addPass(SimplifyCFGPass());
     }
   }
 
   // Speculative execution if the target has divergent branches; otherwise nop.
   if (Level.getSpeedupLevel() > 1) {
-    FPM.addPass(SpeculativeExecutionPass());
+    FPM.addPass(SpeculativeExecutionPass(/*OnlyIfDivergentTarget=*/true));
 
     // Optimize based on known information about branches, and cleanup afterward.
     FPM.addPass(JumpThreadingPass());
     FPM.addPass(CorrelatedValuePropagationPass());
   }
   FPM.addPass(SimplifyCFGPass());
   if (Level == OptimizationLevel::O3)
     FPM.addPass(AggressiveInstCombinePass());