[llvm] [AMDGPU][SIPreEmitPeephole] mustRetainExeczBranch: use BranchProbability and TargetSchedmodel (PR #109818)

[Matt Arsenault via llvm-commits]

Juan Manuel Martinez =?utf-8?q?Caamaño?= <juamarti at amd.com>,
Juan Manuel Martinez =?utf-8?q?Caamaño?= <juamarti at amd.com>
Message-ID:
In-Reply-To: <llvm.org/llvm/llvm-project/pull/109818 at github.com>


================
@@ -326,15 +345,32 @@ bool SIPreEmitPeephole::mustRetainExeczBranch(
       if (TII->hasUnwantedEffectsWhenEXECEmpty(MI))
         return true;
 
-      // These instructions are potentially expensive even if EXEC = 0.
-      if (TII->isSMRD(MI) || TII->isVMEM(MI) || TII->isFLAT(MI) ||
-          TII->isDS(MI) || TII->isWaitcnt(MI.getOpcode()))
-        return true;
-
-      ++NumInstr;
-      if (NumInstr >= SkipThreshold)
+      if (TII->isWaitcnt(MI.getOpcode()))
         return true;
     }
+
+    if (!MinInstr)
+      MinInstr = Traces->getEnsemble(MachineTraceStrategy::TS_Local);
+
+    auto Trace = MinInstr->getTrace(&From);
+    ThenCyclesCost +=
+        std::max(Trace.getCriticalPath(), Trace.getResourceDepth(true));
+
+    // Consider `P = N/D` to be the probability of execz being false (skipping
+    // the then-block) The transformation is profitable if always executing the
+    // 'then' block is cheaper than executing sometimes 'then' and always
+    // executing s_cbranch_execz:
+    // * ThenCost <= P*ThenCost + (1-P)*BranchTakenCost + P*BranchNonTakenCost
+    // * (1-P) * ThenCost <= (1-P)*BranchTakenCost + P*BranchNonTakenCost
+    // * (D-N)/D * ThenCost <= (D-N)/D * BranchTakenCost + N/D *
+    // BranchNonTakenCost
+    uint64_t Numerator = BranchProb.getNumerator();
+    uint64_t Denominator = BranchProb.getDenominator();
+    bool IsProfitable = (Denominator - Numerator) * ThenCyclesCost <=
+                        ((Denominator - Numerator) * BranchTakenCost +
+                         Numerator * BranchNotTakenCost);
+    if (!IsProfitable)
+      return true;
----------------
arsenm wrote:

return !IsProfitable

https://github.com/llvm/llvm-project/pull/109818