[early-ifcvt] Add OptRemarks

[thegameg]

This looks great, thanks! Since it's a CodeGen pass, I would suggest using a MIR test instead of an IR one, but I think it's probably easier to read it as it is right now.

[jroelofs]

In D86797#2244962, @thegameg wrote:

This looks great, thanks! Since it's a CodeGen pass, I would suggest using a MIR test instead of an IR one, but I think it's probably easier to read it as it is right now.

Good call. As-is, this is probably more sensitive to changes in the passes that happen before early-ifcvt than it should be. Making it a MIR test should tighten that up a bit.

[jroelofs]

I'm having trouble getting -stop-before MIR output to trigger these remarks when fed back into llc, so instead I've just relaxed the checks against specific cycle counts to be more forgiving.

[jroelofs]

b15f2bd3ad4ab614e75ff859b03c3c1eb9f1a9d3

llvm/lib/CodeGen/EarlyIfConversion.cpp

 #include "llvm/ADT/SparseSet.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
 #include "llvm/CodeGen/MachineDominators.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/CodeGen/MachineInstr.h"
 #include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/MachineTraceMetrics.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/CodeGen/TargetInstrInfo.h"
 #include "llvm/CodeGen/TargetRegisterInfo.h"
 #include "llvm/CodeGen/TargetSubtargetInfo.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Support/CommandLine.h"

 
 // Adjust cycles with downward saturation.
 static unsigned adjCycles(unsigned Cyc, int Delta) {
   if (Delta < 0 && Cyc + Delta > Cyc)
     return 0;
   return Cyc + Delta;
 }
 
+namespace {
+/// Helper class to simplify emission of cycle counts into optimization remarks.
+struct Cycles {
+  const char *Key;
+  unsigned Value;
+};
+template <typename Remark> Remark &operator<<(Remark &R, Cycles C) {
+  return R << ore::NV(C.Key, C.Value) << (C.Value == 1 ? " cycle" : " cycles");
+}
+} // anonymous namespace
+
 /// Apply cost model and heuristics to the if-conversion in IfConv.
 /// Return true if the conversion is a good idea.
 ///
 bool EarlyIfConverter::shouldConvertIf() {
   // Stress testing mode disables all cost considerations.
   if (Stress)
     return true;
 
   if (!MinInstr)
     MinInstr = Traces->getEnsemble(MachineTraceMetrics::TS_MinInstrCount);
 
   MachineTraceMetrics::Trace TBBTrace = MinInstr->getTrace(IfConv.getTPred());
   MachineTraceMetrics::Trace FBBTrace = MinInstr->getTrace(IfConv.getFPred());
   LLVM_DEBUG(dbgs() << "TBB: " << TBBTrace << "FBB: " << FBBTrace);
   unsigned MinCrit = std::min(TBBTrace.getCriticalPath(),
                               FBBTrace.getCriticalPath());
 
   // Set a somewhat arbitrary limit on the critical path extension we accept.
   unsigned CritLimit = SchedModel.MispredictPenalty/2;
 
+  MachineBasicBlock &MBB = *IfConv.Head;
+  MachineOptimizationRemarkEmitter MORE(*MBB.getParent(), nullptr);
+
   // If-conversion only makes sense when there is unexploited ILP. Compute the
   // maximum-ILP resource length of the trace after if-conversion. Compare it
   // to the shortest critical path.
   SmallVector<const MachineBasicBlock*, 1> ExtraBlocks;
   if (IfConv.TBB != IfConv.Tail)
     ExtraBlocks.push_back(IfConv.TBB);
   unsigned ResLength = FBBTrace.getResourceLength(ExtraBlocks);
   LLVM_DEBUG(dbgs() << "Resource length " << ResLength
                     << ", minimal critical path " << MinCrit << '\n');
   if (ResLength > MinCrit + CritLimit) {
     LLVM_DEBUG(dbgs() << "Not enough available ILP.\n");
+    MORE.emit([&]() {
+      MachineOptimizationRemarkMissed R(DEBUG_TYPE, "IfConversion",
+                                        MBB.findDebugLoc(MBB.back()), &MBB);
+      R << "did not if-convert branch: the resulting critical path ("
+        << Cycles{"ResLength", ResLength}
+        << ") would extend the shorter leg's critical path ("
+        << Cycles{"MinCrit", MinCrit} << ") by more than the threshold of "
+        << Cycles{"CritLimit", CritLimit}
+        << ", which cannot be hidden by available ILP.";
+      return R;
+    });
     return false;
   }
 
   // Assume that the depth of the first head terminator will also be the depth
   // of the select instruction inserted, as determined by the flag dependency.
   // TBB / FBB data dependencies may delay the select even more.
   MachineTraceMetrics::Trace HeadTrace = MinInstr->getTrace(IfConv.Head);
   unsigned BranchDepth =
       HeadTrace.getInstrCycles(*IfConv.Head->getFirstTerminator()).Depth;
   LLVM_DEBUG(dbgs() << "Branch depth: " << BranchDepth << '\n');
 
   // Look at all the tail phis, and compute the critical path extension caused
   // by inserting select instructions.
   MachineTraceMetrics::Trace TailTrace = MinInstr->getTrace(IfConv.Tail);
+  struct CriticalPathInfo {
+    unsigned Extra; //< Count of extra cycles that the component adds.
+    unsigned Depth; //< Absolute depth of the component in cycles.
+  };
+  CriticalPathInfo Cond{};
+  CriticalPathInfo TBlock{};
+  CriticalPathInfo FBlock{};
+  bool ShouldConvert = true;
   for (unsigned i = 0, e = IfConv.PHIs.size(); i != e; ++i) {
     SSAIfConv::PHIInfo &PI = IfConv.PHIs[i];
     unsigned Slack = TailTrace.getInstrSlack(*PI.PHI);
     unsigned MaxDepth = Slack + TailTrace.getInstrCycles(*PI.PHI).Depth;
     LLVM_DEBUG(dbgs() << "Slack " << Slack << ":\t" << *PI.PHI);
 
     // The condition is pulled into the critical path.
     unsigned CondDepth = adjCycles(BranchDepth, PI.CondCycles);
     if (CondDepth > MaxDepth) {
       unsigned Extra = CondDepth - MaxDepth;
       LLVM_DEBUG(dbgs() << "Condition adds " << Extra << " cycles.\n");
+      if (Extra > Cond.Extra)
+        Cond = {Extra, CondDepth};
       if (Extra > CritLimit) {
         LLVM_DEBUG(dbgs() << "Exceeds limit of " << CritLimit << '\n');
-        return false;
+        ShouldConvert = false;
       }
     }
 
     // The TBB value is pulled into the critical path.
     unsigned TDepth = adjCycles(TBBTrace.getPHIDepth(*PI.PHI), PI.TCycles);
     if (TDepth > MaxDepth) {
       unsigned Extra = TDepth - MaxDepth;
       LLVM_DEBUG(dbgs() << "TBB data adds " << Extra << " cycles.\n");
+      if (Extra > TBlock.Extra)
+        TBlock = {Extra, TDepth};
       if (Extra > CritLimit) {
         LLVM_DEBUG(dbgs() << "Exceeds limit of " << CritLimit << '\n');
-        return false;
+        ShouldConvert = false;
       }
     }
 
     // The FBB value is pulled into the critical path.
     unsigned FDepth = adjCycles(FBBTrace.getPHIDepth(*PI.PHI), PI.FCycles);
     if (FDepth > MaxDepth) {
       unsigned Extra = FDepth - MaxDepth;
       LLVM_DEBUG(dbgs() << "FBB data adds " << Extra << " cycles.\n");
+      if (Extra > FBlock.Extra)
+        FBlock = {Extra, FDepth};
       if (Extra > CritLimit) {
         LLVM_DEBUG(dbgs() << "Exceeds limit of " << CritLimit << '\n');
-        return false;
+        ShouldConvert = false;
       }
     }
   }
-  return true;
+
+  // Organize by "short" and "long" legs, since the diagnostics get confusing
+  // when referring to the "true" and "false" sides of the branch, given that
+  // those don't always correlate with what the user wrote in source-terms.
+  const CriticalPathInfo Short = TBlock.Extra > FBlock.Extra ? FBlock : TBlock;
+  const CriticalPathInfo Long = TBlock.Extra > FBlock.Extra ? TBlock : FBlock;
+
+  if (ShouldConvert) {
+    MORE.emit([&]() {
+      MachineOptimizationRemark R(DEBUG_TYPE, "IfConversion",
+                                  MBB.back().getDebugLoc(), &MBB);
+      R << "performing if-conversion on branch: the condition adds "
+        << Cycles{"CondCycles", Cond.Extra} << " to the critical path";
+      if (Short.Extra > 0)
+        R << ", and the short leg adds another "
+          << Cycles{"ShortCycles", Short.Extra};
+      if (Long.Extra > 0)
+        R << ", and the long leg adds another "
+          << Cycles{"LongCycles", Long.Extra};
+      R << ", each staying under the threshold of "
+        << Cycles{"CritLimit", CritLimit} << ".";
+      return R;
+    });
+  } else {
+    MORE.emit([&]() {
+      MachineOptimizationRemarkMissed R(DEBUG_TYPE, "IfConversion",
+                                        MBB.back().getDebugLoc(), &MBB);
+      R << "did not if-convert branch: the condition would add "
+        << Cycles{"CondCycles", Cond.Extra} << " to the critical path";
+      if (Cond.Extra > CritLimit)
+        R << " exceeding the limit of " << Cycles{"CritLimit", CritLimit};
+      if (Short.Extra > 0) {
+        R << ", and the short leg would add another "
+          << Cycles{"ShortCycles", Short.Extra};
+        if (Short.Extra > CritLimit)
+          R << " exceeding the limit of " << Cycles{"CritLimit", CritLimit};
+      }
+      if (Long.Extra > 0) {
+        R << ", and the long leg would add another "
+          << Cycles{"LongCycles", Long.Extra};
+        if (Long.Extra > CritLimit)
+          R << " exceeding the limit of " << Cycles{"CritLimit", CritLimit};
+      }
+      R << ".";
+      return R;
+    });
+  }
+
+  return ShouldConvert;
 }
 
 /// Attempt repeated if-conversion on MBB, return true if successful.
 ///
 bool EarlyIfConverter::tryConvertIf(MachineBasicBlock *MBB) {
   bool Changed = false;
   while (IfConv.canConvertIf(MBB) && shouldConvertIf()) {
     // If-convert MBB and update analyses.

llvm/test/CodeGen/X86/early-ifcvt-remarks.ll

+; RUN: llc %s -x86-early-ifcvt -pass-remarks='early-ifcvt' -pass-remarks-missed='early-ifcvt' -mcpu=k8 -o - 2>&1 | FileCheck %s
+target triple = "x86_64-none-none"
+
+; CHECK: remark: <unknown>:0:0: performing if-conversion on branch:
+; CHECK-SAME: the condition adds {{[0-9]+}} cycle{{s?}} to the critical path,
+; CHECK-SAME: and the short leg adds another {{[0-9]+}} cycles{{s?}},
+; CHECK-SAME: and the long leg adds another {{[0-9]+}} cycles{{s?}},
+; CHECK-SAME: each staying under the threshold of {{[0-9]+}} cycles{{s?}}.
+define i32 @mm1(i1 %pred, i32 %val) {
+entry:
+  br i1 %pred, label %if.true, label %if.else
+
+if.true:
+  %v1 = add i32 1, %val
+  br label %if.else
+
+if.else:
+  %res = phi i32 [ %val, %entry ], [ %v1, %if.true ]
+  ret i32 %res
+}
+
+; CHECK: remark: <unknown>:0:0: did not if-convert branch:
+; CHECK-SAME: the condition would add {{[0-9]+}} cycles{{s?}} to the critical path,
+; CHECK-SAME: and the short leg would add another {{[0-9]+}} cycles{{s?}},
+; CHECK-SAME: and the long leg would add another {{[0-9]+}} cycles{{s?}} exceeding the limit of {{[0-9]+}} cycles{{s?}}.
+define i32 @mm2(i1 %pred, i32 %val, i32 %e1, i32 %e2, i32 %e3, i32 %e4, i32 %e5) {
+entry:
+  br i1 %pred, label %if.true, label %if.else
+
+if.true:
+  %v1 = add i32 %e1, %val
+  %v2 = add i32 %e2, %v1
+  %v3 = add i32 %e3, %v2
+  %v4 = add i32 %e4, %v3
+  br label %if.else
+
+if.else:
+  %res = phi i32 [ %val, %entry ], [ %v4, %if.true ]
+  ret i32 %res
+}
+
+; CHECK: did not if-convert branch:
+; CHECK-SAME: the resulting critical path ({{[0-9]+}} cycles{{s?}})
+; CHECK-SAME: would extend the shorter leg's critical path ({{[0-9]+}} cycle{{s?}})
+; CHECK-SAME: by more than the threshold of {{[0-9]+}} cycles{{s?}},
+; CHECK-SAME: which cannot be hidden by available ILP.
+define i32 @mm3(i1 %pred, i32 %val, i32 %e1, i128 %e2, i128 %e3, i128 %e4, i128 %e5) {
+entry:
+  br i1 %pred, label %if.true, label %if.false
+
+if.true:
+  br label %if.endif
+
+if.false:
+  %f1 = mul i32 %e1, %e1
+  %f3 = sext i32 %f1 to i128
+  %f4 = mul i128 %e2, %f3
+  %f6 = add i128 %e3, %f4
+  %f7 = xor i128 %e4, %f6
+  %f8 = xor i128 %e5, %f7
+  %a1 = ashr i128 %f8, %e5
+  %f5 = trunc i128 %a1 to i32
+  br label %if.endif
+
+if.endif:
+  %r1 = phi i32 [ %val, %if.true ], [ %f1, %if.false ]
+  %r2 = phi i32 [ %val, %if.true ], [ %f5, %if.false ]
+  %res = add i32 %r1, %r2
+  ret i32 %res
+}