[PowerPC] Fix the PPCInstrInfo::getInstrLatency implementation

PowerPC uses itineraries to describe processor pipelines (and dispatch-group
restrictions for P7/P8 cores). Unfortunately, the target-independent
implementation of TII.getInstrLatency calls ItinData->getStageLatency, and that
looks for the largest cycle count in the pipeline for any given instruction.
This, however, yields the wrong answer for the PPC itineraries, because we
don't encode the full pipeline. Because the functional units are fully
pipelined, we only model the initial stages (there are no relevant hazards in
the later stages to model), and so the technique employed by getStageLatency
does not really work. Instead, we should take the maximum output operand
latency, and that's what PPCInstrInfo::getInstrLatency now does.

This caused some test-case churn, including two unfortunate side effects.
First, the new arrangement of copies we get from function parameters now
sometimes blocks VSX FMA mutation (a FIXME has been added to the code and the
test cases), and we have one significant test-suite regression:

SingleSource/Benchmarks/BenchmarkGame/spectral-norm
	56.4185% +/- 18.9398%

In this benchmark we have a loop with a vectorized FP divide, and it with the
new scheduling both divides end up in the same dispatch group (which in this
case seems to cause a problem, although why is not exactly clear). The grouping
structure is hard to predict from the bottom of the loop, and there may not be
much we can do to fix this.

Very few other test-suite performance effects were really significant, but
almost all weakly favor this change. However, in light of the issues
highlighted above, I've left the old behavior available via a
command-line flag.

llvm-svn: 242188

5 files changed, 55 insertions, 0 deletions

diff --git a/llvm/lib/Target/PowerPC/PPCInstrInfo.cpp b/llvm/lib/Target/PowerPC/PPCInstrInfo.cpp
index 696a83860e5..bf6e4029640 100644
--- a/llvm/lib/Target/PowerPC/PPCInstrInfo.cpp
+++ b/llvm/lib/Target/PowerPC/PPCInstrInfo.cpp
@@ -57,6 +57,10 @@ static cl::opt<bool> VSXSelfCopyCrash("crash-on-ppc-vsx-self-copy",
 cl::desc("Causes the backend to crash instead of generating a nop VSX copy"),
 cl::Hidden);
 
+static cl::opt<bool>
+UseOldLatencyCalc("ppc-old-latency-calc", cl::Hidden,
+  cl::desc("Use the old (incorrect) instruction latency calculation"));
+
 // Pin the vtable to this file.
 void PPCInstrInfo::anchor() {}
 
@@ -103,6 +107,35 @@ PPCInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
   return new ScoreboardHazardRecognizer(II, DAG);
 }
 
+unsigned PPCInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
+                                       const MachineInstr *MI,
+                                       unsigned *PredCost) const {
+  if (!ItinData || UseOldLatencyCalc)
+    return PPCGenInstrInfo::getInstrLatency(ItinData, MI, PredCost);
+
+  // The default implementation of getInstrLatency calls getStageLatency, but
+  // getStageLatency does not do the right thing for us. While we have
+  // itinerary, most cores are fully pipelined, and so the itineraries only
+  // express the first part of the pipeline, not every stage. Instead, we need
+  // to use the listed output operand cycle number (using operand 0 here, which
+  // is an output).
+
+  unsigned Latency = 1;
+  unsigned DefClass = MI->getDesc().getSchedClass();
+  for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+    const MachineOperand &MO = MI->getOperand(i);
+    if (!MO.isReg() || !MO.isDef() || MO.isImplicit())
+      continue;
+
+    int Cycle = ItinData->getOperandCycle(DefClass, i);
+    if (Cycle < 0)
+      continue;
+
+    Latency = std::max(Latency, (unsigned) Cycle);
+  }
+
+  return Latency;
+}
 
 int PPCInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
                                     const MachineInstr *DefMI, unsigned DefIdx,
diff --git a/llvm/lib/Target/PowerPC/PPCInstrInfo.h b/llvm/lib/Target/PowerPC/PPCInstrInfo.h
index e2d6346aa53..40badae644d 100644
--- a/llvm/lib/Target/PowerPC/PPCInstrInfo.h
+++ b/llvm/lib/Target/PowerPC/PPCInstrInfo.h
@@ -95,6 +95,10 @@ public:
   CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
                                      const ScheduleDAG *DAG) const override;
 
+  unsigned getInstrLatency(const InstrItineraryData *ItinData,
+                           const MachineInstr *MI,
+                           unsigned *PredCost = nullptr) const override;
+
   int getOperandLatency(const InstrItineraryData *ItinData,
                         const MachineInstr *DefMI, unsigned DefIdx,
                         const MachineInstr *UseMI,
diff --git a/llvm/lib/Target/PowerPC/PPCScheduleP7.td b/llvm/lib/Target/PowerPC/PPCScheduleP7.td
index 635d154d10b..267f5672618 100644
--- a/llvm/lib/Target/PowerPC/PPCScheduleP7.td
+++ b/llvm/lib/Target/PowerPC/PPCScheduleP7.td
@@ -315,6 +315,10 @@ def P7Itineraries : ProcessorItineraries<
                                                   P7_DU3, P7_DU4], 0>,
                                    InstrStage<1, [P7_VS1, P7_VS2]>],
                                   [5, 1, 1]>,
+  InstrItinData<IIC_FPAddSub    , [InstrStage<1, [P7_DU1, P7_DU2,
+                                                  P7_DU3, P7_DU4], 0>,
+                                   InstrStage<1, [P7_VS1, P7_VS2]>],
+                                  [5, 1, 1]>,
   InstrItinData<IIC_FPCompare   , [InstrStage<1, [P7_DU1, P7_DU2,
                                                   P7_DU3, P7_DU4], 0>,
                                    InstrStage<1, [P7_VS1, P7_VS2]>],
diff --git a/llvm/lib/Target/PowerPC/PPCScheduleP8.td b/llvm/lib/Target/PowerPC/PPCScheduleP8.td
index 020739baec3..69e6d05c660 100644
--- a/llvm/lib/Target/PowerPC/PPCScheduleP8.td
+++ b/llvm/lib/Target/PowerPC/PPCScheduleP8.td
@@ -323,6 +323,10 @@ def P8Itineraries : ProcessorItineraries<
                                                   P8_DU4, P8_DU5, P8_DU6], 0>,
                                    InstrStage<1, [P8_FPU1, P8_FPU2]>],
                                   [5, 1, 1]>,
+  InstrItinData<IIC_FPAddSub    , [InstrStage<1, [P8_DU1, P8_DU2, P8_DU3,
+                                                  P8_DU4, P8_DU5, P8_DU6], 0>,
+                                   InstrStage<1, [P8_FPU1, P8_FPU2]>],
+                                  [5, 1, 1]>,
   InstrItinData<IIC_FPCompare   , [InstrStage<1, [P8_DU1, P8_DU2, P8_DU3,
                                                   P8_DU4, P8_DU5, P8_DU6], 0>,
                                    InstrStage<1, [P8_FPU1, P8_FPU2]>],
diff --git a/llvm/lib/Target/PowerPC/PPCVSXFMAMutate.cpp b/llvm/lib/Target/PowerPC/PPCVSXFMAMutate.cpp
index f352fa647ac..58d3c3d3fa2 100644
--- a/llvm/lib/Target/PowerPC/PPCVSXFMAMutate.cpp
+++ b/llvm/lib/Target/PowerPC/PPCVSXFMAMutate.cpp
@@ -136,6 +136,16 @@ protected:
         // source of the copy, it must still be live here.  We can't use
         // interval testing for a physical register, so as long as we're
         // walking the MIs we may as well test liveness here.
+        //
+        // FIXME: There is a case that occurs in practice, like this:
+        //   %vreg9<def> = COPY %F1; VSSRC:%vreg9
+        //   ...
+        //   %vreg6<def> = COPY %vreg9; VSSRC:%vreg6,%vreg9
+        //   %vreg7<def> = COPY %vreg9; VSSRC:%vreg7,%vreg9
+        //   %vreg9<def,tied1> = XSMADDASP %vreg9<tied0>, %vreg1, %vreg4; VSSRC:
+        //   %vreg6<def,tied1> = XSMADDASP %vreg6<tied0>, %vreg1, %vreg2; VSSRC:
+        //   %vreg7<def,tied1> = XSMADDASP %vreg7<tied0>, %vreg1, %vreg3; VSSRC:
+        // which prevents an otherwise-profitable transformation.
         bool OtherUsers = false, KillsAddendSrc = false;
         for (auto J = std::prev(I), JE = MachineBasicBlock::iterator(AddendMI);
              J != JE; --J) {