Support itineraries in TargetSubtargetInfo::getSchedInfoStr

[avt77]

Summary

Now we support itineraries data in calcultion of latency:throughput (when we're going to print out the sched info). As result Atom schedul model shows its current numbers. I'm not sure those numbers are correct but the engine is here and we can move Atom from the old structures to the new ones.

[RKSimon]

lib/CodeGen/TargetSubtargetInfo.cpp

106

Cleanup the the nested ifs?

if (TSchedModel.hasInstrSchedModel())
  Latency = TSchedModel.computeInstrLatency(MCI.getOpcode());
else if (TSchedModel.hasInstrItineraries()) {
  auto *ItinData = TSchedModel.getInstrItineraries();
  Latency = ItinData->getStageLatency(
      getInstrInfo()->get(MCI.getOpcode()).getSchedClass());
} else
    return std::string();

[avt77]

The note mentioned by Simon fixed.

[RKSimon]

LGTM - thanks

[RKSimon]

rL309666

lib/CodeGen/TargetSubtargetInfo.cpp

 //===- TargetSubtargetInfo.cpp - General Target Information ----------------==//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 /// \file This file describes the general parts of a Subtarget.
 //
 //===----------------------------------------------------------------------===//
 
+#include "llvm/Target/TargetSubtargetInfo.h"
 #include "llvm/ADT/Optional.h"
 #include "llvm/CodeGen/MachineInstr.h"
 #include "llvm/CodeGen/TargetSchedule.h"
 #include "llvm/MC/MCInst.h"
-#include "llvm/Target/TargetSubtargetInfo.h"
 #include "llvm/Support/Format.h"
 #include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetInstrInfo.h"
 #include <string>
 
 using namespace llvm;
 
 TargetSubtargetInfo::TargetSubtargetInfo(
     const Triple &TT, StringRef CPU, StringRef FS,
     ArrayRef<SubtargetFeatureKV> PF, ArrayRef<SubtargetFeatureKV> PD,
     const SubtargetInfoKV *ProcSched, const MCWriteProcResEntry *WPR,

 }
 
 /// Returns string representation of scheduler comment
 std::string TargetSubtargetInfo::getSchedInfoStr(MCInst const &MCI) const {
   // We don't cache TSchedModel because it depends on TargetInstrInfo
   // that could be changed during the compilation
   TargetSchedModel TSchedModel;
   TSchedModel.init(getSchedModel(), this, getInstrInfo());
-  if (!TSchedModel.hasInstrSchedModel())
+  unsigned Latency;
+  if (!TSchedModel.hasInstrSchedModel()) {
+    if (TSchedModel.hasInstrItineraries()) {
+      auto *ItinData = TSchedModel.getInstrItineraries();
+      Latency = ItinData->getStageLatency(
+          getInstrInfo()->get(MCI.getOpcode()).getSchedClass());
+    } else
       return std::string();
-  unsigned Latency = TSchedModel.computeInstrLatency(MCI.getOpcode());
+  } else
+    Latency = TSchedModel.computeInstrLatency(MCI.getOpcode());
   Optional<double> RThroughput =
       TSchedModel.computeInstrRThroughput(MCI.getOpcode());
   return createSchedInfoStr(Latency, RThroughput);
 }

test/CodeGen/X86/lea32-schedule.ll

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal -24(%rdi), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_offset:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal -24(%rdi), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal -24(%rdi), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_offset:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal -24(%rdi), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_offset:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal 1024(%rdi), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_offset_big:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal 1024(%rdi), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal 1024(%rdi), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_offset_big:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal 1024(%rdi), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_offset_big:

 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal (%rdi,%rsi), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal (%rdi,%rsi), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal (%rdi,%rsi), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal (%rdi,%rsi), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal 16(%rdi,%rsi), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add_offset:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal 16(%rdi,%rsi), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal 16(%rdi,%rsi), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add_offset:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal 16(%rdi,%rsi), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal -4096(%rdi,%rsi), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add_offset_big:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal -4096(%rdi,%rsi), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal -4096(%rdi,%rsi), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add_offset_big:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal -4096(%rdi,%rsi), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal (%rdi,%rdi,2), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_mul:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal (%rdi,%rdi,2), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal (%rdi,%rdi,2), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_mul:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal (%rdi,%rdi,2), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_mul:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal -32(%rdi,%rdi,2), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_mul_offset:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal -32(%rdi,%rdi,2), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal -32(%rdi,%rdi,2), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_mul_offset:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal -32(%rdi,%rdi,2), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_mul_offset:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal 10000(%rdi,%rdi,8), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_mul_offset_big:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal 10000(%rdi,%rdi,8), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal 10000(%rdi,%rdi,8), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_mul_offset_big:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal 10000(%rdi,%rdi,8), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_mul_offset_big:

 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal (%rdi,%rsi,2), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add_scale:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal (%rdi,%rsi,2), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal (%rdi,%rsi,2), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add_scale:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal (%rdi,%rsi,2), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal 96(%rdi,%rsi,4), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add_scale_offset:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal 96(%rdi,%rsi,4), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal 96(%rdi,%rsi,4), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add_scale_offset:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal 96(%rdi,%rsi,4), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; GENERIC-NEXT:    leal -1200(%rdi,%rsi,8), %eax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add_scale_offset_big:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; ATOM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
-; ATOM-NEXT:    leal -1200(%rdi,%rsi,8), %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leal -1200(%rdi,%rsi,8), %eax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add_scale_offset_big:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    # kill: %ESI<def> %ESI<kill> %RSI<def>
 ; SLM-NEXT:    # kill: %EDI<def> %EDI<kill> %RDI<def>
 ; SLM-NEXT:    leal -1200(%rdi,%rsi,8), %eax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

test/CodeGen/X86/lea64-schedule.ll

 define i64 @test_lea_offset(i64) {
 ; GENERIC-LABEL: test_lea_offset:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq -24(%rdi), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_offset:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq -24(%rdi), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq -24(%rdi), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_offset:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq -24(%rdi), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_offset:
 ; SANDY:       # BB#0:

 define i64 @test_lea_offset_big(i64) {
 ; GENERIC-LABEL: test_lea_offset_big:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq 1024(%rdi), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_offset_big:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq 1024(%rdi), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq 1024(%rdi), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_offset_big:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq 1024(%rdi), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_offset_big:
 ; SANDY:       # BB#0:

 define i64 @test_lea_add(i64, i64) {
 ; GENERIC-LABEL: test_lea_add:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq (%rdi,%rsi), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq (%rdi,%rsi), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq (%rdi,%rsi), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq (%rdi,%rsi), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_add:
 ; SANDY:       # BB#0:

 define i64 @test_lea_add_offset(i64, i64) {
 ; GENERIC-LABEL: test_lea_add_offset:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq 16(%rdi,%rsi), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add_offset:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq 16(%rdi,%rsi), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq 16(%rdi,%rsi), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add_offset:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq 16(%rdi,%rsi), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_add_offset:
 ; SANDY:       # BB#0:

 define i64 @test_lea_add_offset_big(i64, i64) {
 ; GENERIC-LABEL: test_lea_add_offset_big:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq -4096(%rdi,%rsi), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add_offset_big:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq -4096(%rdi,%rsi), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq -4096(%rdi,%rsi), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add_offset_big:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq -4096(%rdi,%rsi), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_add_offset_big:
 ; SANDY:       # BB#0:

 define i64 @test_lea_mul(i64) {
 ; GENERIC-LABEL: test_lea_mul:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq (%rdi,%rdi,2), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_mul:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq (%rdi,%rdi,2), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq (%rdi,%rdi,2), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_mul:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq (%rdi,%rdi,2), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_mul:
 ; SANDY:       # BB#0:

 define i64 @test_lea_mul_offset(i64) {
 ; GENERIC-LABEL: test_lea_mul_offset:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq -32(%rdi,%rdi,2), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_mul_offset:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq -32(%rdi,%rdi,2), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq -32(%rdi,%rdi,2), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_mul_offset:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq -32(%rdi,%rdi,2), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_mul_offset:
 ; SANDY:       # BB#0:

 define i64 @test_lea_mul_offset_big(i64) {
 ; GENERIC-LABEL: test_lea_mul_offset_big:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq 10000(%rdi,%rdi,8), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_mul_offset_big:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq 10000(%rdi,%rdi,8), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq 10000(%rdi,%rdi,8), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_mul_offset_big:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq 10000(%rdi,%rdi,8), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_mul_offset_big:
 ; SANDY:       # BB#0:

 define i64 @test_lea_add_scale(i64, i64) {
 ; GENERIC-LABEL: test_lea_add_scale:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq (%rdi,%rsi,2), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add_scale:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq (%rdi,%rsi,2), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq (%rdi,%rsi,2), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add_scale:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq (%rdi,%rsi,2), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_add_scale:
 ; SANDY:       # BB#0:

 define i64 @test_lea_add_scale_offset(i64, i64) {
 ; GENERIC-LABEL: test_lea_add_scale_offset:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq 96(%rdi,%rsi,4), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add_scale_offset:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq 96(%rdi,%rsi,4), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq 96(%rdi,%rsi,4), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add_scale_offset:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq 96(%rdi,%rsi,4), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_add_scale_offset:
 ; SANDY:       # BB#0:

 define i64 @test_lea_add_scale_offset_big(i64, i64) {
 ; GENERIC-LABEL: test_lea_add_scale_offset_big:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    leaq -1200(%rdi,%rsi,8), %rax # sched: [1:0.50]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_lea_add_scale_offset_big:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    leaq -1200(%rdi,%rsi,8), %rax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    leaq -1200(%rdi,%rsi,8), %rax # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lea_add_scale_offset_big:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    leaq -1200(%rdi,%rsi,8), %rax # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lea_add_scale_offset_big:
 ; SANDY:       # BB#0:

test/CodeGen/X86/sse-schedule.ll

 ; GENERIC-LABEL: test_addps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    addps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_addps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    addps %xmm1, %xmm0
-; ATOM-NEXT:    addps (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    addps (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_addps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    addps (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_addps:

 ; GENERIC-LABEL: test_addss:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    addss %xmm1, %xmm0
 ; GENERIC-NEXT:    addss (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_addss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    addss %xmm1, %xmm0
-; ATOM-NEXT:    addss (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addss %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    addss (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_addss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    addss (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_addss:

 ; GENERIC-NEXT:    andps %xmm1, %xmm0
 ; GENERIC-NEXT:    andps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_andps:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    andps %xmm1, %xmm0
 ; ATOM-NEXT:    andps (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_andps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    andps %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    andps (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_andps:

 ; GENERIC-NEXT:    andnps %xmm1, %xmm0
 ; GENERIC-NEXT:    andnps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_andnotps:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    andnps %xmm1, %xmm0
 ; ATOM-NEXT:    andnps (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_andnotps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    andnps %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    andnps (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_andnotps:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cmpeqps %xmm0, %xmm1
 ; GENERIC-NEXT:    cmpeqps (%rdi), %xmm0
 ; GENERIC-NEXT:    orps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cmpps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cmpeqps %xmm0, %xmm1
-; ATOM-NEXT:    cmpeqps (%rdi), %xmm0
+; ATOM-NEXT:    cmpeqps %xmm0, %xmm1 # sched: [5:5.00]
+; ATOM-NEXT:    cmpeqps (%rdi), %xmm0 # sched: [5:5.00]
 ; ATOM-NEXT:    orps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cmpps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cmpeqps %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    cmpeqps (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_cmpss:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cmpeqss %xmm1, %xmm0
 ; GENERIC-NEXT:    cmpeqss (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cmpss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cmpeqss %xmm1, %xmm0
-; ATOM-NEXT:    cmpeqss (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cmpeqss %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    cmpeqss (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cmpss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cmpeqss %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    cmpeqss (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_cmpss:

 ; GENERIC-NEXT:    sete %dl
 ; GENERIC-NEXT:    andb %al, %dl
 ; GENERIC-NEXT:    orb %cl, %dl
 ; GENERIC-NEXT:    movzbl %dl, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_comiss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    comiss %xmm1, %xmm0
-; ATOM-NEXT:    setnp %al
-; ATOM-NEXT:    sete %cl
-; ATOM-NEXT:    andb %al, %cl
-; ATOM-NEXT:    comiss (%rdi), %xmm0
-; ATOM-NEXT:    setnp %al
-; ATOM-NEXT:    sete %dl
-; ATOM-NEXT:    andb %al, %dl
-; ATOM-NEXT:    orb %cl, %dl
-; ATOM-NEXT:    movzbl %dl, %eax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    comiss %xmm1, %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    setnp %al # sched: [1:0.50]
+; ATOM-NEXT:    sete %cl # sched: [1:0.50]
+; ATOM-NEXT:    andb %al, %cl # sched: [1:0.50]
+; ATOM-NEXT:    comiss (%rdi), %xmm0 # sched: [10:5.00]
+; ATOM-NEXT:    setnp %al # sched: [1:0.50]
+; ATOM-NEXT:    sete %dl # sched: [1:0.50]
+; ATOM-NEXT:    andb %al, %dl # sched: [1:0.50]
+; ATOM-NEXT:    orb %cl, %dl # sched: [1:0.50]
+; ATOM-NEXT:    movzbl %dl, %eax # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_comiss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    comiss %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    setnp %al # sched: [1:0.50]
 ; SLM-NEXT:    sete %cl # sched: [1:0.50]
 ; SLM-NEXT:    andb %al, %cl # sched: [1:0.50]
 ; SLM-NEXT:    comiss (%rdi), %xmm0 # sched: [6:1.00]

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtsi2ssl %edi, %xmm1
 ; GENERIC-NEXT:    cvtsi2ssl (%rsi), %xmm0
 ; GENERIC-NEXT:    addss %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtsi2ss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtsi2ssl (%rsi), %xmm0
-; ATOM-NEXT:    cvtsi2ssl %edi, %xmm1
-; ATOM-NEXT:    addss %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtsi2ssl (%rsi), %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    cvtsi2ssl %edi, %xmm1 # sched: [6:3.00]
+; ATOM-NEXT:    addss %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtsi2ss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtsi2ssl (%rsi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    cvtsi2ssl %edi, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtsi2ssq %rdi, %xmm1
 ; GENERIC-NEXT:    cvtsi2ssq (%rsi), %xmm0
 ; GENERIC-NEXT:    addss %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtsi2ssq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtsi2ssq (%rsi), %xmm0
-; ATOM-NEXT:    cvtsi2ssq %rdi, %xmm1
-; ATOM-NEXT:    addss %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtsi2ssq (%rsi), %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    cvtsi2ssq %rdi, %xmm1 # sched: [6:3.00]
+; ATOM-NEXT:    addss %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtsi2ssq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtsi2ssq (%rsi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    cvtsi2ssq %rdi, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtss2si %xmm0, %ecx
 ; GENERIC-NEXT:    cvtss2si (%rdi), %eax
 ; GENERIC-NEXT:    addl %ecx, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtss2si:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtss2si (%rdi), %eax
-; ATOM-NEXT:    cvtss2si %xmm0, %ecx
-; ATOM-NEXT:    addl %ecx, %eax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtss2si (%rdi), %eax # sched: [9:4.50]
+; ATOM-NEXT:    cvtss2si %xmm0, %ecx # sched: [8:4.00]
+; ATOM-NEXT:    addl %ecx, %eax # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtss2si:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtss2si (%rdi), %eax # sched: [7:1.00]
 ; SLM-NEXT:    cvtss2si %xmm0, %ecx # sched: [4:0.50]
 ; SLM-NEXT:    addl %ecx, %eax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtss2si %xmm0, %rcx
 ; GENERIC-NEXT:    cvtss2si (%rdi), %rax
 ; GENERIC-NEXT:    addq %rcx, %rax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtss2siq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtss2si (%rdi), %rax
-; ATOM-NEXT:    cvtss2si %xmm0, %rcx
-; ATOM-NEXT:    addq %rcx, %rax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtss2si (%rdi), %rax # sched: [10:5.00]
+; ATOM-NEXT:    cvtss2si %xmm0, %rcx # sched: [9:4.50]
+; ATOM-NEXT:    addq %rcx, %rax # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtss2siq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtss2si (%rdi), %rax # sched: [7:1.00]
 ; SLM-NEXT:    cvtss2si %xmm0, %rcx # sched: [4:0.50]
 ; SLM-NEXT:    addq %rcx, %rax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvttss2si %xmm0, %ecx
 ; GENERIC-NEXT:    cvttss2si (%rdi), %eax
 ; GENERIC-NEXT:    addl %ecx, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvttss2si:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvttss2si (%rdi), %eax
-; ATOM-NEXT:    cvttss2si %xmm0, %ecx
-; ATOM-NEXT:    addl %ecx, %eax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvttss2si (%rdi), %eax # sched: [9:4.50]
+; ATOM-NEXT:    cvttss2si %xmm0, %ecx # sched: [8:4.00]
+; ATOM-NEXT:    addl %ecx, %eax # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvttss2si:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvttss2si (%rdi), %eax # sched: [7:1.00]
 ; SLM-NEXT:    cvttss2si %xmm0, %ecx # sched: [4:0.50]
 ; SLM-NEXT:    addl %ecx, %eax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvttss2si %xmm0, %rcx
 ; GENERIC-NEXT:    cvttss2si (%rdi), %rax
 ; GENERIC-NEXT:    addq %rcx, %rax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvttss2siq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvttss2si (%rdi), %rax
-; ATOM-NEXT:    cvttss2si %xmm0, %rcx
-; ATOM-NEXT:    addq %rcx, %rax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvttss2si (%rdi), %rax # sched: [10:5.00]
+; ATOM-NEXT:    cvttss2si %xmm0, %rcx # sched: [9:4.50]
+; ATOM-NEXT:    addq %rcx, %rax # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvttss2siq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvttss2si (%rdi), %rax # sched: [7:1.00]
 ; SLM-NEXT:    cvttss2si %xmm0, %rcx # sched: [4:0.50]
 ; SLM-NEXT:    addq %rcx, %rax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_divps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    divps %xmm1, %xmm0
 ; GENERIC-NEXT:    divps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_divps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    divps %xmm1, %xmm0
-; ATOM-NEXT:    divps (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    divps %xmm1, %xmm0 # sched: [70:35.00]
+; ATOM-NEXT:    divps (%rdi), %xmm0 # sched: [125:62.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_divps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    divps %xmm1, %xmm0 # sched: [34:34.00]
 ; SLM-NEXT:    divps (%rdi), %xmm0 # sched: [37:34.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_divps:

 ; GENERIC-LABEL: test_divss:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    divss %xmm1, %xmm0
 ; GENERIC-NEXT:    divss (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_divss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    divss %xmm1, %xmm0
-; ATOM-NEXT:    divss (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    divss %xmm1, %xmm0 # sched: [34:17.00]
+; ATOM-NEXT:    divss (%rdi), %xmm0 # sched: [62:31.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_divss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    divss %xmm1, %xmm0 # sched: [34:34.00]
 ; SLM-NEXT:    divss (%rdi), %xmm0 # sched: [37:34.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_divss:

 ; GENERIC-LABEL: test_ldmxcsr:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movl %edi, -{{[0-9]+}}(%rsp)
 ; GENERIC-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_ldmxcsr:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movl %edi, -{{[0-9]+}}(%rsp)
-; ATOM-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
+; ATOM-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp) # sched: [5:2.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_ldmxcsr:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movl %edi, -{{[0-9]+}}(%rsp) # sched: [1:1.00]
 ; SLM-NEXT:    ldmxcsr -{{[0-9]+}}(%rsp) # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_ldmxcsr:

 ; GENERIC-LABEL: test_maxps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    maxps %xmm1, %xmm0
 ; GENERIC-NEXT:    maxps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_maxps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    maxps %xmm1, %xmm0
-; ATOM-NEXT:    maxps (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    maxps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    maxps (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_maxps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    maxps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    maxps (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_maxps:

 ; GENERIC-LABEL: test_maxss:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    maxss %xmm1, %xmm0
 ; GENERIC-NEXT:    maxss (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_maxss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    maxss %xmm1, %xmm0
-; ATOM-NEXT:    maxss (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    maxss %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    maxss (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_maxss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    maxss %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    maxss (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_maxss:

 ; GENERIC-LABEL: test_minps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    minps %xmm1, %xmm0
 ; GENERIC-NEXT:    minps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_minps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    minps %xmm1, %xmm0
-; ATOM-NEXT:    minps (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    minps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    minps (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_minps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    minps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    minps (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_minps:

 ; GENERIC-LABEL: test_minss:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    minss %xmm1, %xmm0
 ; GENERIC-NEXT:    minss (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_minss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    minss %xmm1, %xmm0
-; ATOM-NEXT:    minss (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    minss %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    minss (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_minss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    minss %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    minss (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_minss:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movaps (%rdi), %xmm0
 ; GENERIC-NEXT:    addps %xmm0, %xmm0
 ; GENERIC-NEXT:    movaps %xmm0, (%rsi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movaps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movaps (%rdi), %xmm0
-; ATOM-NEXT:    addps %xmm0, %xmm0
-; ATOM-NEXT:    movaps %xmm0, (%rsi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movaps (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    addps %xmm0, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    movaps %xmm0, (%rsi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movaps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movaps (%rdi), %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    movaps %xmm0, (%rsi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 define <4 x float> @test_movhlps(<4 x float> %a0, <4 x float> %a1) {
 ; GENERIC-LABEL: test_movhlps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1]
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movhlps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1]
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movhlps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movhlps {{.*#+}} xmm0 = xmm1[1],xmm0[1] sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_movhlps:
 ; SANDY:       # BB#0:

 ; GENERIC-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0]
 ; GENERIC-NEXT:    addps %xmm0, %xmm1
 ; GENERIC-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
 ; GENERIC-NEXT:    movlps %xmm1, (%rdi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movhps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0]
-; ATOM-NEXT:    addps %xmm0, %xmm1
-; ATOM-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
-; ATOM-NEXT:    movlps %xmm1, (%rdi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [1:1.00]
+; ATOM-NEXT:    addps %xmm0, %xmm1 # sched: [5:5.00]
+; ATOM-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1] sched: [1:1.00]
+; ATOM-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movhps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [4:1.00]
 ; SLM-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    pextrq $1, %xmm1, (%rdi) # sched: [4:2.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_movlhps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movlhps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; ATOM-NEXT:    addps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]
+; ATOM-NEXT:    addps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movlhps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]
 ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_movlhps:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1]
 ; GENERIC-NEXT:    addps %xmm0, %xmm1
 ; GENERIC-NEXT:    movlps %xmm1, (%rdi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movlps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1]
-; ATOM-NEXT:    addps %xmm0, %xmm1
-; ATOM-NEXT:    movlps %xmm1, (%rdi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [1:1.00]
+; ATOM-NEXT:    addps %xmm0, %xmm1 # sched: [5:5.00]
+; ATOM-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movlps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [4:1.00]
 ; SLM-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    movlps %xmm1, (%rdi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 define i32 @test_movmskps(<4 x float> %a0) {
 ; GENERIC-LABEL: test_movmskps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movmskps %xmm0, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movmskps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movmskps %xmm0, %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movmskps %xmm0, %eax # sched: [3:3.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movmskps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movmskps %xmm0, %eax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_movmskps:
 ; SANDY:       # BB#0:

 define void @test_movntps(<4 x float> %a0, <4 x float> *%a1) {
 ; GENERIC-LABEL: test_movntps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movntps %xmm0, (%rdi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movntps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movntps %xmm0, (%rdi)
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movntps %xmm0, (%rdi) # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movntps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movntps %xmm0, (%rdi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_movntps:
 ; SANDY:       # BB#0:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
 ; GENERIC-NEXT:    addss %xmm0, %xmm0
 ; GENERIC-NEXT:    movss %xmm0, (%rsi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movss_mem:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; ATOM-NEXT:    addss %xmm0, %xmm0
-; ATOM-NEXT:    movss %xmm0, (%rsi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [1:1.00]
+; ATOM-NEXT:    addss %xmm0, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    movss %xmm0, (%rsi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movss_mem:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [3:1.00]
 ; SLM-NEXT:    addss %xmm0, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    movss %xmm0, (%rsi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 define <4 x float> @test_movss_reg(<4 x float> %a0, <4 x float> %a1) {
 ; GENERIC-LABEL: test_movss_reg:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movss_reg:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movss_reg:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    blendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3] sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_movss_reg:
 ; SANDY:       # BB#0:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movups (%rdi), %xmm0
 ; GENERIC-NEXT:    addps %xmm0, %xmm0
 ; GENERIC-NEXT:    movups %xmm0, (%rsi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movups:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movups (%rdi), %xmm0
-; ATOM-NEXT:    addps %xmm0, %xmm0
-; ATOM-NEXT:    movups %xmm0, (%rsi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movups (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    addps %xmm0, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    movups %xmm0, (%rsi) # sched: [2:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movups:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movups (%rdi), %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    addps %xmm0, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    movups %xmm0, (%rsi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_mulps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    mulps %xmm1, %xmm0
 ; GENERIC-NEXT:    mulps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_mulps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    mulps %xmm1, %xmm0
-; ATOM-NEXT:    mulps (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    mulps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    mulps (%rdi), %xmm0 # sched: [10:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_mulps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    mulps %xmm1, %xmm0 # sched: [5:2.00]
 ; SLM-NEXT:    mulps (%rdi), %xmm0 # sched: [8:2.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_mulps:

 ; GENERIC-LABEL: test_mulss:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    mulss %xmm1, %xmm0
 ; GENERIC-NEXT:    mulss (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_mulss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    mulss %xmm1, %xmm0
-; ATOM-NEXT:    mulss (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    mulss %xmm1, %xmm0 # sched: [4:4.00]
+; ATOM-NEXT:    mulss (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_mulss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    mulss %xmm1, %xmm0 # sched: [5:2.00]
 ; SLM-NEXT:    mulss (%rdi), %xmm0 # sched: [8:2.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_mulss:

 ; GENERIC-NEXT:    orps %xmm1, %xmm0
 ; GENERIC-NEXT:    orps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_orps:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    orps %xmm1, %xmm0
 ; ATOM-NEXT:    orps (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_orps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    orps %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    orps (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_orps:

 define void @test_prefetchnta(i8* %a0) {
 ; GENERIC-LABEL: test_prefetchnta:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    prefetchnta (%rdi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_prefetchnta:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    prefetchnta (%rdi)
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    prefetchnta (%rdi) # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_prefetchnta:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    prefetchnta (%rdi) # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_prefetchnta:
 ; SANDY:       # BB#0:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    rcpps %xmm0, %xmm1
 ; GENERIC-NEXT:    rcpps (%rdi), %xmm0
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_rcpps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    rcpps (%rdi), %xmm1
-; ATOM-NEXT:    rcpps %xmm0, %xmm0
-; ATOM-NEXT:    addps %xmm0, %xmm1
-; ATOM-NEXT:    movaps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    rcpps (%rdi), %xmm1 # sched: [10:5.00]
+; ATOM-NEXT:    rcpps %xmm0, %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    addps %xmm0, %xmm1 # sched: [5:5.00]
+; ATOM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_rcpps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    rcpps (%rdi), %xmm1 # sched: [8:1.00]
 ; SLM-NEXT:    rcpps %xmm0, %xmm0 # sched: [5:1.00]
 ; SLM-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC-NEXT:    rcpss %xmm0, %xmm0
 ; GENERIC-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
 ; GENERIC-NEXT:    rcpss %xmm1, %xmm1
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_rcpss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; ATOM-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:1.00]
 ; ATOM-NEXT:    rcpss %xmm0, %xmm0
 ; ATOM-NEXT:    rcpss %xmm1, %xmm1
-; ATOM-NEXT:    addps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_rcpss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [3:1.00]
 ; SLM-NEXT:    rcpss %xmm0, %xmm0 # sched: [8:1.00]
 ; SLM-NEXT:    rcpss %xmm1, %xmm1 # sched: [8:1.00]
 ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    rsqrtps %xmm0, %xmm1
 ; GENERIC-NEXT:    rsqrtps (%rdi), %xmm0
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_rsqrtps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    rsqrtps (%rdi), %xmm1
-; ATOM-NEXT:    rsqrtps %xmm0, %xmm0
-; ATOM-NEXT:    addps %xmm0, %xmm1
-; ATOM-NEXT:    movaps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    rsqrtps (%rdi), %xmm1 # sched: [10:5.00]
+; ATOM-NEXT:    rsqrtps %xmm0, %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    addps %xmm0, %xmm1 # sched: [5:5.00]
+; ATOM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_rsqrtps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    rsqrtps (%rdi), %xmm1 # sched: [8:1.00]
 ; SLM-NEXT:    rsqrtps %xmm0, %xmm0 # sched: [5:1.00]
 ; SLM-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC-NEXT:    rsqrtss %xmm0, %xmm0
 ; GENERIC-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
 ; GENERIC-NEXT:    rsqrtss %xmm1, %xmm1
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_rsqrtss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; ATOM-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:1.00]
 ; ATOM-NEXT:    rsqrtss %xmm0, %xmm0
 ; ATOM-NEXT:    rsqrtss %xmm1, %xmm1
-; ATOM-NEXT:    addps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_rsqrtss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [3:1.00]
 ; SLM-NEXT:    rsqrtss %xmm0, %xmm0 # sched: [8:1.00]
 ; SLM-NEXT:    rsqrtss %xmm1, %xmm1 # sched: [8:1.00]
 ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 define void @test_sfence() {
 ; GENERIC-LABEL: test_sfence:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    sfence
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_sfence:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    sfence
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    sfence # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_sfence:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    sfence # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_sfence:
 ; SANDY:       # BB#0:

 ; GENERIC-LABEL: test_shufps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0]
 ; GENERIC-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,3],mem[0,0]
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_shufps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0]
-; ATOM-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,3],mem[0,0]
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00]
+; ATOM-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,3],mem[0,0] sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_shufps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0] sched: [1:1.00]
 ; SLM-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,3],mem[0,0] sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_shufps:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    sqrtps %xmm0, %xmm1
 ; GENERIC-NEXT:    sqrtps (%rdi), %xmm0
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_sqrtps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    sqrtps %xmm0, %xmm1
-; ATOM-NEXT:    sqrtps (%rdi), %xmm0
-; ATOM-NEXT:    addps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    sqrtps %xmm0, %xmm1 # sched: [70:35.00]
+; ATOM-NEXT:    sqrtps (%rdi), %xmm0 # sched: [70:35.00]
+; ATOM-NEXT:    addps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_sqrtps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    sqrtps (%rdi), %xmm1 # sched: [18:1.00]
 ; SLM-NEXT:    sqrtps %xmm0, %xmm0 # sched: [15:1.00]
 ; SLM-NEXT:    addps %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC-NEXT:    sqrtss %xmm0, %xmm0
 ; GENERIC-NEXT:    movaps (%rdi), %xmm1
 ; GENERIC-NEXT:    sqrtss %xmm1, %xmm1
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_sqrtss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movaps (%rdi), %xmm1
+; ATOM-NEXT:    movaps (%rdi), %xmm1 # sched: [1:1.00]
 ; ATOM-NEXT:    sqrtss %xmm0, %xmm0
 ; ATOM-NEXT:    sqrtss %xmm1, %xmm1
-; ATOM-NEXT:    addps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_sqrtss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movaps (%rdi), %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    sqrtss %xmm0, %xmm0 # sched: [18:1.00]
 ; SLM-NEXT:    sqrtss %xmm1, %xmm1 # sched: [18:1.00]
 ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC-LABEL: test_stmxcsr:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    stmxcsr -{{[0-9]+}}(%rsp)
 ; GENERIC-NEXT:    movl -{{[0-9]+}}(%rsp), %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_stmxcsr:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    stmxcsr -{{[0-9]+}}(%rsp)
-; ATOM-NEXT:    movl -{{[0-9]+}}(%rsp), %eax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    stmxcsr -{{[0-9]+}}(%rsp) # sched: [15:7.50]
+; ATOM-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_stmxcsr:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    stmxcsr -{{[0-9]+}}(%rsp) # sched: [1:1.00]
 ; SLM-NEXT:    movl -{{[0-9]+}}(%rsp), %eax # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_stmxcsr:

 ; GENERIC-LABEL: test_subps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    subps %xmm1, %xmm0
 ; GENERIC-NEXT:    subps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_subps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    subps %xmm1, %xmm0
-; ATOM-NEXT:    subps (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    subps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    subps (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_subps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    subps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    subps (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_subps:

 ; GENERIC-LABEL: test_subss:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    subss %xmm1, %xmm0
 ; GENERIC-NEXT:    subss (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_subss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    subss %xmm1, %xmm0
-; ATOM-NEXT:    subss (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    subss %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    subss (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_subss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    subss %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    subss (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_subss:

 ; GENERIC-NEXT:    sete %dl
 ; GENERIC-NEXT:    andb %al, %dl
 ; GENERIC-NEXT:    orb %cl, %dl
 ; GENERIC-NEXT:    movzbl %dl, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_ucomiss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    ucomiss %xmm1, %xmm0
-; ATOM-NEXT:    setnp %al
-; ATOM-NEXT:    sete %cl
-; ATOM-NEXT:    andb %al, %cl
-; ATOM-NEXT:    ucomiss (%rdi), %xmm0
-; ATOM-NEXT:    setnp %al
-; ATOM-NEXT:    sete %dl
-; ATOM-NEXT:    andb %al, %dl
-; ATOM-NEXT:    orb %cl, %dl
-; ATOM-NEXT:    movzbl %dl, %eax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    ucomiss %xmm1, %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    setnp %al # sched: [1:0.50]
+; ATOM-NEXT:    sete %cl # sched: [1:0.50]
+; ATOM-NEXT:    andb %al, %cl # sched: [1:0.50]
+; ATOM-NEXT:    ucomiss (%rdi), %xmm0 # sched: [10:5.00]
+; ATOM-NEXT:    setnp %al # sched: [1:0.50]
+; ATOM-NEXT:    sete %dl # sched: [1:0.50]
+; ATOM-NEXT:    andb %al, %dl # sched: [1:0.50]
+; ATOM-NEXT:    orb %cl, %dl # sched: [1:0.50]
+; ATOM-NEXT:    movzbl %dl, %eax # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_ucomiss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    ucomiss %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    setnp %al # sched: [1:0.50]
 ; SLM-NEXT:    sete %cl # sched: [1:0.50]
 ; SLM-NEXT:    andb %al, %cl # sched: [1:0.50]
 ; SLM-NEXT:    ucomiss (%rdi), %xmm0 # sched: [6:1.00]

 ; GENERIC-LABEL: test_unpckhps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
 ; GENERIC-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_unpckhps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; ATOM-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],mem[2],xmm0[3],mem[3]
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00]
+; ATOM-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],mem[2],xmm0[3],mem[3] sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_unpckhps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00]
 ; SLM-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],mem[2],xmm0[3],mem[3] sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_unpckhps:

 ; GENERIC-LABEL: test_unpcklps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; GENERIC-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_unpcklps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; ATOM-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00]
+; ATOM-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_unpcklps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00]
 ; SLM-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_unpcklps:

 ; GENERIC-NEXT:    xorps %xmm1, %xmm0
 ; GENERIC-NEXT:    xorps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_xorps:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    xorps %xmm1, %xmm0
 ; ATOM-NEXT:    xorps (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_xorps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    xorps %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    xorps (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_xorps:

test/CodeGen/X86/sse2-schedule.ll

 ; GENERIC-LABEL: test_addpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    addpd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_addpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    addpd %xmm1, %xmm0
-; ATOM-NEXT:    addpd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    addpd (%rdi), %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_addpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    addpd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_addpd:

 ; GENERIC-LABEL: test_addsd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    addsd %xmm1, %xmm0
 ; GENERIC-NEXT:    addsd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_addsd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    addsd %xmm1, %xmm0
-; ATOM-NEXT:    addsd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addsd %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    addsd (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_addsd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    addsd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    addsd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_addsd:

 ; GENERIC-NEXT:    andpd (%rdi), %xmm0
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_andpd:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    andpd %xmm1, %xmm0
 ; ATOM-NEXT:    andpd (%rdi), %xmm0
-; ATOM-NEXT:    addpd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_andpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    andpd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    andpd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    andnpd (%rdi), %xmm0
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_andnotpd:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    andnpd %xmm1, %xmm0
 ; ATOM-NEXT:    andnpd (%rdi), %xmm0
-; ATOM-NEXT:    addpd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_andnotpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    andnpd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    andnpd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cmpeqpd %xmm0, %xmm1
 ; GENERIC-NEXT:    cmpeqpd (%rdi), %xmm0
 ; GENERIC-NEXT:    orpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cmppd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cmpeqpd %xmm0, %xmm1
-; ATOM-NEXT:    cmpeqpd (%rdi), %xmm0
+; ATOM-NEXT:    cmpeqpd %xmm0, %xmm1 # sched: [6:3.00]
+; ATOM-NEXT:    cmpeqpd (%rdi), %xmm0 # sched: [7:3.50]
 ; ATOM-NEXT:    orpd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cmppd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cmpeqpd %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    cmpeqpd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    orpd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_cmpsd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cmpeqsd %xmm1, %xmm0
 ; GENERIC-NEXT:    cmpeqsd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cmpsd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cmpeqsd %xmm1, %xmm0
-; ATOM-NEXT:    cmpeqsd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cmpeqsd %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    cmpeqsd (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cmpsd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cmpeqsd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    cmpeqsd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_cmpsd:

 ; GENERIC-NEXT:    sete %dl
 ; GENERIC-NEXT:    andb %al, %dl
 ; GENERIC-NEXT:    orb %cl, %dl
 ; GENERIC-NEXT:    movzbl %dl, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_comisd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    comisd %xmm1, %xmm0
-; ATOM-NEXT:    setnp %al
-; ATOM-NEXT:    sete %cl
-; ATOM-NEXT:    andb %al, %cl
-; ATOM-NEXT:    comisd (%rdi), %xmm0
-; ATOM-NEXT:    setnp %al
-; ATOM-NEXT:    sete %dl
-; ATOM-NEXT:    andb %al, %dl
-; ATOM-NEXT:    orb %cl, %dl
-; ATOM-NEXT:    movzbl %dl, %eax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    comisd %xmm1, %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    setnp %al # sched: [1:0.50]
+; ATOM-NEXT:    sete %cl # sched: [1:0.50]
+; ATOM-NEXT:    andb %al, %cl # sched: [1:0.50]
+; ATOM-NEXT:    comisd (%rdi), %xmm0 # sched: [10:5.00]
+; ATOM-NEXT:    setnp %al # sched: [1:0.50]
+; ATOM-NEXT:    sete %dl # sched: [1:0.50]
+; ATOM-NEXT:    andb %al, %dl # sched: [1:0.50]
+; ATOM-NEXT:    orb %cl, %dl # sched: [1:0.50]
+; ATOM-NEXT:    movzbl %dl, %eax # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_comisd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    comisd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    setnp %al # sched: [1:0.50]
 ; SLM-NEXT:    sete %cl # sched: [1:0.50]
 ; SLM-NEXT:    andb %al, %cl # sched: [1:0.50]
 ; SLM-NEXT:    comisd (%rdi), %xmm0 # sched: [6:1.00]

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtdq2pd %xmm0, %xmm1
 ; GENERIC-NEXT:    cvtdq2pd (%rdi), %xmm0
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtdq2pd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtdq2pd %xmm0, %xmm1
-; ATOM-NEXT:    cvtdq2pd (%rdi), %xmm0
-; ATOM-NEXT:    addpd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtdq2pd %xmm0, %xmm1 # sched: [8:4.00]
+; ATOM-NEXT:    cvtdq2pd (%rdi), %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    addpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtdq2pd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtdq2pd %xmm0, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    cvtdq2pd (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtdq2ps %xmm0, %xmm1
 ; GENERIC-NEXT:    cvtdq2ps (%rdi), %xmm0
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtdq2ps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtdq2ps (%rdi), %xmm1
-; ATOM-NEXT:    cvtdq2ps %xmm0, %xmm0
-; ATOM-NEXT:    addps %xmm0, %xmm1
-; ATOM-NEXT:    movaps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtdq2ps (%rdi), %xmm1 # sched: [7:3.50]
+; ATOM-NEXT:    cvtdq2ps %xmm0, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    addps %xmm0, %xmm1 # sched: [5:5.00]
+; ATOM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtdq2ps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtdq2ps %xmm0, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    cvtdq2ps (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtpd2dq %xmm0, %xmm1
 ; GENERIC-NEXT:    cvtpd2dq (%rdi), %xmm0
 ; GENERIC-NEXT:    paddd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtpd2dq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtpd2dq (%rdi), %xmm1
-; ATOM-NEXT:    cvtpd2dq %xmm0, %xmm0
-; ATOM-NEXT:    paddd %xmm0, %xmm1
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtpd2dq (%rdi), %xmm1 # sched: [8:4.00]
+; ATOM-NEXT:    cvtpd2dq %xmm0, %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    paddd %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtpd2dq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtpd2dq %xmm0, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    cvtpd2dq (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    paddd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtpd2ps %xmm0, %xmm1
 ; GENERIC-NEXT:    cvtpd2ps (%rdi), %xmm0
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtpd2ps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtpd2ps (%rdi), %xmm1
-; ATOM-NEXT:    cvtpd2ps %xmm0, %xmm0
-; ATOM-NEXT:    addps %xmm0, %xmm1
-; ATOM-NEXT:    movaps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtpd2ps (%rdi), %xmm1 # sched: [8:4.00]
+; ATOM-NEXT:    cvtpd2ps %xmm0, %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    addps %xmm0, %xmm1 # sched: [5:5.00]
+; ATOM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtpd2ps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtpd2ps %xmm0, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    cvtpd2ps (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtps2dq %xmm0, %xmm1
 ; GENERIC-NEXT:    cvtps2dq (%rdi), %xmm0
 ; GENERIC-NEXT:    paddd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtps2dq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtps2dq (%rdi), %xmm1
-; ATOM-NEXT:    cvtps2dq %xmm0, %xmm0
-; ATOM-NEXT:    paddd %xmm0, %xmm1
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtps2dq (%rdi), %xmm1 # sched: [7:3.50]
+; ATOM-NEXT:    cvtps2dq %xmm0, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    paddd %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtps2dq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtps2dq %xmm0, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    cvtps2dq (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    paddd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtps2pd %xmm0, %xmm1
 ; GENERIC-NEXT:    cvtps2pd (%rdi), %xmm0
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtps2pd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtps2pd (%rdi), %xmm1
-; ATOM-NEXT:    cvtps2pd %xmm0, %xmm0
-; ATOM-NEXT:    addpd %xmm0, %xmm1
-; ATOM-NEXT:    movapd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtps2pd (%rdi), %xmm1 # sched: [8:4.00]
+; ATOM-NEXT:    cvtps2pd %xmm0, %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    addpd %xmm0, %xmm1 # sched: [6:3.00]
+; ATOM-NEXT:    movapd %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtps2pd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtps2pd %xmm0, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    cvtps2pd (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtsd2si %xmm0, %ecx
 ; GENERIC-NEXT:    cvtsd2si (%rdi), %eax
 ; GENERIC-NEXT:    addl %ecx, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtsd2si:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtsd2si (%rdi), %eax
-; ATOM-NEXT:    cvtsd2si %xmm0, %ecx
-; ATOM-NEXT:    addl %ecx, %eax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtsd2si (%rdi), %eax # sched: [9:4.50]
+; ATOM-NEXT:    cvtsd2si %xmm0, %ecx # sched: [8:4.00]
+; ATOM-NEXT:    addl %ecx, %eax # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtsd2si:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtsd2si (%rdi), %eax # sched: [7:1.00]
 ; SLM-NEXT:    cvtsd2si %xmm0, %ecx # sched: [4:0.50]
 ; SLM-NEXT:    addl %ecx, %eax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtsd2si %xmm0, %rcx
 ; GENERIC-NEXT:    cvtsd2si (%rdi), %rax
 ; GENERIC-NEXT:    addq %rcx, %rax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtsd2siq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtsd2si (%rdi), %rax
-; ATOM-NEXT:    cvtsd2si %xmm0, %rcx
-; ATOM-NEXT:    addq %rcx, %rax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtsd2si (%rdi), %rax # sched: [9:4.50]
+; ATOM-NEXT:    cvtsd2si %xmm0, %rcx # sched: [8:4.00]
+; ATOM-NEXT:    addq %rcx, %rax # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtsd2siq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtsd2si (%rdi), %rax # sched: [7:1.00]
 ; SLM-NEXT:    cvtsd2si %xmm0, %rcx # sched: [4:0.50]
 ; SLM-NEXT:    addq %rcx, %rax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    cvtsd2ss %xmm0, %xmm1
 ; GENERIC-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
 ; GENERIC-NEXT:    cvtsd2ss %xmm0, %xmm0
 ; GENERIC-NEXT:    addss %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtsd2ss:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero
-; ATOM-NEXT:    cvtsd2ss %xmm0, %xmm2
+; ATOM-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero sched: [1:1.00]
+; ATOM-NEXT:    cvtsd2ss %xmm0, %xmm2 # sched: [6:3.00]
 ; ATOM-NEXT:    xorps %xmm0, %xmm0
-; ATOM-NEXT:    cvtsd2ss %xmm1, %xmm0
-; ATOM-NEXT:    addss %xmm2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtsd2ss %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    addss %xmm2, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtsd2ss:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtsd2ss %xmm0, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero sched: [3:1.00]
 ; SLM-NEXT:    cvtsd2ss %xmm0, %xmm0 # sched: [4:0.50]
 ; SLM-NEXT:    addss %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtsi2sdl %edi, %xmm1
 ; GENERIC-NEXT:    cvtsi2sdl (%rsi), %xmm0
 ; GENERIC-NEXT:    addsd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtsi2sd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtsi2sdl (%rsi), %xmm0
-; ATOM-NEXT:    cvtsi2sdl %edi, %xmm1
-; ATOM-NEXT:    addsd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtsi2sdl (%rsi), %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    cvtsi2sdl %edi, %xmm1 # sched: [6:3.00]
+; ATOM-NEXT:    addsd %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtsi2sd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtsi2sdl (%rsi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    cvtsi2sdl %edi, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    addsd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvtsi2sdq %rdi, %xmm1
 ; GENERIC-NEXT:    cvtsi2sdq (%rsi), %xmm0
 ; GENERIC-NEXT:    addsd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtsi2sdq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvtsi2sdq (%rsi), %xmm0
-; ATOM-NEXT:    cvtsi2sdq %rdi, %xmm1
-; ATOM-NEXT:    addsd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtsi2sdq (%rsi), %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    cvtsi2sdq %rdi, %xmm1 # sched: [6:3.00]
+; ATOM-NEXT:    addsd %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtsi2sdq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtsi2sdq (%rsi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    cvtsi2sdq %rdi, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    addsd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    cvtss2sd %xmm0, %xmm1
 ; GENERIC-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
 ; GENERIC-NEXT:    cvtss2sd %xmm0, %xmm0
 ; GENERIC-NEXT:    addsd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvtss2sd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; ATOM-NEXT:    cvtss2sd %xmm0, %xmm2
+; ATOM-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:1.00]
+; ATOM-NEXT:    cvtss2sd %xmm0, %xmm2 # sched: [6:3.00]
 ; ATOM-NEXT:    xorps %xmm0, %xmm0
-; ATOM-NEXT:    cvtss2sd %xmm1, %xmm0
-; ATOM-NEXT:    addsd %xmm2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvtss2sd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    addsd %xmm2, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvtss2sd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvtss2sd %xmm0, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero sched: [3:1.00]
 ; SLM-NEXT:    cvtss2sd %xmm0, %xmm0 # sched: [4:0.50]
 ; SLM-NEXT:    addsd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvttpd2dq %xmm0, %xmm1
 ; GENERIC-NEXT:    cvttpd2dq (%rdi), %xmm0
 ; GENERIC-NEXT:    paddd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvttpd2dq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvttpd2dq (%rdi), %xmm1
-; ATOM-NEXT:    cvttpd2dq %xmm0, %xmm0
-; ATOM-NEXT:    paddd %xmm0, %xmm1
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvttpd2dq (%rdi), %xmm1 # sched: [8:4.00]
+; ATOM-NEXT:    cvttpd2dq %xmm0, %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    paddd %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvttpd2dq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvttpd2dq %xmm0, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    cvttpd2dq (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    paddd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvttps2dq %xmm0, %xmm1
 ; GENERIC-NEXT:    cvttps2dq (%rdi), %xmm0
 ; GENERIC-NEXT:    paddd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvttps2dq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvttps2dq (%rdi), %xmm1
-; ATOM-NEXT:    cvttps2dq %xmm0, %xmm0
-; ATOM-NEXT:    paddd %xmm0, %xmm1
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvttps2dq (%rdi), %xmm1 # sched: [7:3.50]
+; ATOM-NEXT:    cvttps2dq %xmm0, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    paddd %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvttps2dq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvttps2dq %xmm0, %xmm1 # sched: [4:0.50]
 ; SLM-NEXT:    cvttps2dq (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    paddd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvttsd2si %xmm0, %ecx
 ; GENERIC-NEXT:    cvttsd2si (%rdi), %eax
 ; GENERIC-NEXT:    addl %ecx, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvttsd2si:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvttsd2si (%rdi), %eax
-; ATOM-NEXT:    cvttsd2si %xmm0, %ecx
-; ATOM-NEXT:    addl %ecx, %eax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvttsd2si (%rdi), %eax # sched: [9:4.50]
+; ATOM-NEXT:    cvttsd2si %xmm0, %ecx # sched: [8:4.00]
+; ATOM-NEXT:    addl %ecx, %eax # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvttsd2si:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvttsd2si (%rdi), %eax # sched: [7:1.00]
 ; SLM-NEXT:    cvttsd2si %xmm0, %ecx # sched: [4:0.50]
 ; SLM-NEXT:    addl %ecx, %eax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    cvttsd2si %xmm0, %rcx
 ; GENERIC-NEXT:    cvttsd2si (%rdi), %rax
 ; GENERIC-NEXT:    addq %rcx, %rax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_cvttsd2siq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    cvttsd2si (%rdi), %rax
-; ATOM-NEXT:    cvttsd2si %xmm0, %rcx
-; ATOM-NEXT:    addq %rcx, %rax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    cvttsd2si (%rdi), %rax # sched: [9:4.50]
+; ATOM-NEXT:    cvttsd2si %xmm0, %rcx # sched: [8:4.00]
+; ATOM-NEXT:    addq %rcx, %rax # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_cvttsd2siq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    cvttsd2si (%rdi), %rax # sched: [7:1.00]
 ; SLM-NEXT:    cvttsd2si %xmm0, %rcx # sched: [4:0.50]
 ; SLM-NEXT:    addq %rcx, %rax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_divpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    divpd %xmm1, %xmm0
 ; GENERIC-NEXT:    divpd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_divpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    divpd %xmm1, %xmm0
-; ATOM-NEXT:    divpd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    divpd %xmm1, %xmm0 # sched: [125:62.50]
+; ATOM-NEXT:    divpd (%rdi), %xmm0 # sched: [125:62.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_divpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    divpd %xmm1, %xmm0 # sched: [34:34.00]
 ; SLM-NEXT:    divpd (%rdi), %xmm0 # sched: [37:34.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_divpd:

 ; GENERIC-LABEL: test_divsd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    divsd %xmm1, %xmm0
 ; GENERIC-NEXT:    divsd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_divsd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    divsd %xmm1, %xmm0
-; ATOM-NEXT:    divsd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    divsd %xmm1, %xmm0 # sched: [62:31.00]
+; ATOM-NEXT:    divsd (%rdi), %xmm0 # sched: [62:31.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_divsd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    divsd %xmm1, %xmm0 # sched: [34:34.00]
 ; SLM-NEXT:    divsd (%rdi), %xmm0 # sched: [37:34.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_divsd:

 define void @test_lfence() {
 ; GENERIC-LABEL: test_lfence:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    lfence
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_lfence:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    lfence
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    lfence # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lfence:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    lfence # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lfence:
 ; SANDY:       # BB#0:

 define void @test_mfence() {
 ; GENERIC-LABEL: test_mfence:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    mfence
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_mfence:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    mfence
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    mfence # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_mfence:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    mfence # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_mfence:
 ; SANDY:       # BB#0:

 define void @test_maskmovdqu(<16 x i8> %a0, <16 x i8> %a1, i8* %a2) {
 ; GENERIC-LABEL: test_maskmovdqu:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    maskmovdqu %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_maskmovdqu:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    maskmovdqu %xmm1, %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    maskmovdqu %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_maskmovdqu:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    maskmovdqu %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_maskmovdqu:
 ; SANDY:       # BB#0:

 ; GENERIC-LABEL: test_maxpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    maxpd %xmm1, %xmm0
 ; GENERIC-NEXT:    maxpd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_maxpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    maxpd %xmm1, %xmm0
-; ATOM-NEXT:    maxpd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    maxpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    maxpd (%rdi), %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_maxpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    maxpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    maxpd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_maxpd:

 ; GENERIC-LABEL: test_maxsd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    maxsd %xmm1, %xmm0
 ; GENERIC-NEXT:    maxsd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_maxsd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    maxsd %xmm1, %xmm0
-; ATOM-NEXT:    maxsd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    maxsd %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    maxsd (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_maxsd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    maxsd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    maxsd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_maxsd:

 ; GENERIC-LABEL: test_minpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    minpd %xmm1, %xmm0
 ; GENERIC-NEXT:    minpd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_minpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    minpd %xmm1, %xmm0
-; ATOM-NEXT:    minpd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    minpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    minpd (%rdi), %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_minpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    minpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    minpd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_minpd:

 ; GENERIC-LABEL: test_minsd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    minsd %xmm1, %xmm0
 ; GENERIC-NEXT:    minsd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_minsd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    minsd %xmm1, %xmm0
-; ATOM-NEXT:    minsd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    minsd %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    minsd (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_minsd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    minsd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    minsd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_minsd:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movapd (%rdi), %xmm0
 ; GENERIC-NEXT:    addpd %xmm0, %xmm0
 ; GENERIC-NEXT:    movapd %xmm0, (%rsi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movapd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movapd (%rdi), %xmm0
-; ATOM-NEXT:    addpd %xmm0, %xmm0
-; ATOM-NEXT:    movapd %xmm0, (%rsi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movapd (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    addpd %xmm0, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    movapd %xmm0, (%rsi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movapd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movapd (%rdi), %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    addpd %xmm0, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    movapd %xmm0, (%rsi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movdqa (%rdi), %xmm0
 ; GENERIC-NEXT:    paddq %xmm0, %xmm0
 ; GENERIC-NEXT:    movdqa %xmm0, (%rsi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movdqa:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movdqa (%rdi), %xmm0
-; ATOM-NEXT:    paddq %xmm0, %xmm0
-; ATOM-NEXT:    movdqa %xmm0, (%rsi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movdqa (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    paddq %xmm0, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    movdqa %xmm0, (%rsi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movdqa:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movdqa (%rdi), %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    paddq %xmm0, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    movdqa %xmm0, (%rsi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movdqu (%rdi), %xmm0
 ; GENERIC-NEXT:    paddq %xmm0, %xmm0
 ; GENERIC-NEXT:    movdqu %xmm0, (%rsi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movdqu:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movdqu (%rdi), %xmm0
-; ATOM-NEXT:    paddq %xmm0, %xmm0
-; ATOM-NEXT:    movdqu %xmm0, (%rsi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movdqu (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    paddq %xmm0, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    movdqu %xmm0, (%rsi) # sched: [2:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movdqu:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movdqu (%rdi), %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    paddq %xmm0, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    movdqu %xmm0, (%rsi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    paddd %xmm0, %xmm1
 ; GENERIC-NEXT:    paddd %xmm0, %xmm2
 ; GENERIC-NEXT:    movd %xmm2, %eax
 ; GENERIC-NEXT:    movd %xmm1, (%rsi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; ATOM-NEXT:    paddd %xmm0, %xmm1
-; ATOM-NEXT:    movd %xmm1, %eax
-; ATOM-NEXT:    movd %edi, %xmm1
-; ATOM-NEXT:    paddd %xmm0, %xmm1
-; ATOM-NEXT:    movd %xmm1, (%rsi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movd {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [1:1.00]
+; ATOM-NEXT:    paddd %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movd %xmm1, %eax # sched: [3:3.00]
+; ATOM-NEXT:    movd %edi, %xmm1 # sched: [1:1.00]
+; ATOM-NEXT:    paddd %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movd %xmm1, (%rsi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movd {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [3:1.00]
 ; SLM-NEXT:    movd %edi, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    paddd %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    movd %xmm1, (%rsi) # sched: [1:1.00]
 ; SLM-NEXT:    paddd %xmm0, %xmm2 # sched: [1:0.50]

 ; GENERIC-NEXT:    paddq %xmm0, %xmm1
 ; GENERIC-NEXT:    paddq %xmm0, %xmm2
 ; GENERIC-NEXT:    movq %xmm2, %rax
 ; GENERIC-NEXT:    movq %xmm1, (%rsi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movd_64:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
-; ATOM-NEXT:    movq %rdi, %xmm2
-; ATOM-NEXT:    paddq %xmm0, %xmm2
-; ATOM-NEXT:    paddq %xmm0, %xmm1
-; ATOM-NEXT:    movq %xmm2, (%rsi)
-; ATOM-NEXT:    movq %xmm1, %rax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero sched: [1:1.00]
+; ATOM-NEXT:    movq %rdi, %xmm2 # sched: [1:1.00]
+; ATOM-NEXT:    paddq %xmm0, %xmm2 # sched: [2:1.00]
+; ATOM-NEXT:    paddq %xmm0, %xmm1 # sched: [2:1.00]
+; ATOM-NEXT:    movq %xmm2, (%rsi) # sched: [1:1.00]
+; ATOM-NEXT:    movq %xmm1, %rax # sched: [3:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movd_64:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movq {{.*#+}} xmm2 = mem[0],zero sched: [3:1.00]
 ; SLM-NEXT:    movq %rdi, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    paddq %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    movq %xmm1, (%rsi) # sched: [1:1.00]
 ; SLM-NEXT:    paddq %xmm0, %xmm2 # sched: [1:0.50]

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0]
 ; GENERIC-NEXT:    addpd %xmm0, %xmm1
 ; GENERIC-NEXT:    movhpd %xmm1, (%rdi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movhpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0]
-; ATOM-NEXT:    addpd %xmm0, %xmm1
-; ATOM-NEXT:    movhpd %xmm1, (%rdi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [1:1.00]
+; ATOM-NEXT:    addpd %xmm0, %xmm1 # sched: [6:3.00]
+; ATOM-NEXT:    movhpd %xmm1, (%rdi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movhpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [4:1.00]
 ; SLM-NEXT:    addpd %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    movhpd %xmm1, (%rdi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1]
 ; GENERIC-NEXT:    addpd %xmm0, %xmm1
 ; GENERIC-NEXT:    movlpd %xmm1, (%rdi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movlpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1]
-; ATOM-NEXT:    addpd %xmm0, %xmm1
-; ATOM-NEXT:    movlpd %xmm1, (%rdi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [1:1.00]
+; ATOM-NEXT:    addpd %xmm0, %xmm1 # sched: [6:3.00]
+; ATOM-NEXT:    movlpd %xmm1, (%rdi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movlpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movlpd {{.*#+}} xmm1 = mem[0],xmm1[1] sched: [4:1.00]
 ; SLM-NEXT:    addpd %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    movlpd %xmm1, (%rdi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 define i32 @test_movmskpd(<2 x double> %a0) {
 ; GENERIC-LABEL: test_movmskpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movmskpd %xmm0, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movmskpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movmskpd %xmm0, %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movmskpd %xmm0, %eax # sched: [3:3.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movmskpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movmskpd %xmm0, %eax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_movmskpd:
 ; SANDY:       # BB#0:

 ; GENERIC-LABEL: test_movntdqa:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    paddq %xmm0, %xmm0
 ; GENERIC-NEXT:    movntdq %xmm0, (%rdi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movntdqa:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    paddq %xmm0, %xmm0
-; ATOM-NEXT:    movntdq %xmm0, (%rdi)
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    paddq %xmm0, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    movntdq %xmm0, (%rdi) # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movntdqa:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    paddq %xmm0, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    movntdq %xmm0, (%rdi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_movntdqa:

 ; GENERIC-LABEL: test_movntpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    addpd %xmm0, %xmm0
 ; GENERIC-NEXT:    movntpd %xmm0, (%rdi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movntpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    addpd %xmm0, %xmm0
-; ATOM-NEXT:    movntpd %xmm0, (%rdi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addpd %xmm0, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    movntpd %xmm0, (%rdi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movntpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    addpd %xmm0, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    movntpd %xmm0, (%rdi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_movntpd:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
 ; GENERIC-NEXT:    paddq %xmm1, %xmm0
 ; GENERIC-NEXT:    movq %xmm0, (%rdi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movq_mem:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
-; ATOM-NEXT:    paddq %xmm1, %xmm0
-; ATOM-NEXT:    movq %xmm0, (%rdi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero sched: [1:1.00]
+; ATOM-NEXT:    paddq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    movq %xmm0, (%rdi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movq_mem:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero sched: [3:1.00]
 ; SLM-NEXT:    paddq %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    movq %xmm0, (%rdi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_movq_reg:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movq {{.*#+}} xmm0 = xmm0[0],zero
 ; GENERIC-NEXT:    paddq %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movq_reg:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movq {{.*#+}} xmm0 = xmm0[0],zero
-; ATOM-NEXT:    paddq %xmm1, %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movq {{.*#+}} xmm0 = xmm0[0],zero sched: [1:0.50]
+; ATOM-NEXT:    paddq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movq_reg:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movq {{.*#+}} xmm0 = xmm0[0],zero sched: [1:0.50]
 ; SLM-NEXT:    paddq %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_movq_reg:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
 ; GENERIC-NEXT:    addsd %xmm0, %xmm0
 ; GENERIC-NEXT:    movsd %xmm0, (%rsi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movsd_mem:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
-; ATOM-NEXT:    addsd %xmm0, %xmm0
-; ATOM-NEXT:    movsd %xmm0, (%rsi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero sched: [1:1.00]
+; ATOM-NEXT:    addsd %xmm0, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    movsd %xmm0, (%rsi) # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movsd_mem:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero sched: [3:1.00]
 ; SLM-NEXT:    addsd %xmm0, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    movsd %xmm0, (%rsi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_movsd_reg:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    unpcklpd {{.*#+}} xmm1 = xmm1[0],xmm0[0]
 ; GENERIC-NEXT:    movapd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movsd_reg:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    unpcklpd {{.*#+}} xmm1 = xmm1[0],xmm0[0]
-; ATOM-NEXT:    movapd %xmm1, %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    unpcklpd {{.*#+}} xmm1 = xmm1[0],xmm0[0] sched: [1:1.00]
+; ATOM-NEXT:    movapd %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movsd_reg:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    unpcklpd {{.*#+}} xmm1 = xmm1[0],xmm0[0] sched: [1:1.00]
 ; SLM-NEXT:    movapd %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_movsd_reg:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movupd (%rdi), %xmm0
 ; GENERIC-NEXT:    addpd %xmm0, %xmm0
 ; GENERIC-NEXT:    movupd %xmm0, (%rsi)
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movupd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movupd (%rdi), %xmm0
-; ATOM-NEXT:    addpd %xmm0, %xmm0
-; ATOM-NEXT:    movupd %xmm0, (%rsi)
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movupd (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    addpd %xmm0, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    movupd %xmm0, (%rsi) # sched: [2:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movupd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movupd (%rdi), %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    addpd %xmm0, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    movupd %xmm0, (%rsi) # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_mulpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    mulpd %xmm1, %xmm0
 ; GENERIC-NEXT:    mulpd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_mulpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    mulpd %xmm1, %xmm0
-; ATOM-NEXT:    mulpd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    mulpd %xmm1, %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    mulpd (%rdi), %xmm0 # sched: [10:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_mulpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    mulpd %xmm1, %xmm0 # sched: [5:2.00]
 ; SLM-NEXT:    mulpd (%rdi), %xmm0 # sched: [8:2.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_mulpd:

 ; GENERIC-LABEL: test_mulsd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    mulsd %xmm1, %xmm0
 ; GENERIC-NEXT:    mulsd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_mulsd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    mulsd %xmm1, %xmm0
-; ATOM-NEXT:    mulsd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    mulsd %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    mulsd (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_mulsd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    mulsd %xmm1, %xmm0 # sched: [5:2.00]
 ; SLM-NEXT:    mulsd (%rdi), %xmm0 # sched: [8:2.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_mulsd:

 ; GENERIC-NEXT:    orpd (%rdi), %xmm0
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_orpd:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    orpd %xmm1, %xmm0
 ; ATOM-NEXT:    orpd (%rdi), %xmm0
-; ATOM-NEXT:    addpd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_orpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    orpd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    orpd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    packssdw %xmm1, %xmm0
 ; GENERIC-NEXT:    packssdw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_packssdw:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    packssdw %xmm1, %xmm0
 ; ATOM-NEXT:    packssdw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_packssdw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    packssdw %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    packssdw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_packssdw:

 ; GENERIC-NEXT:    packsswb %xmm1, %xmm0
 ; GENERIC-NEXT:    packsswb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_packsswb:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    packsswb %xmm1, %xmm0
 ; ATOM-NEXT:    packsswb (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_packsswb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    packsswb %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    packsswb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_packsswb:

 ; GENERIC-NEXT:    packuswb %xmm1, %xmm0
 ; GENERIC-NEXT:    packuswb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_packuswb:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    packuswb %xmm1, %xmm0
 ; ATOM-NEXT:    packuswb (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_packuswb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    packuswb %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    packuswb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_packuswb:

 ; GENERIC-LABEL: test_paddb:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    paddb %xmm1, %xmm0
 ; GENERIC-NEXT:    paddb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_paddb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    paddb %xmm1, %xmm0
-; ATOM-NEXT:    paddb (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    paddb %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    paddb (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_paddb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    paddb %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    paddb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_paddb:

 ; GENERIC-LABEL: test_paddd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    paddd %xmm1, %xmm0
 ; GENERIC-NEXT:    paddd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_paddd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    paddd %xmm1, %xmm0
-; ATOM-NEXT:    paddd (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    paddd %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    paddd (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_paddd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    paddd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    paddd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_paddd:

 ; GENERIC-LABEL: test_paddq:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    paddq %xmm1, %xmm0
 ; GENERIC-NEXT:    paddq (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_paddq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    paddq %xmm1, %xmm0
-; ATOM-NEXT:    paddq (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    paddq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    paddq (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_paddq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    paddq %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    paddq (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_paddq:

 ; GENERIC-LABEL: test_paddsb:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    paddsb %xmm1, %xmm0
 ; GENERIC-NEXT:    paddsb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_paddsb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    paddsb %xmm1, %xmm0
-; ATOM-NEXT:    paddsb (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    paddsb %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    paddsb (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_paddsb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    paddsb %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    paddsb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_paddsb:

 ; GENERIC-LABEL: test_paddsw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    paddsw %xmm1, %xmm0
 ; GENERIC-NEXT:    paddsw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_paddsw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    paddsw %xmm1, %xmm0
-; ATOM-NEXT:    paddsw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    paddsw %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    paddsw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_paddsw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    paddsw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    paddsw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_paddsw:

 ; GENERIC-LABEL: test_paddusb:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    paddusb %xmm1, %xmm0
 ; GENERIC-NEXT:    paddusb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_paddusb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    paddusb %xmm1, %xmm0
-; ATOM-NEXT:    paddusb (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    paddusb %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    paddusb (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_paddusb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    paddusb %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    paddusb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_paddusb:

 ; GENERIC-LABEL: test_paddusw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    paddusw %xmm1, %xmm0
 ; GENERIC-NEXT:    paddusw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_paddusw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    paddusw %xmm1, %xmm0
-; ATOM-NEXT:    paddusw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    paddusw %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    paddusw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_paddusw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    paddusw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    paddusw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_paddusw:

 ; GENERIC-LABEL: test_paddw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    paddw %xmm1, %xmm0
 ; GENERIC-NEXT:    paddw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_paddw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    paddw %xmm1, %xmm0
-; ATOM-NEXT:    paddw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    paddw %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    paddw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_paddw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    paddw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    paddw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_paddw:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pand %xmm1, %xmm0
 ; GENERIC-NEXT:    pand (%rdi), %xmm0
 ; GENERIC-NEXT:    paddq %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pand:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pand %xmm1, %xmm0
-; ATOM-NEXT:    pand (%rdi), %xmm0
-; ATOM-NEXT:    paddq %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pand %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    pand (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    paddq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pand:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pand %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    pand (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    paddq %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    movdqa %xmm0, %xmm1
 ; GENERIC-NEXT:    pandn (%rdi), %xmm1
 ; GENERIC-NEXT:    paddq %xmm0, %xmm1
 ; GENERIC-NEXT:    movdqa %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pandn:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pandn %xmm1, %xmm0
-; ATOM-NEXT:    movdqa %xmm0, %xmm1
-; ATOM-NEXT:    pandn (%rdi), %xmm1
-; ATOM-NEXT:    paddq %xmm0, %xmm1
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pandn %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    movdqa %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    pandn (%rdi), %xmm1 # sched: [1:1.00]
+; ATOM-NEXT:    paddq %xmm0, %xmm1 # sched: [2:1.00]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pandn:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pandn %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    movdqa %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    pandn (%rdi), %xmm1 # sched: [4:1.00]
 ; SLM-NEXT:    paddq %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]

 ; GENERIC-LABEL: test_pavgb:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pavgb %xmm1, %xmm0
 ; GENERIC-NEXT:    pavgb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pavgb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pavgb %xmm1, %xmm0
-; ATOM-NEXT:    pavgb (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pavgb %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    pavgb (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pavgb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pavgb %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    pavgb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pavgb:

 ; GENERIC-LABEL: test_pavgw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pavgw %xmm1, %xmm0
 ; GENERIC-NEXT:    pavgw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pavgw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pavgw %xmm1, %xmm0
-; ATOM-NEXT:    pavgw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pavgw %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    pavgw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pavgw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pavgw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    pavgw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pavgw:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pcmpeqb %xmm0, %xmm1
 ; GENERIC-NEXT:    pcmpeqb (%rdi), %xmm0
 ; GENERIC-NEXT:    por %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pcmpeqb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pcmpeqb %xmm0, %xmm1
-; ATOM-NEXT:    pcmpeqb (%rdi), %xmm0
-; ATOM-NEXT:    por %xmm1, %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pcmpeqb %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    pcmpeqb (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    por %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pcmpeqb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pcmpeqb %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    pcmpeqb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    por %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pcmpeqd %xmm0, %xmm1
 ; GENERIC-NEXT:    pcmpeqd (%rdi), %xmm0
 ; GENERIC-NEXT:    por %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pcmpeqd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pcmpeqd %xmm0, %xmm1
-; ATOM-NEXT:    pcmpeqd (%rdi), %xmm0
-; ATOM-NEXT:    por %xmm1, %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pcmpeqd %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    pcmpeqd (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    por %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pcmpeqd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pcmpeqd %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    pcmpeqd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    por %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pcmpeqw %xmm0, %xmm1
 ; GENERIC-NEXT:    pcmpeqw (%rdi), %xmm0
 ; GENERIC-NEXT:    por %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pcmpeqw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pcmpeqw %xmm0, %xmm1
-; ATOM-NEXT:    pcmpeqw (%rdi), %xmm0
-; ATOM-NEXT:    por %xmm1, %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pcmpeqw %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    pcmpeqw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    por %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pcmpeqw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pcmpeqw %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    pcmpeqw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    por %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    movdqa %xmm0, %xmm2
 ; GENERIC-NEXT:    pcmpgtb %xmm1, %xmm2
 ; GENERIC-NEXT:    pcmpgtb (%rdi), %xmm0
 ; GENERIC-NEXT:    por %xmm2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pcmpgtb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movdqa %xmm0, %xmm2
-; ATOM-NEXT:    pcmpgtb (%rdi), %xmm0
-; ATOM-NEXT:    pcmpgtb %xmm1, %xmm2
-; ATOM-NEXT:    por %xmm2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movdqa %xmm0, %xmm2 # sched: [1:0.50]
+; ATOM-NEXT:    pcmpgtb (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    pcmpgtb %xmm1, %xmm2 # sched: [1:0.50]
+; ATOM-NEXT:    por %xmm2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pcmpgtb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movdqa %xmm0, %xmm2 # sched: [1:0.50]
 ; SLM-NEXT:    pcmpgtb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    pcmpgtb %xmm1, %xmm2 # sched: [1:0.50]
 ; SLM-NEXT:    por %xmm2, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC-NEXT:    movdqa %xmm0, %xmm2
 ; GENERIC-NEXT:    pcmpgtd %xmm1, %xmm2
 ; GENERIC-NEXT:    pcmpeqd (%rdi), %xmm0
 ; GENERIC-NEXT:    por %xmm2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pcmpgtd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movdqa %xmm0, %xmm2
-; ATOM-NEXT:    pcmpeqd (%rdi), %xmm0
-; ATOM-NEXT:    pcmpgtd %xmm1, %xmm2
-; ATOM-NEXT:    por %xmm2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movdqa %xmm0, %xmm2 # sched: [1:0.50]
+; ATOM-NEXT:    pcmpeqd (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    pcmpgtd %xmm1, %xmm2 # sched: [1:0.50]
+; ATOM-NEXT:    por %xmm2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pcmpgtd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movdqa %xmm0, %xmm2 # sched: [1:0.50]
 ; SLM-NEXT:    pcmpeqd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    pcmpgtd %xmm1, %xmm2 # sched: [1:0.50]
 ; SLM-NEXT:    por %xmm2, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC-NEXT:    movdqa %xmm0, %xmm2
 ; GENERIC-NEXT:    pcmpgtw %xmm1, %xmm2
 ; GENERIC-NEXT:    pcmpgtw (%rdi), %xmm0
 ; GENERIC-NEXT:    por %xmm2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pcmpgtw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movdqa %xmm0, %xmm2
-; ATOM-NEXT:    pcmpgtw (%rdi), %xmm0
-; ATOM-NEXT:    pcmpgtw %xmm1, %xmm2
-; ATOM-NEXT:    por %xmm2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movdqa %xmm0, %xmm2 # sched: [1:0.50]
+; ATOM-NEXT:    pcmpgtw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    pcmpgtw %xmm1, %xmm2 # sched: [1:0.50]
+; ATOM-NEXT:    por %xmm2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pcmpgtw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movdqa %xmm0, %xmm2 # sched: [1:0.50]
 ; SLM-NEXT:    pcmpgtw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    pcmpgtw %xmm1, %xmm2 # sched: [1:0.50]
 ; SLM-NEXT:    por %xmm2, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC-LABEL: test_pextrw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pextrw $6, %xmm0, %eax
 ; GENERIC-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pextrw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pextrw $6, %xmm0, %eax
+; ATOM-NEXT:    pextrw $6, %xmm0, %eax # sched: [4:2.00]
 ; ATOM-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pextrw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pextrw $6, %xmm0, %eax # sched: [4:1.00]
 ; SLM-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pextrw:

 ; GENERIC-LABEL: test_pinsrw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pinsrw $1, %edi, %xmm0
 ; GENERIC-NEXT:    pinsrw $3, (%rsi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pinsrw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pinsrw $1, %edi, %xmm0
-; ATOM-NEXT:    pinsrw $3, (%rsi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pinsrw $1, %edi, %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    pinsrw $3, (%rsi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pinsrw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pinsrw $1, %edi, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    pinsrw $3, (%rsi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pinsrw:

 ; GENERIC-NEXT:    pmaddwd %xmm1, %xmm0
 ; GENERIC-NEXT:    pmaddwd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pmaddwd:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    pmaddwd %xmm1, %xmm0
 ; ATOM-NEXT:    pmaddwd (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pmaddwd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pmaddwd %xmm1, %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    pmaddwd (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pmaddwd:

 ; GENERIC-LABEL: test_pmaxsw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pmaxsw %xmm1, %xmm0
 ; GENERIC-NEXT:    pmaxsw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pmaxsw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pmaxsw %xmm1, %xmm0
-; ATOM-NEXT:    pmaxsw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pmaxsw %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    pmaxsw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pmaxsw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pmaxsw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    pmaxsw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pmaxsw:

 ; GENERIC-LABEL: test_pmaxub:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pmaxub %xmm1, %xmm0
 ; GENERIC-NEXT:    pmaxub (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pmaxub:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pmaxub %xmm1, %xmm0
-; ATOM-NEXT:    pmaxub (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pmaxub %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    pmaxub (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pmaxub:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pmaxub %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    pmaxub (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pmaxub:

 ; GENERIC-LABEL: test_pminsw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pminsw %xmm1, %xmm0
 ; GENERIC-NEXT:    pminsw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pminsw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pminsw %xmm1, %xmm0
-; ATOM-NEXT:    pminsw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pminsw %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    pminsw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pminsw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pminsw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    pminsw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pminsw:

 ; GENERIC-LABEL: test_pminub:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pminub %xmm1, %xmm0
 ; GENERIC-NEXT:    pminub (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pminub:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pminub %xmm1, %xmm0
-; ATOM-NEXT:    pminub (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pminub %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    pminub (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pminub:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pminub %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    pminub (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pminub:

 define i32 @test_pmovmskb(<16 x i8> %a0) {
 ; GENERIC-LABEL: test_pmovmskb:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pmovmskb %xmm0, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pmovmskb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pmovmskb %xmm0, %eax
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pmovmskb %xmm0, %eax # sched: [3:3.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pmovmskb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pmovmskb %xmm0, %eax # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pmovmskb:
 ; SANDY:       # BB#0:

 ; GENERIC-LABEL: test_pmulhuw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pmulhuw %xmm1, %xmm0
 ; GENERIC-NEXT:    pmulhuw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pmulhuw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pmulhuw %xmm1, %xmm0
-; ATOM-NEXT:    pmulhuw (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pmulhuw %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    pmulhuw (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pmulhuw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pmulhuw %xmm1, %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    pmulhuw (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pmulhuw:

 ; GENERIC-LABEL: test_pmulhw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pmulhw %xmm1, %xmm0
 ; GENERIC-NEXT:    pmulhw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pmulhw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pmulhw %xmm1, %xmm0
-; ATOM-NEXT:    pmulhw (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pmulhw %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    pmulhw (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pmulhw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pmulhw %xmm1, %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    pmulhw (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pmulhw:

 ; GENERIC-LABEL: test_pmullw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pmullw %xmm1, %xmm0
 ; GENERIC-NEXT:    pmullw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pmullw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pmullw %xmm1, %xmm0
-; ATOM-NEXT:    pmullw (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pmullw %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    pmullw (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pmullw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pmullw %xmm1, %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    pmullw (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pmullw:

 ; GENERIC-NEXT:    pmuludq %xmm1, %xmm0
 ; GENERIC-NEXT:    pmuludq (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pmuludq:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    pmuludq %xmm1, %xmm0
 ; ATOM-NEXT:    pmuludq (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pmuludq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pmuludq %xmm1, %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    pmuludq (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pmuludq:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    por %xmm1, %xmm0
 ; GENERIC-NEXT:    por (%rdi), %xmm0
 ; GENERIC-NEXT:    paddq %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_por:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    por %xmm1, %xmm0
-; ATOM-NEXT:    por (%rdi), %xmm0
-; ATOM-NEXT:    paddq %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    por %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    por (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    paddq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_por:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    por %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    por (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    paddq %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    psadbw %xmm1, %xmm0
 ; GENERIC-NEXT:    psadbw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psadbw:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    psadbw %xmm1, %xmm0
 ; ATOM-NEXT:    psadbw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psadbw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psadbw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psadbw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psadbw:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
 ; GENERIC-NEXT:    pshufd {{.*#+}} xmm0 = mem[3,2,1,0]
 ; GENERIC-NEXT:    paddd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pshufd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pshufd {{.*#+}} xmm1 = mem[3,2,1,0]
-; ATOM-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,0,3,2]
-; ATOM-NEXT:    paddd %xmm0, %xmm1
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pshufd {{.*#+}} xmm1 = mem[3,2,1,0] sched: [1:1.00]
+; ATOM-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,0,3,2] sched: [1:1.00]
+; ATOM-NEXT:    paddd %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pshufd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pshufd {{.*#+}} xmm1 = mem[3,2,1,0] sched: [4:1.00]
 ; SLM-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,0,3,2] sched: [1:1.00]
 ; SLM-NEXT:    paddd %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,5,4,7,6]
 ; GENERIC-NEXT:    pshufhw {{.*#+}} xmm0 = mem[0,1,2,3,7,6,5,4]
 ; GENERIC-NEXT:    paddw %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pshufhw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pshufhw {{.*#+}} xmm1 = mem[0,1,2,3,7,6,5,4]
-; ATOM-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,4,7,6]
-; ATOM-NEXT:    paddw %xmm0, %xmm1
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pshufhw {{.*#+}} xmm1 = mem[0,1,2,3,7,6,5,4] sched: [1:1.00]
+; ATOM-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,4,7,6] sched: [1:1.00]
+; ATOM-NEXT:    paddw %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pshufhw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pshufhw {{.*#+}} xmm1 = mem[0,1,2,3,7,6,5,4] sched: [4:1.00]
 ; SLM-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,4,7,6] sched: [1:1.00]
 ; SLM-NEXT:    paddw %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pshuflw {{.*#+}} xmm1 = xmm0[1,0,3,2,4,5,6,7]
 ; GENERIC-NEXT:    pshuflw {{.*#+}} xmm0 = mem[3,2,1,0,4,5,6,7]
 ; GENERIC-NEXT:    paddw %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pshuflw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pshuflw {{.*#+}} xmm1 = mem[3,2,1,0,4,5,6,7]
-; ATOM-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7]
-; ATOM-NEXT:    paddw %xmm0, %xmm1
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pshuflw {{.*#+}} xmm1 = mem[3,2,1,0,4,5,6,7] sched: [1:1.00]
+; ATOM-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7] sched: [1:1.00]
+; ATOM-NEXT:    paddw %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pshuflw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pshuflw {{.*#+}} xmm1 = mem[3,2,1,0,4,5,6,7] sched: [4:1.00]
 ; SLM-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7] sched: [1:1.00]
 ; SLM-NEXT:    paddw %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pslld %xmm1, %xmm0
 ; GENERIC-NEXT:    pslld (%rdi), %xmm0
 ; GENERIC-NEXT:    pslld $2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pslld:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pslld %xmm1, %xmm0
-; ATOM-NEXT:    pslld (%rdi), %xmm0
-; ATOM-NEXT:    pslld $2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pslld %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    pslld (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    pslld $2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pslld:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pslld %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    pslld (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    pslld $2, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 define <4 x i32> @test_pslldq(<4 x i32> %a0) {
 ; GENERIC-LABEL: test_pslldq:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pslldq {{.*#+}} xmm0 = zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7,8,9,10,11]
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pslldq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pslldq {{.*#+}} xmm0 = zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7,8,9,10,11]
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pslldq {{.*#+}} xmm0 = zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7,8,9,10,11] sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pslldq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pslldq {{.*#+}} xmm0 = zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7,8,9,10,11] sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pslldq:
 ; SANDY:       # BB#0:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psllq %xmm1, %xmm0
 ; GENERIC-NEXT:    psllq (%rdi), %xmm0
 ; GENERIC-NEXT:    psllq $2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psllq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psllq %xmm1, %xmm0
-; ATOM-NEXT:    psllq (%rdi), %xmm0
-; ATOM-NEXT:    psllq $2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psllq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    psllq (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    psllq $2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psllq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psllq %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    psllq (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    psllq $2, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psllw %xmm1, %xmm0
 ; GENERIC-NEXT:    psllw (%rdi), %xmm0
 ; GENERIC-NEXT:    psllw $2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psllw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psllw %xmm1, %xmm0
-; ATOM-NEXT:    psllw (%rdi), %xmm0
-; ATOM-NEXT:    psllw $2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psllw %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    psllw (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    psllw $2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psllw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psllw %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    psllw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    psllw $2, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psrad %xmm1, %xmm0
 ; GENERIC-NEXT:    psrad (%rdi), %xmm0
 ; GENERIC-NEXT:    psrad $2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psrad:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psrad %xmm1, %xmm0
-; ATOM-NEXT:    psrad (%rdi), %xmm0
-; ATOM-NEXT:    psrad $2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psrad %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    psrad (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    psrad $2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psrad:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psrad %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    psrad (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    psrad $2, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psraw %xmm1, %xmm0
 ; GENERIC-NEXT:    psraw (%rdi), %xmm0
 ; GENERIC-NEXT:    psraw $2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psraw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psraw %xmm1, %xmm0
-; ATOM-NEXT:    psraw (%rdi), %xmm0
-; ATOM-NEXT:    psraw $2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psraw %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    psraw (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    psraw $2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psraw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psraw %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    psraw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    psraw $2, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psrld %xmm1, %xmm0
 ; GENERIC-NEXT:    psrld (%rdi), %xmm0
 ; GENERIC-NEXT:    psrld $2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psrld:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psrld %xmm1, %xmm0
-; ATOM-NEXT:    psrld (%rdi), %xmm0
-; ATOM-NEXT:    psrld $2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psrld %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    psrld (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    psrld $2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psrld:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psrld %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    psrld (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    psrld $2, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 define <4 x i32> @test_psrldq(<4 x i32> %a0) {
 ; GENERIC-LABEL: test_psrldq:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[4,5,6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psrldq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[4,5,6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[4,5,6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psrldq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[4,5,6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psrldq:
 ; SANDY:       # BB#0:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psrlq %xmm1, %xmm0
 ; GENERIC-NEXT:    psrlq (%rdi), %xmm0
 ; GENERIC-NEXT:    psrlq $2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psrlq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psrlq %xmm1, %xmm0
-; ATOM-NEXT:    psrlq (%rdi), %xmm0
-; ATOM-NEXT:    psrlq $2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psrlq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    psrlq (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    psrlq $2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psrlq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psrlq %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    psrlq (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    psrlq $2, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psrlw %xmm1, %xmm0
 ; GENERIC-NEXT:    psrlw (%rdi), %xmm0
 ; GENERIC-NEXT:    psrlw $2, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psrlw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psrlw %xmm1, %xmm0
-; ATOM-NEXT:    psrlw (%rdi), %xmm0
-; ATOM-NEXT:    psrlw $2, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psrlw %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    psrlw (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    psrlw $2, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psrlw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psrlw %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    psrlw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    psrlw $2, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_psubb:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psubb %xmm1, %xmm0
 ; GENERIC-NEXT:    psubb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psubb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psubb %xmm1, %xmm0
-; ATOM-NEXT:    psubb (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psubb %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    psubb (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psubb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psubb %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psubb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psubb:

 ; GENERIC-LABEL: test_psubd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psubd %xmm1, %xmm0
 ; GENERIC-NEXT:    psubd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psubd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psubd %xmm1, %xmm0
-; ATOM-NEXT:    psubd (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psubd %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    psubd (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psubd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psubd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psubd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psubd:

 ; GENERIC-LABEL: test_psubq:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psubq %xmm1, %xmm0
 ; GENERIC-NEXT:    psubq (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psubq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psubq %xmm1, %xmm0
-; ATOM-NEXT:    psubq (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psubq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    psubq (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psubq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psubq %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psubq (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psubq:

 ; GENERIC-LABEL: test_psubsb:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psubsb %xmm1, %xmm0
 ; GENERIC-NEXT:    psubsb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psubsb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psubsb %xmm1, %xmm0
-; ATOM-NEXT:    psubsb (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psubsb %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    psubsb (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psubsb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psubsb %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psubsb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psubsb:

 ; GENERIC-LABEL: test_psubsw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psubsw %xmm1, %xmm0
 ; GENERIC-NEXT:    psubsw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psubsw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psubsw %xmm1, %xmm0
-; ATOM-NEXT:    psubsw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psubsw %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    psubsw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psubsw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psubsw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psubsw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psubsw:

 ; GENERIC-LABEL: test_psubusb:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psubusb %xmm1, %xmm0
 ; GENERIC-NEXT:    psubusb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psubusb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psubusb %xmm1, %xmm0
-; ATOM-NEXT:    psubusb (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psubusb %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    psubusb (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psubusb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psubusb %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psubusb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psubusb:

 ; GENERIC-LABEL: test_psubusw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psubusw %xmm1, %xmm0
 ; GENERIC-NEXT:    psubusw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psubusw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psubusw %xmm1, %xmm0
-; ATOM-NEXT:    psubusw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psubusw %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    psubusw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psubusw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psubusw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psubusw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psubusw:

 ; GENERIC-LABEL: test_psubw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    psubw %xmm1, %xmm0
 ; GENERIC-NEXT:    psubw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psubw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    psubw %xmm1, %xmm0
-; ATOM-NEXT:    psubw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    psubw %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    psubw (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psubw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psubw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psubw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psubw:

 ; GENERIC-LABEL: test_punpckhbw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
 ; GENERIC-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15]
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_punpckhbw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15]
-; ATOM-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15]
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15] sched: [1:1.00]
+; ATOM-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15] sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_punpckhbw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15] sched: [1:1.00]
 ; SLM-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15] sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_punpckhbw:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
 ; GENERIC-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]
 ; GENERIC-NEXT:    paddd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_punpckhdq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; ATOM-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]
-; ATOM-NEXT:    paddd %xmm1, %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00]
+; ATOM-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3] sched: [1:1.00]
+; ATOM-NEXT:    paddd %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_punpckhdq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00]
 ; SLM-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],mem[2],xmm1[3],mem[3] sched: [4:1.00]
 ; SLM-NEXT:    paddd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
 ; GENERIC-NEXT:    punpckhqdq {{.*#+}} xmm1 = xmm1[1],mem[1]
 ; GENERIC-NEXT:    paddq %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_punpckhqdq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
-; ATOM-NEXT:    punpckhqdq {{.*#+}} xmm1 = xmm1[1],mem[1]
-; ATOM-NEXT:    paddq %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1] sched: [1:1.00]
+; ATOM-NEXT:    punpckhqdq {{.*#+}} xmm1 = xmm1[1],mem[1] sched: [1:1.00]
+; ATOM-NEXT:    paddq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_punpckhqdq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1] sched: [1:1.00]
 ; SLM-NEXT:    punpckhqdq {{.*#+}} xmm1 = xmm1[1],mem[1] sched: [4:1.00]
 ; SLM-NEXT:    paddq %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_punpckhwd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
 ; GENERIC-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_punpckhwd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; ATOM-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] sched: [1:1.00]
+; ATOM-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7] sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_punpckhwd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] sched: [1:1.00]
 ; SLM-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7] sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_punpckhwd:

 ; GENERIC-LABEL: test_punpcklbw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
 ; GENERIC-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_punpcklbw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
-; ATOM-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] sched: [1:1.00]
+; ATOM-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7] sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_punpcklbw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] sched: [1:1.00]
 ; SLM-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7] sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_punpcklbw:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; GENERIC-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
 ; GENERIC-NEXT:    paddd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_punpckldq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; ATOM-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; ATOM-NEXT:    paddd %xmm1, %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00]
+; ATOM-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] sched: [1:1.00]
+; ATOM-NEXT:    paddd %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_punpckldq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] sched: [1:1.00]
 ; SLM-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] sched: [4:1.00]
 ; SLM-NEXT:    paddd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; GENERIC-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0]
 ; GENERIC-NEXT:    paddq %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_punpcklqdq:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; ATOM-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0]
-; ATOM-NEXT:    paddq %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]
+; ATOM-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [1:1.00]
+; ATOM-NEXT:    paddq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_punpcklqdq:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]
 ; SLM-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [4:1.00]
 ; SLM-NEXT:    paddq %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_punpcklwd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
 ; GENERIC-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_punpcklwd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; ATOM-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00]
+; ATOM-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] sched: [1:1.00]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_punpcklwd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] sched: [1:1.00]
 ; SLM-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_punpcklwd:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pxor %xmm1, %xmm0
 ; GENERIC-NEXT:    pxor (%rdi), %xmm0
 ; GENERIC-NEXT:    paddq %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pxor:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pxor %xmm1, %xmm0
-; ATOM-NEXT:    pxor (%rdi), %xmm0
-; ATOM-NEXT:    paddq %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pxor %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    pxor (%rdi), %xmm0 # sched: [1:1.00]
+; ATOM-NEXT:    paddq %xmm1, %xmm0 # sched: [2:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pxor:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pxor %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    pxor (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    paddq %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    shufpd {{.*#+}} xmm0 = xmm0[1],xmm1[0]
 ; GENERIC-NEXT:    shufpd {{.*#+}} xmm1 = xmm1[1],mem[0]
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_shufpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    shufpd {{.*#+}} xmm0 = xmm0[1],xmm1[0]
-; ATOM-NEXT:    shufpd {{.*#+}} xmm1 = xmm1[1],mem[0]
-; ATOM-NEXT:    addpd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    shufpd {{.*#+}} xmm0 = xmm0[1],xmm1[0] sched: [1:1.00]
+; ATOM-NEXT:    shufpd {{.*#+}} xmm1 = xmm1[1],mem[0] sched: [1:1.00]
+; ATOM-NEXT:    addpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_shufpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    shufpd {{.*#+}} xmm0 = xmm0[1],xmm1[0] sched: [1:1.00]
 ; SLM-NEXT:    shufpd {{.*#+}} xmm1 = xmm1[1],mem[0] sched: [4:1.00]
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    sqrtpd %xmm0, %xmm1
 ; GENERIC-NEXT:    sqrtpd (%rdi), %xmm0
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_sqrtpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    sqrtpd %xmm0, %xmm1
-; ATOM-NEXT:    sqrtpd (%rdi), %xmm0
-; ATOM-NEXT:    addpd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    sqrtpd %xmm0, %xmm1 # sched: [125:62.50]
+; ATOM-NEXT:    sqrtpd (%rdi), %xmm0 # sched: [125:62.50]
+; ATOM-NEXT:    addpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_sqrtpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    sqrtpd (%rdi), %xmm1 # sched: [18:1.00]
 ; SLM-NEXT:    sqrtpd %xmm0, %xmm0 # sched: [15:1.00]
 ; SLM-NEXT:    addpd %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    movapd %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC-NEXT:    sqrtsd %xmm0, %xmm0
 ; GENERIC-NEXT:    movapd (%rdi), %xmm1
 ; GENERIC-NEXT:    sqrtsd %xmm1, %xmm1
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_sqrtsd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movapd (%rdi), %xmm1
+; ATOM-NEXT:    movapd (%rdi), %xmm1 # sched: [1:1.00]
 ; ATOM-NEXT:    sqrtsd %xmm0, %xmm0
 ; ATOM-NEXT:    sqrtsd %xmm1, %xmm1
-; ATOM-NEXT:    addpd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_sqrtsd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movapd (%rdi), %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    sqrtsd %xmm0, %xmm0 # sched: [18:1.00]
 ; SLM-NEXT:    sqrtsd %xmm1, %xmm1 # sched: [18:1.00]
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]

 ; GENERIC-LABEL: test_subpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    subpd %xmm1, %xmm0
 ; GENERIC-NEXT:    subpd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_subpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    subpd %xmm1, %xmm0
-; ATOM-NEXT:    subpd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    subpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    subpd (%rdi), %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_subpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    subpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    subpd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_subpd:

 ; GENERIC-LABEL: test_subsd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    subsd %xmm1, %xmm0
 ; GENERIC-NEXT:    subsd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_subsd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    subsd %xmm1, %xmm0
-; ATOM-NEXT:    subsd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    subsd %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    subsd (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_subsd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    subsd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    subsd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_subsd:

 ; GENERIC-NEXT:    sete %dl
 ; GENERIC-NEXT:    andb %al, %dl
 ; GENERIC-NEXT:    orb %cl, %dl
 ; GENERIC-NEXT:    movzbl %dl, %eax
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_ucomisd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    ucomisd %xmm1, %xmm0
-; ATOM-NEXT:    setnp %al
-; ATOM-NEXT:    sete %cl
-; ATOM-NEXT:    andb %al, %cl
-; ATOM-NEXT:    ucomisd (%rdi), %xmm0
-; ATOM-NEXT:    setnp %al
-; ATOM-NEXT:    sete %dl
-; ATOM-NEXT:    andb %al, %dl
-; ATOM-NEXT:    orb %cl, %dl
-; ATOM-NEXT:    movzbl %dl, %eax
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    ucomisd %xmm1, %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    setnp %al # sched: [1:0.50]
+; ATOM-NEXT:    sete %cl # sched: [1:0.50]
+; ATOM-NEXT:    andb %al, %cl # sched: [1:0.50]
+; ATOM-NEXT:    ucomisd (%rdi), %xmm0 # sched: [10:5.00]
+; ATOM-NEXT:    setnp %al # sched: [1:0.50]
+; ATOM-NEXT:    sete %dl # sched: [1:0.50]
+; ATOM-NEXT:    andb %al, %dl # sched: [1:0.50]
+; ATOM-NEXT:    orb %cl, %dl # sched: [1:0.50]
+; ATOM-NEXT:    movzbl %dl, %eax # sched: [1:1.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_ucomisd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    ucomisd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    setnp %al # sched: [1:0.50]
 ; SLM-NEXT:    sete %cl # sched: [1:0.50]
 ; SLM-NEXT:    andb %al, %cl # sched: [1:0.50]
 ; SLM-NEXT:    ucomisd (%rdi), %xmm0 # sched: [6:1.00]

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
 ; GENERIC-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],mem[1]
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_unpckhpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
-; ATOM-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],mem[1]
-; ATOM-NEXT:    addpd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1] sched: [1:1.00]
+; ATOM-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],mem[1] sched: [1:1.00]
+; ATOM-NEXT:    addpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_unpckhpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1] sched: [1:1.00]
 ; SLM-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],mem[1] sched: [4:1.00]
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-NEXT:    movapd %xmm0, %xmm1
 ; GENERIC-NEXT:    unpcklpd {{.*#+}} xmm1 = xmm1[0],mem[0]
 ; GENERIC-NEXT:    addpd %xmm0, %xmm1
 ; GENERIC-NEXT:    movapd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_unpcklpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; ATOM-NEXT:    movapd %xmm0, %xmm1
-; ATOM-NEXT:    unpcklpd {{.*#+}} xmm1 = xmm1[0],mem[0]
-; ATOM-NEXT:    addpd %xmm0, %xmm1
-; ATOM-NEXT:    movapd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]
+; ATOM-NEXT:    movapd %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    unpcklpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [1:1.00]
+; ATOM-NEXT:    addpd %xmm0, %xmm1 # sched: [6:3.00]
+; ATOM-NEXT:    movapd %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_unpcklpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0] sched: [1:1.00]
 ; SLM-NEXT:    movapd %xmm0, %xmm1 # sched: [1:1.00]
 ; SLM-NEXT:    unpcklpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [4:1.00]
 ; SLM-NEXT:    addpd %xmm0, %xmm1 # sched: [3:1.00]
 ; SLM-NEXT:    movapd %xmm1, %xmm0 # sched: [1:1.00]

 ; GENERIC-NEXT:    xorpd (%rdi), %xmm0
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_xorpd:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    xorpd %xmm1, %xmm0
 ; ATOM-NEXT:    xorpd (%rdi), %xmm0
-; ATOM-NEXT:    addpd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_xorpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    xorpd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    xorpd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

test/CodeGen/X86/sse3-schedule.ll

 ; GENERIC-LABEL: test_addsubpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    addsubpd %xmm1, %xmm0
 ; GENERIC-NEXT:    addsubpd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_addsubpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    addsubpd %xmm1, %xmm0
-; ATOM-NEXT:    addsubpd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addsubpd %xmm1, %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    addsubpd (%rdi), %xmm0 # sched: [6:3.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_addsubpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    addsubpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    addsubpd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_addsubpd:

 ; GENERIC-LABEL: test_addsubps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    addsubps %xmm1, %xmm0
 ; GENERIC-NEXT:    addsubps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_addsubps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    addsubps %xmm1, %xmm0
-; ATOM-NEXT:    addsubps (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    addsubps %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    addsubps (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_addsubps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    addsubps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    addsubps (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_addsubps:

 ; GENERIC-LABEL: test_haddpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    haddpd %xmm1, %xmm0
 ; GENERIC-NEXT:    haddpd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_haddpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    haddpd %xmm1, %xmm0
-; ATOM-NEXT:    haddpd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    haddpd %xmm1, %xmm0 # sched: [8:4.00]
+; ATOM-NEXT:    haddpd (%rdi), %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_haddpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    haddpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    haddpd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_haddpd:

 ; GENERIC-LABEL: test_haddps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    haddps %xmm1, %xmm0
 ; GENERIC-NEXT:    haddps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_haddps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    haddps %xmm1, %xmm0
-; ATOM-NEXT:    haddps (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    haddps %xmm1, %xmm0 # sched: [8:4.00]
+; ATOM-NEXT:    haddps (%rdi), %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_haddps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    haddps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    haddps (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_haddps:

 ; GENERIC-LABEL: test_hsubpd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    hsubpd %xmm1, %xmm0
 ; GENERIC-NEXT:    hsubpd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_hsubpd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    hsubpd %xmm1, %xmm0
-; ATOM-NEXT:    hsubpd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    hsubpd %xmm1, %xmm0 # sched: [8:4.00]
+; ATOM-NEXT:    hsubpd (%rdi), %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_hsubpd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    hsubpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    hsubpd (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_hsubpd:

 ; GENERIC-LABEL: test_hsubps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    hsubps %xmm1, %xmm0
 ; GENERIC-NEXT:    hsubps (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_hsubps:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    hsubps %xmm1, %xmm0
-; ATOM-NEXT:    hsubps (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    hsubps %xmm1, %xmm0 # sched: [8:4.00]
+; ATOM-NEXT:    hsubps (%rdi), %xmm0 # sched: [9:4.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_hsubps:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    hsubps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    hsubps (%rdi), %xmm0 # sched: [6:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_hsubps:

 define <16 x i8> @test_lddqu(i8* %a0) {
 ; GENERIC-LABEL: test_lddqu:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    lddqu (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_lddqu:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    lddqu (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    lddqu (%rdi), %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_lddqu:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    lddqu (%rdi), %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_lddqu:
 ; SANDY:       # BB#0:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movddup {{.*#+}} xmm1 = xmm0[0,0]
 ; GENERIC-NEXT:    movddup {{.*#+}} xmm0 = mem[0,0]
 ; GENERIC-NEXT:    addpd %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movddup:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movddup {{.*#+}} xmm1 = mem[0,0]
-; ATOM-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
-; ATOM-NEXT:    addpd %xmm0, %xmm1
-; ATOM-NEXT:    movapd %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movddup {{.*#+}} xmm1 = mem[0,0] sched: [1:1.00]
+; ATOM-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0] sched: [1:1.00]
+; ATOM-NEXT:    addpd %xmm0, %xmm1 # sched: [6:3.00]
+; ATOM-NEXT:    movapd %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movddup:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movddup {{.*#+}} xmm1 = xmm0[0,0] sched: [1:1.00]
 ; SLM-NEXT:    movddup {{.*#+}} xmm0 = mem[0,0] sched: [3:1.00]
 ; SLM-NEXT:    addpd %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
 ; GENERIC-NEXT:    movshdup {{.*#+}} xmm0 = mem[1,1,3,3]
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movshdup:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movshdup {{.*#+}} xmm1 = mem[1,1,3,3]
-; ATOM-NEXT:    movshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; ATOM-NEXT:    addps %xmm0, %xmm1
-; ATOM-NEXT:    movaps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movshdup {{.*#+}} xmm1 = mem[1,1,3,3] sched: [1:1.00]
+; ATOM-NEXT:    movshdup {{.*#+}} xmm0 = xmm0[1,1,3,3] sched: [1:1.00]
+; ATOM-NEXT:    addps %xmm0, %xmm1 # sched: [5:5.00]
+; ATOM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movshdup:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3] sched: [1:1.00]
 ; SLM-NEXT:    movshdup {{.*#+}} xmm0 = mem[1,1,3,3] sched: [3:1.00]
 ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    movsldup {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; GENERIC-NEXT:    movsldup {{.*#+}} xmm0 = mem[0,0,2,2]
 ; GENERIC-NEXT:    addps %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_movsldup:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    movsldup {{.*#+}} xmm1 = mem[0,0,2,2]
-; ATOM-NEXT:    movsldup {{.*#+}} xmm0 = xmm0[0,0,2,2]
-; ATOM-NEXT:    addps %xmm0, %xmm1
-; ATOM-NEXT:    movaps %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    movsldup {{.*#+}} xmm1 = mem[0,0,2,2] sched: [1:1.00]
+; ATOM-NEXT:    movsldup {{.*#+}} xmm0 = xmm0[0,0,2,2] sched: [1:1.00]
+; ATOM-NEXT:    addps %xmm0, %xmm1 # sched: [5:5.00]
+; ATOM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movsldup:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    movsldup {{.*#+}} xmm1 = xmm0[0,0,2,2] sched: [1:1.00]
 ; SLM-NEXT:    movsldup {{.*#+}} xmm0 = mem[0,0,2,2] sched: [3:1.00]
 ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

test/CodeGen/X86/ssse3-schedule.ll

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pabsb %xmm0, %xmm1
 ; GENERIC-NEXT:    pabsb (%rdi), %xmm0
 ; GENERIC-NEXT:    por %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pabsb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pabsb (%rdi), %xmm1
-; ATOM-NEXT:    pabsb %xmm0, %xmm0
-; ATOM-NEXT:    por %xmm0, %xmm1
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pabsb (%rdi), %xmm1 # sched: [1:1.00]
+; ATOM-NEXT:    pabsb %xmm0, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    por %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pabsb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pabsb %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    pabsb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    por %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pabsd %xmm0, %xmm1
 ; GENERIC-NEXT:    pabsd (%rdi), %xmm0
 ; GENERIC-NEXT:    por %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pabsd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pabsd (%rdi), %xmm1
-; ATOM-NEXT:    pabsd %xmm0, %xmm0
-; ATOM-NEXT:    por %xmm0, %xmm1
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pabsd (%rdi), %xmm1 # sched: [1:1.00]
+; ATOM-NEXT:    pabsd %xmm0, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    por %xmm0, %xmm1 # sched: [1:0.50]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pabsd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pabsd %xmm0, %xmm1 # sched: [1:0.50]
 ; SLM-NEXT:    pabsd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    por %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 define <8 x i16> @test_pabsw(<8 x i16> %a0, <8 x i16> *%a1) {
 ; GENERIC-LABEL: test_pabsw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pabsw %xmm0, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pabsw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pabsw %xmm0, %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pabsw %xmm0, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pabsw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pabsw %xmm0, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pabsw:
 ; SANDY:       # BB#0:

 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    palignr {{.*#+}} xmm1 = xmm0[6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5]
 ; GENERIC-NEXT:    palignr {{.*#+}} xmm1 = mem[14,15],xmm1[0,1,2,3,4,5,6,7,8,9,10,11,12,13]
 ; GENERIC-NEXT:    movdqa %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_palignr:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    palignr {{.*#+}} xmm1 = xmm0[6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5]
-; ATOM-NEXT:    palignr {{.*#+}} xmm1 = mem[14,15],xmm1[0,1,2,3,4,5,6,7,8,9,10,11,12,13]
-; ATOM-NEXT:    movdqa %xmm1, %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    palignr {{.*#+}} xmm1 = xmm0[6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5] sched: [1:1.00]
+; ATOM-NEXT:    palignr {{.*#+}} xmm1 = mem[14,15],xmm1[0,1,2,3,4,5,6,7,8,9,10,11,12,13] sched: [1:1.00]
+; ATOM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_palignr:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    palignr {{.*#+}} xmm1 = xmm0[6,7,8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5] sched: [1:1.00]
 ; SLM-NEXT:    palignr {{.*#+}} xmm1 = mem[14,15],xmm1[0,1,2,3,4,5,6,7,8,9,10,11,12,13] sched: [4:1.00]
 ; SLM-NEXT:    movdqa %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;

 ; GENERIC-LABEL: test_phaddd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    phaddd %xmm1, %xmm0
 ; GENERIC-NEXT:    phaddd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_phaddd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    phaddd %xmm1, %xmm0
-; ATOM-NEXT:    phaddd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    phaddd %xmm1, %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    phaddd (%rdi), %xmm0 # sched: [4:2.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_phaddd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    phaddd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    phaddd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_phaddd:

 ; GENERIC-NEXT:    phaddsw %xmm1, %xmm0
 ; GENERIC-NEXT:    phaddsw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_phaddsw:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    phaddsw %xmm1, %xmm0
 ; ATOM-NEXT:    phaddsw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_phaddsw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    phaddsw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    phaddsw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_phaddsw:

 ; GENERIC-LABEL: test_phaddw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    phaddw %xmm1, %xmm0
 ; GENERIC-NEXT:    phaddw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_phaddw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    phaddw %xmm1, %xmm0
-; ATOM-NEXT:    phaddw (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    phaddw %xmm1, %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    phaddw (%rdi), %xmm0 # sched: [8:4.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_phaddw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    phaddw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    phaddw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_phaddw:

 ; GENERIC-LABEL: test_phsubd:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    phsubd %xmm1, %xmm0
 ; GENERIC-NEXT:    phsubd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_phsubd:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    phsubd %xmm1, %xmm0
-; ATOM-NEXT:    phsubd (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    phsubd %xmm1, %xmm0 # sched: [3:1.50]
+; ATOM-NEXT:    phsubd (%rdi), %xmm0 # sched: [4:2.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_phsubd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    phsubd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    phsubd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_phsubd:

 ; GENERIC-NEXT:    phsubsw %xmm1, %xmm0
 ; GENERIC-NEXT:    phsubsw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_phsubsw:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    phsubsw %xmm1, %xmm0
 ; ATOM-NEXT:    phsubsw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_phsubsw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    phsubsw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    phsubsw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_phsubsw:

 ; GENERIC-LABEL: test_phsubw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    phsubw %xmm1, %xmm0
 ; GENERIC-NEXT:    phsubw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_phsubw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    phsubw %xmm1, %xmm0
-; ATOM-NEXT:    phsubw (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    phsubw %xmm1, %xmm0 # sched: [7:3.50]
+; ATOM-NEXT:    phsubw (%rdi), %xmm0 # sched: [8:4.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_phsubw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    phsubw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    phsubw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_phsubw:

 ; GENERIC-LABEL: test_pmaddubsw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pmaddubsw %xmm1, %xmm0
 ; GENERIC-NEXT:    pmaddubsw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pmaddubsw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pmaddubsw %xmm1, %xmm0
-; ATOM-NEXT:    pmaddubsw (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pmaddubsw %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    pmaddubsw (%rdi), %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pmaddubsw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pmaddubsw %xmm1, %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    pmaddubsw (%rdi), %xmm0 # sched: [7:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pmaddubsw:

 define <8 x i16> @test_pmulhrsw(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> *%a2) {
 ; GENERIC-LABEL: test_pmulhrsw:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pmulhrsw %xmm1, %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pmulhrsw:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pmulhrsw %xmm1, %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pmulhrsw %xmm1, %xmm0 # sched: [5:5.00]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pmulhrsw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pmulhrsw %xmm1, %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pmulhrsw:
 ; SANDY:       # BB#0:

 ; GENERIC-LABEL: test_pshufb:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pshufb %xmm1, %xmm0
 ; GENERIC-NEXT:    pshufb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_pshufb:
 ; ATOM:       # BB#0:
-; ATOM-NEXT:    pshufb %xmm1, %xmm0
-; ATOM-NEXT:    pshufb (%rdi), %xmm0
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    pshufb %xmm1, %xmm0 # sched: [4:2.00]
+; ATOM-NEXT:    pshufb (%rdi), %xmm0 # sched: [5:2.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_pshufb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    pshufb %xmm1, %xmm0 # sched: [1:1.00]
 ; SLM-NEXT:    pshufb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pshufb:

 ; GENERIC-NEXT:    psignb %xmm1, %xmm0
 ; GENERIC-NEXT:    psignb (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psignb:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    psignb %xmm1, %xmm0
 ; ATOM-NEXT:    psignb (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psignb:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psignb %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psignb (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psignb:

 ; GENERIC-NEXT:    psignd %xmm1, %xmm0
 ; GENERIC-NEXT:    psignd (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psignd:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    psignd %xmm1, %xmm0
 ; ATOM-NEXT:    psignd (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psignd:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psignd %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psignd (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psignd:

 ; GENERIC-NEXT:    psignw %xmm1, %xmm0
 ; GENERIC-NEXT:    psignw (%rdi), %xmm0
 ; GENERIC-NEXT:    retq
 ;
 ; ATOM-LABEL: test_psignw:
 ; ATOM:       # BB#0:
 ; ATOM-NEXT:    psignw %xmm1, %xmm0
 ; ATOM-NEXT:    psignw (%rdi), %xmm0
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    nop
-; ATOM-NEXT:    retq
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    nop # sched: [1:0.50]
+; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_psignw:
 ; SLM:       # BB#0:
 ; SLM-NEXT:    psignw %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    psignw (%rdi), %xmm0 # sched: [4:1.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_psignw: