On recent Intel u-arch's, folding loads into some unary SSE instructions can

be non-optimal. To be precise, we should avoid folding loads if the instructions
only update part of the destination register, and the non-updated part is not
needed. e.g. cvtss2sd, sqrtss. Unfolding the load from these instructions breaks
the partial register dependency and it can improve performance. e.g.

movss (%rdi), %xmm0
cvtss2sd %xmm0, %xmm0

instead of
cvtss2sd (%rdi), %xmm0

An alternative method to break dependency is to clear the register first. e.g.
xorps %xmm0, %xmm0
cvtss2sd (%rdi), %xmm0


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@91672 91177308-0d34-0410-b5e6-96231b3b80d8

7 files changed, 110 insertions, 17 deletions

diff --git a/lib/Target/X86/X86.td b/lib/Target/X86/X86.td
index a6e1ca3..313177e 100644
--- a/lib/Target/X86/X86.td
+++ b/lib/Target/X86/X86.td
@@ -57,6 +57,8 @@ def Feature64Bit   : SubtargetFeature<"64bit", "HasX86_64", "true",
                                       "Support 64-bit instructions">;
 def FeatureSlowBTMem : SubtargetFeature<"slow-bt-mem", "IsBTMemSlow", "true",
                                        "Bit testing of memory is slow">;
+def FeatureBreakSSEDep : SubtargetFeature<"break-sse-dep", "BreakSSEDep","true",
+                       "Should break SSE partial update dep with load / xorps">;
 def FeatureSSE4A   : SubtargetFeature<"sse4a", "HasSSE4A", "true",
                                       "Support SSE 4a instructions">;
 
@@ -86,17 +88,27 @@ def : Proc<"pentium2",        [FeatureMMX, FeatureCMOV]>;
 def : Proc<"pentium3",        [FeatureSSE1]>;
 def : Proc<"pentium-m",       [FeatureSSE2, FeatureSlowBTMem]>;
 def : Proc<"pentium4",        [FeatureSSE2]>;
-def : Proc<"x86-64",          [FeatureSSE2,   Feature64Bit, FeatureSlowBTMem]>;
-def : Proc<"yonah",           [FeatureSSE3, FeatureSlowBTMem]>;
-def : Proc<"prescott",        [FeatureSSE3, FeatureSlowBTMem]>;
-def : Proc<"nocona",          [FeatureSSE3,   Feature64Bit, FeatureSlowBTMem]>;
-def : Proc<"core2",           [FeatureSSSE3,  Feature64Bit, FeatureSlowBTMem]>;
-def : Proc<"penryn",          [FeatureSSE41,  Feature64Bit, FeatureSlowBTMem]>;
-def : Proc<"atom",            [FeatureSSE3,   Feature64Bit, FeatureSlowBTMem]>;
-def : Proc<"corei7",          [FeatureSSE42,  Feature64Bit, FeatureSlowBTMem]>;
-def : Proc<"nehalem",         [FeatureSSE42,  Feature64Bit, FeatureSlowBTMem]>;
+def : Proc<"x86-64",          [FeatureSSE2,   Feature64Bit, FeatureSlowBTMem,
+                               FeatureBreakSSEDep]>;
+def : Proc<"yonah",           [FeatureSSE3, FeatureSlowBTMem,
+                               FeatureBreakSSEDep]>;
+def : Proc<"prescott",        [FeatureSSE3, FeatureSlowBTMem,
+                               FeatureBreakSSEDep]>;
+def : Proc<"nocona",          [FeatureSSE3,   Feature64Bit, FeatureSlowBTMem,
+                               FeatureBreakSSEDep]>;
+def : Proc<"core2",           [FeatureSSSE3,  Feature64Bit, FeatureSlowBTMem,
+                               FeatureBreakSSEDep]>;
+def : Proc<"penryn",          [FeatureSSE41,  Feature64Bit, FeatureSlowBTMem,
+                               FeatureBreakSSEDep]>;
+def : Proc<"atom",            [FeatureSSE3,   Feature64Bit, FeatureSlowBTMem,
+                               FeatureBreakSSEDep]>;
+def : Proc<"corei7",          [FeatureSSE42,  Feature64Bit, FeatureSlowBTMem,
+                               FeatureBreakSSEDep]>;
+def : Proc<"nehalem",         [FeatureSSE42,  Feature64Bit, FeatureSlowBTMem,
+                               FeatureBreakSSEDep]>;
 // Sandy Bridge does not have FMA
-def : Proc<"sandybridge",     [FeatureSSE42,  FeatureAVX,   Feature64Bit]>;
+def : Proc<"sandybridge",     [FeatureSSE42,  FeatureAVX,   Feature64Bit,
+                               FeatureBreakSSEDep]>;
 
 def : Proc<"k6",              [FeatureMMX]>;
 def : Proc<"k6-2",            [FeatureMMX,    Feature3DNow]>;
diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp
index bc72f63..e1e6ff3 100644
--- a/lib/Target/X86/X86InstrInfo.cpp
+++ b/lib/Target/X86/X86InstrInfo.cpp
@@ -2370,6 +2370,23 @@ MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
   // Check switch flag 
   if (NoFusing) return NULL;
 
+  if (TM.getSubtarget<X86Subtarget>().shouldBreakSSEDep())
+    switch (MI->getOpcode()) {
+    case X86::CVTSD2SSrr:
+    case X86::Int_CVTSD2SSrr:
+    case X86::CVTSS2SDrr:
+    case X86::Int_CVTSS2SDrr:
+    case X86::RCPSSr:
+    case X86::RCPSSr_Int:
+    case X86::ROUNDSDr_Int:
+    case X86::ROUNDSSr_Int:
+    case X86::RSQRTSSr:
+    case X86::RSQRTSSr_Int:
+    case X86::SQRTSSr:
+    case X86::SQRTSSr_Int:
+      return 0;
+    }
+
   const MachineFrameInfo *MFI = MF.getFrameInfo();
   unsigned Size = MFI->getObjectSize(FrameIndex);
   unsigned Alignment = MFI->getObjectAlignment(FrameIndex);
@@ -2405,6 +2422,23 @@ MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
   // Check switch flag 
   if (NoFusing) return NULL;
 
+  if (TM.getSubtarget<X86Subtarget>().shouldBreakSSEDep())
+    switch (MI->getOpcode()) {
+    case X86::CVTSD2SSrr:
+    case X86::Int_CVTSD2SSrr:
+    case X86::CVTSS2SDrr:
+    case X86::Int_CVTSS2SDrr:
+    case X86::RCPSSr:
+    case X86::RCPSSr_Int:
+    case X86::ROUNDSDr_Int:
+    case X86::ROUNDSSr_Int:
+    case X86::RSQRTSSr:
+    case X86::RSQRTSSr_Int:
+    case X86::SQRTSSr:
+    case X86::SQRTSSr_Int:
+      return 0;
+    }
+
   // Determine the alignment of the load.
   unsigned Alignment = 0;
   if (LoadMI->hasOneMemOperand())
diff --git a/lib/Target/X86/X86InstrInfo.td b/lib/Target/X86/X86InstrInfo.td
index 7411dde..0b6efaa 100644
--- a/lib/Target/X86/X86InstrInfo.td
+++ b/lib/Target/X86/X86InstrInfo.td
@@ -301,6 +301,8 @@ def IsStatic     : Predicate<"TM.getRelocationModel() == Reloc::Static">;
 def OptForSpeed  : Predicate<"!OptForSize">;
 def FastBTMem    : Predicate<"!Subtarget->isBTMemSlow()">;
 def CallImmAddr  : Predicate<"Subtarget->IsLegalToCallImmediateAddr(TM)">;
+def SSEBreakDep  : Predicate<"Subtarget->shouldBreakSSEDep() && !OptForSize">;
+def NoSSEBreakDep: Predicate<"!Subtarget->shouldBreakSSEDep() || OptForSize">;
 
 //===----------------------------------------------------------------------===//
 // X86 Instruction Format Definitions.
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index ae1a68a..694b91e 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -824,9 +824,10 @@ multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
   }
 
   // Scalar operation, mem.
-  def SSm : SSI<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
+  def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
                 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
-                [(set FR32:$dst, (OpNode (load addr:$src)))]>;
+                [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
+            Requires<[HasSSE1, NoSSEBreakDep]>;
 
   // Vector operation, reg.
   def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
@@ -1116,9 +1117,10 @@ def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
 def CVTSD2SSrr  : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
                       "cvtsd2ss\t{$src, $dst|$dst, $src}",
                       [(set FR32:$dst, (fround FR64:$src))]>;
-def CVTSD2SSrm  : SDI<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
+def CVTSD2SSrm  : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
                       "cvtsd2ss\t{$src, $dst|$dst, $src}",
-                      [(set FR32:$dst, (fround (loadf64 addr:$src)))]>;
+                      [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
+                      Requires<[HasSSE2, NoSSEBreakDep]>;
 def CVTSI2SDrr  : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
                       "cvtsi2sd\t{$src, $dst|$dst, $src}",
                       [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
@@ -1155,7 +1157,10 @@ def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
                    "cvtss2sd\t{$src, $dst|$dst, $src}",
                    [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
-                 Requires<[HasSSE2]>;
+                 Requires<[HasSSE2, NoSSEBreakDep]>;
+
+def : Pat<(extloadf32 addr:$src),
+          (CVTSS2SDrr (MOVSSrm addr:$src))>, Requires<[SSEBreakDep]>;
 
 // Match intrinsics which expect XMM operand(s).
 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
@@ -3220,13 +3225,14 @@ multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
                     OpSize;
 
   // Vector intrinsic operation, mem
-  def PSm_Int : SS4AIi8<opcps, MRMSrcMem,
+  def PSm_Int : Ii8<opcps, MRMSrcMem,
                     (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
                     !strconcat(OpcodeStr,
                     "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
                     [(set VR128:$dst,
                           (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
-                    OpSize;
+                    TA, OpSize,
+                Requires<[HasSSE41, NoSSEBreakDep]>;
 
   // Vector intrinsic operation, reg
   def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
diff --git a/lib/Target/X86/X86Subtarget.cpp b/lib/Target/X86/X86Subtarget.cpp
index 75cdbad..4db3fdb 100644
--- a/lib/Target/X86/X86Subtarget.cpp
+++ b/lib/Target/X86/X86Subtarget.cpp
@@ -266,6 +266,7 @@ void X86Subtarget::AutoDetectSubtargetFeatures() {
     unsigned Model  = 0;
     DetectFamilyModel(EAX, Family, Model);
     IsBTMemSlow = IsAMD || (Family == 6 && Model >= 13);
+    BreakSSEDep = IsIntel;
 
     GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
     HasX86_64 = (EDX >> 29) & 0x1;
@@ -286,6 +287,7 @@ X86Subtarget::X86Subtarget(const std::string &TT, const std::string &FS,
   , HasFMA3(false)
   , HasFMA4(false)
   , IsBTMemSlow(false)
+  , BreakSSEDep(false)
   , DarwinVers(0)
   , stackAlignment(8)
   // FIXME: this is a known good value for Yonah. How about others?
diff --git a/lib/Target/X86/X86Subtarget.h b/lib/Target/X86/X86Subtarget.h
index fb457dd..b2b48ed 100644
--- a/lib/Target/X86/X86Subtarget.h
+++ b/lib/Target/X86/X86Subtarget.h
@@ -77,6 +77,14 @@ protected:
 
   /// IsBTMemSlow - True if BT (bit test) of memory instructions are slow.
   bool IsBTMemSlow;
+
+  /// BreakSSEDep - True if codegen should unfold load or insert xorps / pxor
+  /// to break register dependency for a partial register update SSE
+  /// instruction. This is needed for instructions such as CVTSS2SD which
+  /// only update the lower part of the register, and the result of the updated
+  /// part does not depend on the contents of the destination before the
+  /// instruction, and the non-updated portion of the register is not used.
+  bool BreakSSEDep;
   
   /// DarwinVers - Nonzero if this is a darwin platform: the numeric
   /// version of the platform, e.g. 8 = 10.4 (Tiger), 9 = 10.5 (Leopard), etc.
@@ -142,6 +150,7 @@ public:
   bool hasFMA3() const { return HasFMA3; }
   bool hasFMA4() const { return HasFMA4; }
   bool isBTMemSlow() const { return IsBTMemSlow; }
+  bool shouldBreakSSEDep() const { return BreakSSEDep; }
 
   bool isTargetDarwin() const { return TargetType == isDarwin; }
   bool isTargetELF() const { return TargetType == isELF; }
diff --git a/test/CodeGen/X86/break-sse-dep.ll b/test/CodeGen/X86/break-sse-dep.ll
new file mode 100644
index 0000000..00c943f
--- /dev/null
+++ b/test/CodeGen/X86/break-sse-dep.ll
@@ -0,0 +1,28 @@
+; RUN: llc < %s -march=x86-64 -mattr=+sse2,+break-sse-dep | FileCheck %s --check-prefix=YES
+; RUN: llc < %s -march=x86-64 -mattr=+sse2,-break-sse-dep | FileCheck %s --check-prefix=NO
+
+define double @t1(float* nocapture %x) nounwind readonly ssp {
+entry:
+; YES: t1:
+; YES: movss (%rdi), %xmm0
+; YES; cvtss2sd %xmm0, %xmm0
+
+; NO: t1:
+; NO; cvtss2sd (%rdi), %xmm0
+  %0 = load float* %x, align 4
+  %1 = fpext float %0 to double
+  ret double %1
+}
+
+define float @t2(double* nocapture %x) nounwind readonly ssp {
+entry:
+; YES: t2:
+; YES: movsd (%rdi), %xmm0
+; YES; cvtsd2ss %xmm0, %xmm0
+
+; NO: t2:
+; NO; cvtsd2ss (%rdi), %xmm0
+  %0 = load double* %x, align 8
+  %1 = fptrunc double %0 to float
+  ret float %1
+}