Add 256-bit load/store recognition and matching in several places.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@135171 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp
index 702331d8ed99345b913642a16b3ec512c0f73bfe..0509a1515a74bc09a69d29bb3827eee2fa2a9dad 100644 (file)
--- a/lib/Target/X86/X86InstrInfo.cpp
+++ b/lib/Target/X86/X86InstrInfo.cpp
@@ -301,12 +301,17 @@ X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
     { X86::MOVAPDrr,    X86::MOVAPDmr, 0, 16 },
     { X86::MOVAPSrr,    X86::MOVAPSmr, 0, 16 },
     { X86::MOVDQArr,    X86::MOVDQAmr, 0, 16 },
+    { X86::VMOVAPDYrr,  X86::VMOVAPDYmr, 0, 32 },
+    { X86::VMOVAPSYrr,  X86::VMOVAPSYmr, 0, 32 },
+    { X86::VMOVDQAYrr,  X86::VMOVDQAYmr, 0, 32 },
     { X86::MOVPDI2DIrr, X86::MOVPDI2DImr, 0, 0 },
     { X86::MOVPQIto64rr,X86::MOVPQI2QImr, 0, 0 },
     { X86::MOVSDto64rr, X86::MOVSDto64mr, 0, 0 },
     { X86::MOVSS2DIrr,  X86::MOVSS2DImr, 0, 0 },
     { X86::MOVUPDrr,    X86::MOVUPDmr, 0, 0 },
     { X86::MOVUPSrr,    X86::MOVUPSmr, 0, 0 },
+    { X86::VMOVUPDYrr,  X86::VMOVUPDYmr, 0, 0 },
+    { X86::VMOVUPSYrr,  X86::VMOVUPSYmr, 0, 0 },
     { X86::MUL16r,      X86::MUL16m, 1, 0 },
     { X86::MUL32r,      X86::MUL32m, 1, 0 },
     { X86::MUL64r,      X86::MUL64m, 1, 0 },
@@ -411,10 +416,13 @@ X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
     { X86::MOV8rr,          X86::MOV8rm, 0 },
     { X86::MOVAPDrr,        X86::MOVAPDrm, 16 },
     { X86::MOVAPSrr,        X86::MOVAPSrm, 16 },
+    { X86::VMOVAPDYrr,      X86::VMOVAPDYrm, 32 },
+    { X86::VMOVAPSYrr,      X86::VMOVAPSYrm, 32 },
     { X86::MOVDDUPrr,       X86::MOVDDUPrm, 0 },
     { X86::MOVDI2PDIrr,     X86::MOVDI2PDIrm, 0 },
     { X86::MOVDI2SSrr,      X86::MOVDI2SSrm, 0 },
     { X86::MOVDQArr,        X86::MOVDQArm, 16 },
+    { X86::VMOVDQAYrr,      X86::VMOVDQAYrm, 16 },
     { X86::MOVSHDUPrr,      X86::MOVSHDUPrm, 16 },
     { X86::MOVSLDUPrr,      X86::MOVSLDUPrm, 16 },
     { X86::MOVSX16rr8,      X86::MOVSX16rm8, 0 },
@@ -425,6 +433,8 @@ X86InstrInfo::X86InstrInfo(X86TargetMachine &tm)
     { X86::MOVSX64rr8,      X86::MOVSX64rm8, 0 },
     { X86::MOVUPDrr,        X86::MOVUPDrm, 16 },
     { X86::MOVUPSrr,        X86::MOVUPSrm, 0 },
+    { X86::VMOVUPDYrr,      X86::VMOVUPDYrm, 0 },
+    { X86::VMOVUPSYrr,      X86::VMOVUPSYrm, 0 },
     { X86::MOVZDI2PDIrr,    X86::MOVZDI2PDIrm, 0 },
     { X86::MOVZQI2PQIrr,    X86::MOVZQI2PQIrm, 0 },
     { X86::MOVZPQILo2PQIrr, X86::MOVZPQILo2PQIrm, 16 },
@@ -787,6 +797,9 @@ static bool isFrameLoadOpcode(int Opcode) {
   case X86::MOVAPSrm:
   case X86::MOVAPDrm:
   case X86::MOVDQArm:
+  case X86::VMOVAPSYrm:
+  case X86::VMOVAPDYrm:
+  case X86::VMOVDQAYrm:
   case X86::MMX_MOVD64rm:
   case X86::MMX_MOVQ64rm:
     return true;
@@ -808,6 +821,9 @@ static bool isFrameStoreOpcode(int Opcode) {
   case X86::MOVAPSmr:
   case X86::MOVAPDmr:
   case X86::MOVDQAmr:
+  case X86::VMOVAPSYmr:
+  case X86::VMOVAPDYmr:
+  case X86::VMOVDQAYmr:
   case X86::MMX_MOVD64mr:
   case X86::MMX_MOVQ64mr:
   case X86::MMX_MOVNTQmr:
@@ -926,6 +942,10 @@ X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
     case X86::MOVUPSrm:
     case X86::MOVAPDrm:
     case X86::MOVDQArm:
+    case X86::VMOVAPSYrm:
+    case X86::VMOVUPSYrm:
+    case X86::VMOVAPDYrm:
+    case X86::VMOVDQAYrm:
     case X86::MMX_MOVD64rm:
     case X86::MMX_MOVQ64rm:
     case X86::FsMOVAPSrm:
@@ -1975,6 +1995,8 @@ void X86InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
       Opc = X86::MOV8rr;
   } else if (X86::VR128RegClass.contains(DestReg, SrcReg))
     Opc = X86::MOVAPSrr;
+  else if (X86::VR256RegClass.contains(DestReg, SrcReg))
+    Opc = X86::VMOVAPSYrr;
   else if (X86::VR64RegClass.contains(DestReg, SrcReg))
     Opc = X86::MMX_MOVQ64rr;
   else
@@ -2064,6 +2086,13 @@ static unsigned getLoadStoreRegOpcode(unsigned Reg,
       return load ? X86::MOVAPSrm : X86::MOVAPSmr;
     else
       return load ? X86::MOVUPSrm : X86::MOVUPSmr;
+  case 32:
+    assert(X86::VR256RegClass.hasSubClassEq(RC) && "Unknown 32-byte regclass");
+    // If stack is realigned we can use aligned stores.
+    if (isStackAligned)
+      return load ? X86::VMOVAPSYrm : X86::VMOVAPSYmr;
+    else
+      return load ? X86::VMOVUPSYrm : X86::VMOVUPSYmr;
   }
 }
 
@@ -2853,6 +2882,11 @@ X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
   case X86::MOVAPDrm:
   case X86::MOVDQArm:
   case X86::MOVDQUrm:
+  case X86::VMOVAPSYrm:
+  case X86::VMOVUPSYrm:
+  case X86::VMOVAPDYrm:
+  case X86::VMOVDQAYrm:
+  case X86::VMOVDQUYrm:
     break;
   }
   switch (Opc2) {
@@ -2875,6 +2909,11 @@ X86InstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
   case X86::MOVAPDrm:
   case X86::MOVDQArm:
   case X86::MOVDQUrm:
+  case X86::VMOVAPSYrm:
+  case X86::VMOVUPSYrm:
+  case X86::VMOVAPDYrm:
+  case X86::VMOVDQAYrm:
+  case X86::VMOVDQUYrm:
     break;
   }
 
@@ -3053,6 +3092,13 @@ static const unsigned ReplaceableInstrs[][3] = {
   { X86::AVX_SET0PS, X86::AVX_SET0PD, X86::AVX_SET0PI },
   { X86::VXORPSrm,   X86::VXORPDrm,   X86::VPXORrm    },
   { X86::VXORPSrr,   X86::VXORPDrr,   X86::VPXORrr    },
+  // AVX 256-bit support
+  { X86::VMOVAPSYmr,   X86::VMOVAPDYmr,   X86::VMOVDQAYmr  },
+  { X86::VMOVAPSYrm,   X86::VMOVAPDYrm,   X86::VMOVDQAYrm  },
+  { X86::VMOVAPSYrr,   X86::VMOVAPDYrr,   X86::VMOVDQAYrr  },
+  { X86::VMOVUPSYmr,   X86::VMOVUPDYmr,   X86::VMOVDQUYmr  },
+  { X86::VMOVUPSYrm,   X86::VMOVUPDYrm,   X86::VMOVDQUYrm  },
+  { X86::VMOVNTPSYmr,  X86::VMOVNTPDYmr,  X86::VMOVNTDQYmr },
 };
 
 // FIXME: Some shuffle and unpack instructions have equivalents in different

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index e6167c73d3094aaa60b12abf72bfea29f1b57a45..da66e87a59fc73eef926c6936be5081230f846bb 100644 (file)
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -2022,7 +2022,10 @@ def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
 }
 
 // The same as done above but for AVX. The 128-bit versions are the
-// same, but re-encoded. The 256-bit does not support PI version.
+// same, but re-encoded. The 256-bit does not support PI version, and
+// doesn't need it because on sandy bridge the register is set to zero
+// at the rename stage without using any execution unit, so SET0PSY
+// and SET0PDY can be used for vector int instructions without penalty
 // FIXME: Change encoding to pseudo! This is blocked right now by the x86
 // JIT implementatioan, it does not expand the instructions below like
 // X86MCInstLower does.
@@ -2037,8 +2040,8 @@ def AVX_SET0PSY : PSI<0x57, MRMInitReg, (outs VR256:$dst), (ins), "",
 def AVX_SET0PDY : PDI<0x57, MRMInitReg, (outs VR256:$dst), (ins), "",
                    [(set VR256:$dst, (v4f64 immAllZerosV))]>, VEX_4V;
 let ExeDomain = SSEPackedInt in
-def AVX_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
-                 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
+def AVX_SET0PI  : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
+                   [(set VR128:$dst, (v4i32 immAllZerosV))]>;
 }
 
 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
@@ -3831,6 +3834,8 @@ def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
           (CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
 
 // Use movaps / movups for SSE integer load / store (one byte shorter).
+// The instructions selected below are then converted to MOVDQA/MOVDQU
+// during the SSE domain pass.
 let Predicates = [HasSSE1] in {
   def : Pat<(alignedloadv4i32 addr:$src),
             (MOVAPSrm addr:$src)>;
@@ -3859,8 +3864,9 @@ let Predicates = [HasSSE1] in {
             (MOVUPSmr addr:$dst, VR128:$src)>;
 }
 
-// Use vmovaps/vmovups for AVX 128-bit integer load/store (one byte shorter).
+// Use vmovaps/vmovups for AVX integer load/store.
 let Predicates = [HasAVX] in {
+  // 128-bit load/store
   def : Pat<(alignedloadv4i32 addr:$src),
             (VMOVAPSrm addr:$src)>;
   def : Pat<(loadv4i32 addr:$src),
@@ -3886,6 +3892,24 @@ let Predicates = [HasAVX] in {
             (VMOVUPSmr addr:$dst, VR128:$src)>;
   def : Pat<(store (v16i8 VR128:$src), addr:$dst),
             (VMOVUPSmr addr:$dst, VR128:$src)>;
+
+  // 256-bit load/store
+  def : Pat<(alignedloadv4i64 addr:$src),
+            (VMOVAPSYrm addr:$src)>;
+  def : Pat<(loadv4i64 addr:$src),
+            (VMOVUPSYrm addr:$src)>;
+  def : Pat<(alignedloadv8i32 addr:$src),
+            (VMOVAPSYrm addr:$src)>;
+  def : Pat<(loadv8i32 addr:$src),
+            (VMOVUPSYrm addr:$src)>;
+  def : Pat<(alignedstore (v4i64 VR256:$src), addr:$dst),
+            (VMOVAPSYmr addr:$dst, VR256:$src)>;
+  def : Pat<(alignedstore (v8i32 VR256:$src), addr:$dst),
+            (VMOVAPSYmr addr:$dst, VR256:$src)>;
+  def : Pat<(store (v4i64 VR256:$src), addr:$dst),
+            (VMOVUPSYmr addr:$dst, VR256:$src)>;
+  def : Pat<(store (v8i32 VR256:$src), addr:$dst),
+            (VMOVUPSYmr addr:$dst, VR256:$src)>;
 }
 
 //===----------------------------------------------------------------------===//

diff --git a/test/CodeGen/X86/avx-load-store.ll b/test/CodeGen/X86/avx-load-store.ll
new file mode 100644 (file)
index 0000000..5196089
--- /dev/null
+++ b/test/CodeGen/X86/avx-load-store.ll
@@ -0,0 +1,24 @@
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=corei7-avx -mattr=+avx | FileCheck %s
+
+; CHECK: vmovaps
+; CHECK: vmovaps
+; CHECK: vmovapd
+; CHECK: vmovapd
+; CHECK: vmovaps
+; CHECK: vmovaps
+define void @test_256_load(double* nocapture %d, float* nocapture %f, <4 x i64>* nocapture %i) nounwind uwtable ssp {
+entry:
+  %0 = bitcast double* %d to <4 x double>*
+  %tmp1.i = load <4 x double>* %0, align 32
+  %1 = bitcast float* %f to <8 x float>*
+  %tmp1.i17 = load <8 x float>* %1, align 32
+  %tmp1.i16 = load <4 x i64>* %i, align 32
+  tail call void @dummy(<4 x double> %tmp1.i, <8 x float> %tmp1.i17, <4 x i64> %tmp1.i16) nounwind
+  store <4 x double> %tmp1.i, <4 x double>* %0, align 32
+  store <8 x float> %tmp1.i17, <8 x float>* %1, align 32
+  store <4 x i64> %tmp1.i16, <4 x i64>* %i, align 32
+  ret void
+}
+
+declare void @dummy(<4 x double>, <8 x float>, <4 x i64>)
+