R600: Emit native instructions for tex

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178452 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 197 insertions, 115 deletions

diff --git a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
index d20716000d..a945fe9ef8 100644
--- a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
+++ b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
@@ -66,8 +66,6 @@ private:
   void EmitSrcISA(const MCInst &MI, unsigned RegOpIdx, unsigned SelOpIdx,
                     raw_ostream &OS) const;
   void EmitDst(const MCInst &MI, raw_ostream &OS) const;
-  void EmitTexInstr(const MCInst &MI, SmallVectorImpl<MCFixup> &Fixups,
-                    raw_ostream &OS) const;
   void EmitFCInstr(const MCInst &MI, raw_ostream &OS) const;
 
   void EmitNullBytes(unsigned int byteCount, raw_ostream &OS) const;
@@ -140,9 +138,7 @@ MCCodeEmitter *llvm::createR600MCCodeEmitter(const MCInstrInfo &MCII,
 
 void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
                                        SmallVectorImpl<MCFixup> &Fixups) const {
-  if (isTexOp(MI.getOpcode())) {
-    EmitTexInstr(MI, Fixups, OS);
-  } else if (isFCOp(MI.getOpcode())){
+  if (isFCOp(MI.getOpcode())){
     EmitFCInstr(MI, OS);
   } else if (MI.getOpcode() == AMDGPU::RETURN ||
     MI.getOpcode() == AMDGPU::BUNDLE ||
@@ -175,6 +171,77 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
       Emit(InstWord2, OS);
       break;
     }
+    case AMDGPU::TEX_LD:
+    case AMDGPU::TEX_GET_TEXTURE_RESINFO:
+    case AMDGPU::TEX_SAMPLE:
+    case AMDGPU::TEX_SAMPLE_C:
+    case AMDGPU::TEX_SAMPLE_L:
+    case AMDGPU::TEX_SAMPLE_C_L:
+    case AMDGPU::TEX_SAMPLE_LB:
+    case AMDGPU::TEX_SAMPLE_C_LB:
+    case AMDGPU::TEX_SAMPLE_G:
+    case AMDGPU::TEX_SAMPLE_C_G:
+    case AMDGPU::TEX_GET_GRADIENTS_H:
+    case AMDGPU::TEX_GET_GRADIENTS_V:
+    case AMDGPU::TEX_SET_GRADIENTS_H:
+    case AMDGPU::TEX_SET_GRADIENTS_V: {
+      unsigned Opcode = MI.getOpcode();
+      bool HasOffsets = (Opcode == AMDGPU::TEX_LD);
+      unsigned OpOffset = HasOffsets ? 3 : 0;
+      int64_t Sampler = MI.getOperand(OpOffset + 3).getImm();
+      int64_t TextureType = MI.getOperand(OpOffset + 4).getImm();
+
+      uint32_t SrcSelect[4] = {0, 1, 2, 3};
+      uint32_t Offsets[3] = {0, 0, 0};
+      uint64_t CoordType[4] = {1, 1, 1, 1};
+
+      if (HasOffsets)
+        for (unsigned i = 0; i < 3; i++)
+          Offsets[i] = MI.getOperand(i + 2).getImm();
+
+      if (TextureType == TEXTURE_RECT ||
+          TextureType == TEXTURE_SHADOWRECT) {
+        CoordType[ELEMENT_X] = 0;
+        CoordType[ELEMENT_Y] = 0;
+      }
+
+      if (TextureType == TEXTURE_1D_ARRAY ||
+          TextureType == TEXTURE_SHADOW1D_ARRAY) {
+        if (Opcode == AMDGPU::TEX_SAMPLE_C_L ||
+            Opcode == AMDGPU::TEX_SAMPLE_C_LB) {
+          CoordType[ELEMENT_Y] = 0;
+        } else {
+          CoordType[ELEMENT_Z] = 0;
+          SrcSelect[ELEMENT_Z] = ELEMENT_Y;
+        }
+      } else if (TextureType == TEXTURE_2D_ARRAY ||
+          TextureType == TEXTURE_SHADOW2D_ARRAY) {
+        CoordType[ELEMENT_Z] = 0;
+      }
+
+
+      if ((TextureType == TEXTURE_SHADOW1D ||
+          TextureType == TEXTURE_SHADOW2D ||
+          TextureType == TEXTURE_SHADOWRECT ||
+          TextureType == TEXTURE_SHADOW1D_ARRAY) &&
+          Opcode != AMDGPU::TEX_SAMPLE_C_L &&
+          Opcode != AMDGPU::TEX_SAMPLE_C_LB) {
+        SrcSelect[ELEMENT_W] = ELEMENT_Z;
+      }
+
+      uint64_t Word01 = getBinaryCodeForInstr(MI, Fixups) |
+          CoordType[ELEMENT_X] << 60 | CoordType[ELEMENT_Y] << 61 |
+          CoordType[ELEMENT_Z] << 62 | CoordType[ELEMENT_W] << 63;
+      uint32_t Word2 = Sampler << 15 | SrcSelect[ELEMENT_X] << 20 |
+          SrcSelect[ELEMENT_Y] << 23 | SrcSelect[ELEMENT_Z] << 26 |
+          SrcSelect[ELEMENT_W] << 29 | Offsets[0] << 0 | Offsets[1] << 5 |
+          Offsets[2] << 10;
+
+      EmitByte(INSTR_TEX, OS);
+      Emit(Word01, OS);
+      Emit(Word2, OS);
+      break;
+    }
     case AMDGPU::EG_ExportSwz:
     case AMDGPU::R600_ExportSwz:
     case AMDGPU::EG_ExportBuf:
@@ -334,99 +401,6 @@ void R600MCCodeEmitter::EmitSrcISA(const MCInst &MI, unsigned RegOpIdx,
   Emit(InlineConstant.i, OS);
 }
 
-void R600MCCodeEmitter::EmitTexInstr(const MCInst &MI,
-                                     SmallVectorImpl<MCFixup> &Fixups,
-                                     raw_ostream &OS) const {
-
-  unsigned Opcode = MI.getOpcode();
-  bool hasOffsets = (Opcode == AMDGPU::TEX_LD);
-  unsigned OpOffset = hasOffsets ? 3 : 0;
-  int64_t Resource = MI.getOperand(OpOffset + 2).getImm();
-  int64_t Sampler = MI.getOperand(OpOffset + 3).getImm();
-  int64_t TextureType = MI.getOperand(OpOffset + 4).getImm();
-  unsigned srcSelect[4] = {0, 1, 2, 3};
-
-  // Emit instruction type
-  EmitByte(1, OS);
-
-  // Emit instruction
-  EmitByte(getBinaryCodeForInstr(MI, Fixups), OS);
-
-  // Emit resource id
-  EmitByte(Resource, OS);
-
-  // Emit source register
-  EmitByte(getHWReg(MI.getOperand(1).getReg()), OS);
-
-  // XXX: Emit src isRelativeAddress
-  EmitByte(0, OS);
-
-  // Emit destination register
-  EmitByte(getHWReg(MI.getOperand(0).getReg()), OS);
-
-  // XXX: Emit dst isRealtiveAddress
-  EmitByte(0, OS);
-
-  // XXX: Emit dst select
-  EmitByte(0, OS); // X
-  EmitByte(1, OS); // Y
-  EmitByte(2, OS); // Z
-  EmitByte(3, OS); // W
-
-  // XXX: Emit lod bias
-  EmitByte(0, OS);
-
-  // XXX: Emit coord types
-  unsigned coordType[4] = {1, 1, 1, 1};
-
-  if (TextureType == TEXTURE_RECT
-      || TextureType == TEXTURE_SHADOWRECT) {
-    coordType[ELEMENT_X] = 0;
-    coordType[ELEMENT_Y] = 0;
-  }
-
-  if (TextureType == TEXTURE_1D_ARRAY
-      || TextureType == TEXTURE_SHADOW1D_ARRAY) {
-    if (Opcode == AMDGPU::TEX_SAMPLE_C_L || Opcode == AMDGPU::TEX_SAMPLE_C_LB) {
-      coordType[ELEMENT_Y] = 0;
-    } else {
-      coordType[ELEMENT_Z] = 0;
-      srcSelect[ELEMENT_Z] = ELEMENT_Y;
-    }
-  } else if (TextureType == TEXTURE_2D_ARRAY
-             || TextureType == TEXTURE_SHADOW2D_ARRAY) {
-    coordType[ELEMENT_Z] = 0;
-  }
-
-  for (unsigned i = 0; i < 4; i++) {
-    EmitByte(coordType[i], OS);
-  }
-
-  // XXX: Emit offsets
-  if (hasOffsets)
-	  for (unsigned i = 2; i < 5; i++)
-		  EmitByte(MI.getOperand(i).getImm()<<1, OS);
-  else
-	  EmitNullBytes(3, OS);
-
-  // Emit sampler id
-  EmitByte(Sampler, OS);
-
-  // XXX:Emit source select
-  if ((TextureType == TEXTURE_SHADOW1D
-      || TextureType == TEXTURE_SHADOW2D
-      || TextureType == TEXTURE_SHADOWRECT
-      || TextureType == TEXTURE_SHADOW1D_ARRAY)
-      && Opcode != AMDGPU::TEX_SAMPLE_C_L
-      && Opcode != AMDGPU::TEX_SAMPLE_C_LB) {
-    srcSelect[ELEMENT_W] = ELEMENT_Z;
-  }
-
-  for (unsigned i = 0; i < 4; i++) {
-    EmitByte(srcSelect[i], OS);
-  }
-}
-
 void R600MCCodeEmitter::EmitFCInstr(const MCInst &MI, raw_ostream &OS) const {
 
   // Emit instruction type
diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td
index a6daadfffa..7550ecec6f 100644
--- a/lib/Target/R600/R600Instructions.td
+++ b/lib/Target/R600/R600Instructions.td
@@ -234,6 +234,80 @@ class VTX_WORD1_GPR {
   let Word1{31}    = SRF_MODE_ALL;
 }
 
+class TEX_WORD0 {
+  field bits<32> Word0;
+
+  bits<5> TEX_INST;
+  bits<2> INST_MOD;
+  bits<1> FETCH_WHOLE_QUAD;
+  bits<8> RESOURCE_ID;
+  bits<7> SRC_GPR;
+  bits<1> SRC_REL;
+  bits<1> ALT_CONST;
+  bits<2> RESOURCE_INDEX_MODE;
+  bits<2> SAMPLER_INDEX_MODE;
+
+  let Word0{4-0} = TEX_INST;
+  let Word0{6-5} = INST_MOD;
+  let Word0{7} = FETCH_WHOLE_QUAD;
+  let Word0{15-8} = RESOURCE_ID;
+  let Word0{22-16} = SRC_GPR;
+  let Word0{23} = SRC_REL;
+  let Word0{24} = ALT_CONST;
+  let Word0{26-25} = RESOURCE_INDEX_MODE;
+  let Word0{28-27} = SAMPLER_INDEX_MODE;
+}
+
+class TEX_WORD1 {
+  field bits<32> Word1;
+
+  bits<7> DST_GPR;
+  bits<1> DST_REL;
+  bits<3> DST_SEL_X;
+  bits<3> DST_SEL_Y;
+  bits<3> DST_SEL_Z;
+  bits<3> DST_SEL_W;
+  bits<7> LOD_BIAS;
+  bits<1> COORD_TYPE_X;
+  bits<1> COORD_TYPE_Y;
+  bits<1> COORD_TYPE_Z;
+  bits<1> COORD_TYPE_W;
+
+  let Word1{6-0} = DST_GPR;
+  let Word1{7} = DST_REL;
+  let Word1{11-9} = DST_SEL_X;
+  let Word1{14-12} = DST_SEL_Y;
+  let Word1{17-15} = DST_SEL_Z;
+  let Word1{20-18} = DST_SEL_W;
+  let Word1{27-21} = LOD_BIAS;
+  let Word1{28} = COORD_TYPE_X;
+  let Word1{29} = COORD_TYPE_Y;
+  let Word1{30} = COORD_TYPE_Z;
+  let Word1{31} = COORD_TYPE_W;
+}
+
+class TEX_WORD2 {
+  field bits<32> Word2;
+
+  bits<5> OFFSET_X;
+  bits<5> OFFSET_Y;
+  bits<5> OFFSET_Z;
+  bits<5> SAMPLER_ID;
+  bits<3> SRC_SEL_X;
+  bits<3> SRC_SEL_Y;
+  bits<3> SRC_SEL_Z;
+  bits<3> SRC_SEL_W;
+
+  let Word2{4-0} = OFFSET_X;
+  let Word2{9-5} = OFFSET_Y;
+  let Word2{14-10} = OFFSET_Z;
+  let Word2{19-15} = SAMPLER_ID;
+  let Word2{22-20} = SRC_SEL_X;
+  let Word2{25-23} = SRC_SEL_Y;
+  let Word2{28-26} = SRC_SEL_Z;
+  let Word2{31-29} = SRC_SEL_W;
+}
+
 /*
 XXX: R600 subtarget uses a slightly different encoding than the other
 subtargets.  We currently handle this in R600MCCodeEmitter, but we may
@@ -386,12 +460,32 @@ class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
 class R600_TEX <bits<11> inst, string opName, list<dag> pattern,
                 InstrItinClass itin = AnyALU> :
   InstR600 <inst,
-          (outs R600_Reg128:$dst),
-          (ins R600_Reg128:$src0, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
-          !strconcat(opName, "$dst, $src0, $resourceId, $samplerId, $textureTarget"),
+          (outs R600_Reg128:$DST_GPR),
+          (ins R600_Reg128:$SRC_GPR, i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID, i32imm:$textureTarget),
+          !strconcat(opName, "$DST_GPR, $SRC_GPR, $RESOURCE_ID, $SAMPLER_ID, $textureTarget"),
           pattern,
-          itin>{
-    let Inst {10-0} = inst;
+          itin>, TEX_WORD0, TEX_WORD1, TEX_WORD2 {
+    let Inst{31-0} = Word0;
+    let Inst{63-32} = Word1;
+
+    let TEX_INST = inst{4-0};
+    let SRC_REL = 0;
+    let DST_REL = 0;
+    let DST_SEL_X = 0;
+    let DST_SEL_Y = 1;
+    let DST_SEL_Z = 2;
+    let DST_SEL_W = 3;
+    let LOD_BIAS = 0;
+
+    let INST_MOD = 0;
+    let FETCH_WHOLE_QUAD = 0;
+    let ALT_CONST = 0;
+    let SAMPLER_INDEX_MODE = 0;
+
+    let COORD_TYPE_X = 0;
+    let COORD_TYPE_Y = 0;
+    let COORD_TYPE_Z = 0;
+    let COORD_TYPE_W = 0;
   }
 
 } // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
@@ -867,25 +961,33 @@ def CNDGT_INT : R600_3OP <
 
 def TEX_LD : R600_TEX <
   0x03, "TEX_LD",
-  [(set R600_Reg128:$dst, (int_AMDGPU_txf R600_Reg128:$src0, imm:$src1, imm:$src2, imm:$src3, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+  [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txf R600_Reg128:$SRC_GPR,
+      imm:$OFFSET_X, imm:$OFFSET_Y, imm:$OFFSET_Z, imm:$RESOURCE_ID,
+      imm:$SAMPLER_ID, imm:$textureTarget))]
 > {
-let AsmString = "TEX_LD $dst, $src0, $src1, $src2, $src3, $resourceId, $samplerId, $textureTarget";
-let InOperandList = (ins R600_Reg128:$src0, i32imm:$src1, i32imm:$src2, i32imm:$src3, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget);
+let AsmString = "TEX_LD $DST_GPR, $SRC_GPR, $OFFSET_X, $OFFSET_Y, $OFFSET_Z,"
+    "$RESOURCE_ID, $SAMPLER_ID, $textureTarget";
+let InOperandList = (ins R600_Reg128:$SRC_GPR, i32imm:$OFFSET_X,
+    i32imm:$OFFSET_Y, i32imm:$OFFSET_Z, i32imm:$RESOURCE_ID, i32imm:$SAMPLER_ID,
+    i32imm:$textureTarget);
 }
 
 def TEX_GET_TEXTURE_RESINFO : R600_TEX <
   0x04, "TEX_GET_TEXTURE_RESINFO",
-  [(set R600_Reg128:$dst, (int_AMDGPU_txq R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+  [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txq R600_Reg128:$SRC_GPR,
+      imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
 >;
 
 def TEX_GET_GRADIENTS_H : R600_TEX <
   0x07, "TEX_GET_GRADIENTS_H",
-  [(set R600_Reg128:$dst, (int_AMDGPU_ddx R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+  [(set R600_Reg128:$DST_GPR, (int_AMDGPU_ddx R600_Reg128:$SRC_GPR,
+      imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
 >;
 
 def TEX_GET_GRADIENTS_V : R600_TEX <
   0x08, "TEX_GET_GRADIENTS_V",
-  [(set R600_Reg128:$dst, (int_AMDGPU_ddy R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+  [(set R600_Reg128:$DST_GPR, (int_AMDGPU_ddy R600_Reg128:$SRC_GPR,
+      imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
 >;
 
 def TEX_SET_GRADIENTS_H : R600_TEX <
@@ -900,32 +1002,38 @@ def TEX_SET_GRADIENTS_V : R600_TEX <
 
 def TEX_SAMPLE : R600_TEX <
   0x10, "TEX_SAMPLE",
-  [(set R600_Reg128:$dst, (int_AMDGPU_tex R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+  [(set R600_Reg128:$DST_GPR, (int_AMDGPU_tex R600_Reg128:$SRC_GPR,
+      imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
 >;
 
 def TEX_SAMPLE_C : R600_TEX <
   0x18, "TEX_SAMPLE_C",
-  [(set R600_Reg128:$dst, (int_AMDGPU_tex R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
+  [(set R600_Reg128:$DST_GPR, (int_AMDGPU_tex R600_Reg128:$SRC_GPR,
+      imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
 >;
 
 def TEX_SAMPLE_L : R600_TEX <
   0x11, "TEX_SAMPLE_L",
-  [(set R600_Reg128:$dst, (int_AMDGPU_txl R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+  [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txl R600_Reg128:$SRC_GPR,
+      imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
 >;
 
 def TEX_SAMPLE_C_L : R600_TEX <
   0x19, "TEX_SAMPLE_C_L",
-  [(set R600_Reg128:$dst, (int_AMDGPU_txl R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
+  [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txl R600_Reg128:$SRC_GPR,
+      imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
 >;
 
 def TEX_SAMPLE_LB : R600_TEX <
   0x12, "TEX_SAMPLE_LB",
-  [(set R600_Reg128:$dst, (int_AMDGPU_txb R600_Reg128:$src0,imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+  [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txb R600_Reg128:$SRC_GPR,
+      imm:$RESOURCE_ID, imm:$SAMPLER_ID, imm:$textureTarget))]
 >;
 
 def TEX_SAMPLE_C_LB : R600_TEX <
   0x1A, "TEX_SAMPLE_C_LB",
-  [(set R600_Reg128:$dst, (int_AMDGPU_txb R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
+  [(set R600_Reg128:$DST_GPR, (int_AMDGPU_txb R600_Reg128:$SRC_GPR,
+      imm:$RESOURCE_ID, imm:$SAMPLER_ID, TEX_SHADOW:$textureTarget))]
 >;
 
 def TEX_SAMPLE_G : R600_TEX <