[AMDGPU] Add pseudo SDWA instructions

Summary: This is needed for later SDWA support in CodeGen.

Reviewers: vpykhtin, tstellarAMD

Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, tony-tye

Differential Revision: https://reviews.llvm.org/D27412

llvm-svn: 290338

5 files changed, 159 insertions, 85 deletions

diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index 68ab8d676ed..74b188d6b20 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -1567,8 +1567,8 @@ unsigned AMDGPUAsmParser::checkTargetMatchPredicate(MCInst &Inst) {
       getForcedEncodingSize() != 64)
     return Match_PreferE32;
 
-  if (Inst.getOpcode() == AMDGPU::V_MAC_F32_sdwa ||
-      Inst.getOpcode() == AMDGPU::V_MAC_F16_sdwa) {
+  if (Inst.getOpcode() == AMDGPU::V_MAC_F32_sdwa_vi ||
+      Inst.getOpcode() == AMDGPU::V_MAC_F16_sdwa_vi) {
     // v_mac_f32/16 allow only dst_sel == DWORD;
     auto OpNum =
         AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::dst_sel);
@@ -3445,8 +3445,8 @@ void AMDGPUAsmParser::cvtSDWA(MCInst &Inst, const OperandVector &Operands,
 
   addOptionalImmOperand(Inst, Operands, OptionalIdx, AMDGPUOperand::ImmTyClampSI, 0);
 
-  if (Inst.getOpcode() != AMDGPU::V_NOP_sdwa) {
-    // V_NOP_sdwa has no optional sdwa arguments
+  if (Inst.getOpcode() != AMDGPU::V_NOP_sdwa_vi) {
+    // V_NOP_sdwa_vi has no optional sdwa arguments
     switch (BasicInstType) {
     case SIInstrFlags::VOP1:
       addOptionalImmOperand(Inst, Operands, OptionalIdx, AMDGPUOperand::ImmTySdwaDstSel, 6);
@@ -3473,8 +3473,8 @@ void AMDGPUAsmParser::cvtSDWA(MCInst &Inst, const OperandVector &Operands,
 
   // special case v_mac_{f16, f32}:
   // it has src2 register operand that is tied to dst operand
-  if (Inst.getOpcode() == AMDGPU::V_MAC_F32_sdwa ||
-      Inst.getOpcode() == AMDGPU::V_MAC_F16_sdwa)  {
+  if (Inst.getOpcode() == AMDGPU::V_MAC_F32_sdwa_vi ||
+      Inst.getOpcode() == AMDGPU::V_MAC_F16_sdwa_vi)  {
     auto it = Inst.begin();
     std::advance(
         it, AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::src2));
diff --git a/llvm/lib/Target/AMDGPU/VOP1Instructions.td b/llvm/lib/Target/AMDGPU/VOP1Instructions.td
index 9d876c39674..bff706cdc1d 100644
--- a/llvm/lib/Target/AMDGPU/VOP1Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP1Instructions.td
@@ -21,6 +21,15 @@ class VOP1e <bits<8> op, VOPProfile P> : Enc32 {
   let Inst{31-25} = 0x3f; //encoding
 }
 
+class VOP1_SDWAe <bits<8> op, VOPProfile P> : VOP_SDWAe <P> {
+  bits<8> vdst;
+  
+  let Inst{8-0}   = 0xf9; // sdwa
+  let Inst{16-9}  = op;
+  let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
+  let Inst{31-25} = 0x3f; // encoding
+}
+
 class VOP1_Pseudo <string opName, VOPProfile P, list<dag> pattern=[]> :
   InstSI <P.Outs32, P.Ins32, "", pattern>,
   VOP <opName>,
@@ -68,6 +77,11 @@ class VOP1_Real <VOP1_Pseudo ps, int EncodingFamily> :
   let TSFlags            = ps.TSFlags;
 }
 
+class VOP1_SDWA_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
+  VOP_SDWA_Pseudo <OpName, P, pattern> {
+  let AsmMatchConverter = "cvtSdwaVOP1";
+}
+
 class getVOP1Pat64 <SDPatternOperator node, VOPProfile P> : LetDummies {
   list<dag> ret = !if(P.HasModifiers,
     [(set P.DstVT:$vdst, (node (P.Src0VT (VOP3Mods0 P.Src0VT:$src0,
@@ -79,6 +93,7 @@ multiclass VOP1Inst <string opName, VOPProfile P,
                      SDPatternOperator node = null_frag> {
   def _e32 : VOP1_Pseudo <opName, P>;
   def _e64 : VOP3_Pseudo <opName, P, getVOP1Pat64<node, P>.ret>;
+  def _sdwa : VOP1_SDWA_Pseudo <opName, P>;
 }
 
 //===----------------------------------------------------------------------===//
@@ -422,23 +437,6 @@ defm V_EXP_LEGACY_F32    : VOP1_Real_ci <0x46>;
 // VI
 //===----------------------------------------------------------------------===//
 
-class VOP1_SDWA <bits<8> op, VOP1_Pseudo ps, VOPProfile P = ps.Pfl> :
-  VOP_SDWA <ps.OpName, P> {
-  let Defs = ps.Defs;
-  let Uses = ps.Uses;
-  let SchedRW = ps.SchedRW;
-  let hasSideEffects = ps.hasSideEffects;
-  let Constraints = ps.Constraints;
-  let DisableEncoding = ps.DisableEncoding;
-  let AsmMatchConverter = "cvtSdwaVOP1";
-
-  bits<8> vdst;
-  let Inst{8-0}   = 0xf9; // sdwa
-  let Inst{16-9}  = op;
-  let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
-  let Inst{31-25} = 0x3f; // encoding
-}
-
 class VOP1_DPP <bits<8> op, VOP1_Pseudo ps, VOPProfile P = ps.Pfl> :
   VOP_DPP <ps.OpName, P> {
   let Defs = ps.Defs;
@@ -465,9 +463,12 @@ multiclass VOP1_Real_vi <bits<10> op> {
       VOP3e_vi <!add(0x140, op), !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
   }
 
-  // for now left sdwa/dpp only for asm/dasm
+  def _sdwa_vi :
+    VOP_SDWA_Real <!cast<VOP1_SDWA_Pseudo>(NAME#"_sdwa")>,
+    VOP1_SDWAe <op{7-0}, !cast<VOP1_SDWA_Pseudo>(NAME#"_sdwa").Pfl>;
+
+  // For now left dpp only for asm/dasm
   // TODO: add corresponding pseudo
-  def _sdwa : VOP1_SDWA<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32")>;
   def _dpp : VOP1_DPP<op{7-0}, !cast<VOP1_Pseudo>(NAME#"_e32")>;
 }
 
diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
index a3025ffe2b3..0b18e0021f7 100644
--- a/llvm/lib/Target/AMDGPU/VOP2Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
@@ -37,6 +37,17 @@ class VOP2_MADKe <bits<6> op, VOPProfile P> : Enc64 {
   let Inst{63-32} = imm;
 }
 
+class VOP2_SDWAe <bits<6> op, VOPProfile P> : VOP_SDWAe <P> {
+  bits<8> vdst;
+  bits<8> src1;
+  
+  let Inst{8-0}   = 0xf9; // sdwa
+  let Inst{16-9}  = !if(P.HasSrc1, src1{7-0}, 0);
+  let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
+  let Inst{30-25} = op;
+  let Inst{31}    = 0x0; // encoding
+}
+
 class VOP2_Pseudo <string opName, VOPProfile P, list<dag> pattern=[], string suffix = "_e32"> :
   InstSI <P.Outs32, P.Ins32, "", pattern>,
   VOP <opName>,
@@ -84,6 +95,11 @@ class VOP2_Real <VOP2_Pseudo ps, int EncodingFamily> :
   let TSFlags            = ps.TSFlags;
 }
 
+class VOP2_SDWA_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
+  VOP_SDWA_Pseudo <OpName, P, pattern> {
+  let AsmMatchConverter = "cvtSdwaVOP2";
+}
+
 class getVOP2Pat64 <SDPatternOperator node, VOPProfile P> : LetDummies {
   list<dag> ret = !if(P.HasModifiers,
     [(set P.DstVT:$vdst,
@@ -102,8 +118,12 @@ multiclass VOP2Inst <string opName,
 
   def _e64 : VOP3_Pseudo <opName, P, getVOP2Pat64<node, P>.ret>,
              Commutable_REV<revOp#"_e64", !eq(revOp, opName)>;
+
+  def _sdwa : VOP2_SDWA_Pseudo <opName, P>,
+              Commutable_REV<revOp#"_sdwa", !eq(revOp, opName)>;
 }
 
+// TODO: add SDWA pseudo instructions for VOP2bInst and VOP2eInst
 multiclass VOP2bInst <string opName,
                       VOPProfile P,
                       SDPatternOperator node = null_frag,
@@ -554,25 +574,6 @@ defm V_CVT_PK_I16_I32     : VOP2_Real_e32e64_si <0x31>;
 // VI
 //===----------------------------------------------------------------------===//
 
-class VOP2_SDWA <bits<6> op, VOP2_Pseudo ps, VOPProfile P = ps.Pfl> :
-  VOP_SDWA <ps.OpName, P> {
-  let Defs = ps.Defs;
-  let Uses = ps.Uses;
-  let SchedRW = ps.SchedRW;
-  let hasSideEffects = ps.hasSideEffects;
-  let Constraints = ps.Constraints;
-  let DisableEncoding = ps.DisableEncoding;
-  let AsmMatchConverter = "cvtSdwaVOP2";
-
-  bits<8> vdst;
-  bits<8> src1;
-  let Inst{8-0}   = 0xf9; // sdwa
-  let Inst{16-9}  = !if(P.HasSrc1, src1{7-0}, 0);
-  let Inst{24-17} = !if(P.EmitDst, vdst{7-0}, 0);
-  let Inst{30-25} = op;
-  let Inst{31}    = 0x0; // encoding
-}
-
 class VOP2_DPP <bits<6> op, VOP2_Pseudo ps, VOPProfile P = ps.Pfl> :
   VOP_DPP <ps.OpName, P> {
   let Defs = ps.Defs;
@@ -627,12 +628,17 @@ multiclass Base_VOP2_Real_e32e64_vi <bits<6> op> :
   VOP2_Real_e64_vi<{0, 1, 0, 0, op{5-0}}>;
 
 } // End AssemblerPredicates = [isVI], DecoderNamespace = "VI"
+ 
+multiclass VOP2_SDWA_Real <bits<6> op> {
+  def _sdwa_vi :
+    VOP_SDWA_Real <!cast<VOP2_SDWA_Pseudo>(NAME#"_sdwa")>,
+    VOP2_SDWAe <op{5-0}, !cast<VOP2_SDWA_Pseudo>(NAME#"_sdwa").Pfl>;
+}
 
 multiclass VOP2_Real_e32e64_vi <bits<6> op> :
-  Base_VOP2_Real_e32e64_vi<op> {
-  // for now left sdwa/dpp only for asm/dasm
+  Base_VOP2_Real_e32e64_vi<op>, VOP2_SDWA_Real<op> {
+  // For now left dpp only for asm/dasm
   // TODO: add corresponding pseudo
-  def _sdwa : VOP2_SDWA<op, !cast<VOP2_Pseudo>(NAME#"_e32")>;
   def _dpp : VOP2_DPP<op, !cast<VOP2_Pseudo>(NAME#"_e32")>;
 }
 
diff --git a/llvm/lib/Target/AMDGPU/VOPCInstructions.td b/llvm/lib/Target/AMDGPU/VOPCInstructions.td
index 32137888ef1..f042404acaf 100644
--- a/llvm/lib/Target/AMDGPU/VOPCInstructions.td
+++ b/llvm/lib/Target/AMDGPU/VOPCInstructions.td
@@ -21,6 +21,19 @@ class VOPCe <bits<8> op> : Enc32 {
   let Inst{31-25} = 0x3e;
 }
 
+class VOPC_SDWAe <bits<8> op, VOPProfile P> : VOP_SDWAe <P> {
+  bits<8> src1;
+
+  let Inst{8-0}   = 0xf9; // sdwa
+  let Inst{16-9}  = !if(P.HasSrc1, src1{7-0}, 0);
+  let Inst{24-17} = op;
+  let Inst{31-25} = 0x3e; // encoding
+
+  // VOPC disallows dst_sel and dst_unused as they have no effect on destination
+  let Inst{42-40} = SDWA.DWORD;
+  let Inst{44-43} = SDWA.UNUSED_PRESERVE;
+}
+
 //===----------------------------------------------------------------------===//
 // VOPC classes
 //===----------------------------------------------------------------------===//
@@ -82,6 +95,11 @@ class VOPC_Real <VOPC_Pseudo ps, int EncodingFamily> :
   let TSFlags            = ps.TSFlags;
 }
 
+class VOPC_SDWA_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
+  VOP_SDWA_Pseudo <OpName, P, pattern> {
+  let AsmMatchConverter = "cvtSdwaVOPC";
+}
+
 // This class is used only with VOPC instructions. Use $sdst for out operand
 class VOPCInstAlias <VOP3_Pseudo ps, Instruction inst, VOPProfile p = ps.Pfl> :
   InstAlias <ps.OpName#" "#p.Asm32, (inst)>, PredicateControl {
@@ -131,6 +149,7 @@ multiclass VOPC_Pseudos <string opName,
     let isCompare = 1;
     let isCommutable = 1;
   }
+
   def _e64 : VOP3_Pseudo<opName, P,
     !if(P.HasModifiers,
       [(set i1:$sdst,
@@ -145,6 +164,15 @@ multiclass VOPC_Pseudos <string opName,
     let isCompare = 1;
     let isCommutable = 1;
   }
+
+  def _sdwa : VOPC_SDWA_Pseudo <opName, P>,
+              Commutable_REV<revOp#"_sdwa", !eq(revOp, opName)> {
+    let Defs = !if(DefExec, [VCC, EXEC], [VCC]);
+    let SchedRW = P.Schedule;
+    let isConvergent = DefExec;
+    let isCompare = 1;
+    let isCommutable = 1;
+  }
 }
 
 def VOPC_I1_F16_F16 : VOPC_Profile<[Write32Bit], f16>;
@@ -471,10 +499,17 @@ multiclass VOPC_Class_Pseudos <string opName, VOPC_Profile p, bit DefExec> {
     let SchedRW = p.Schedule;
     let isConvergent = DefExec;
   }
+
   def _e64 : VOP3_Pseudo<opName, p, getVOPCClassPat64<p>.ret> {
     let Defs = !if(DefExec, [EXEC], []);
     let SchedRW = p.Schedule;
   }
+
+  def _sdwa : VOPC_SDWA_Pseudo <opName, p> {
+    let Defs = !if(DefExec, [VCC, EXEC], [VCC]);
+    let SchedRW = p.Schedule;
+    let isConvergent = DefExec;
+  }
 }
 
 def VOPC_I1_F16_I32 : VOPC_Class_Profile<[Write32Bit], f16>;
@@ -822,27 +857,6 @@ defm V_CMPX_CLASS_F64 : VOPC_Real_si <0xb8>;
 // VI
 //===----------------------------------------------------------------------===//
 
-class VOPC_SDWA<bits<8> op, VOPC_Pseudo ps, VOPProfile P = ps.Pfl> :
-  VOP_SDWA <ps.OpName, P> {
-  let Defs = ps.Defs;
-  let hasSideEffects = ps.hasSideEffects;
-  let AsmMatchConverter = "cvtSdwaVOPC";
-  let isCompare = ps.isCompare;
-  let isCommutable = ps.isCommutable;
-  let Constraints = ps.Constraints;
-  let DisableEncoding = ps.DisableEncoding;
-
-  bits<8> src1;
-  let Inst{8-0}   = 0xf9; // sdwa
-  let Inst{16-9}  = !if(P.HasSrc1, src1{7-0}, 0);
-  let Inst{24-17} = op;
-  let Inst{31-25} = 0x3e; // encoding
-
-  // VOPC disallows dst_sel and dst_unused as they have no effect on destination
-  let Inst{42-40} = SDWA_DWORD;
-  let Inst{44-43} = SDWA_UNUSED_PRESERVE;
-}
-
 multiclass VOPC_Real_vi <bits<10> op> {
   let AssemblerPredicates = [isVI], DecoderNamespace = "VI" in {
     def _e32_vi :
@@ -859,9 +873,9 @@ multiclass VOPC_Real_vi <bits<10> op> {
     }
   }
 
-  // for now left sdwa only for asm/dasm
-  // TODO: add corresponding pseudo
-  def _sdwa : VOPC_SDWA<op{7-0}, !cast<VOPC_Pseudo>(NAME#"_e32")>;
+  def _sdwa_vi :
+    VOP_SDWA_Real <!cast<VOPC_SDWA_Pseudo>(NAME#"_sdwa")>,
+    VOPC_SDWAe <op{7-0}, !cast<VOPC_SDWA_Pseudo>(NAME#"_sdwa").Pfl>;
 
   def : VOPCInstAlias <!cast<VOP3_Pseudo>(NAME#"_e64"),
                        !cast<Instruction>(NAME#"_e32_vi")> {
diff --git a/llvm/lib/Target/AMDGPU/VOPInstructions.td b/llvm/lib/Target/AMDGPU/VOPInstructions.td
index 3af16b32695..5f72f97d9e2 100644
--- a/llvm/lib/Target/AMDGPU/VOPInstructions.td
+++ b/llvm/lib/Target/AMDGPU/VOPInstructions.td
@@ -207,6 +207,22 @@ class VOP3be_vi <bits<10> op, VOPProfile P> : VOP3be<P> {
   let Inst{15}    = !if(P.HasClamp, clamp{0}, 0);
 }
 
+def SDWA {
+  // sdwa_sel
+  int BYTE_0 = 0;
+  int BYTE_1 = 1;
+  int BYTE_2 = 2;
+  int BYTE_3 = 3;
+  int WORD_0 = 4;
+  int WORD_1 = 5;
+  int DWORD = 6;
+
+  // dst_unused
+  int UNUSED_PAD = 0;
+  int UNUSED_SEXT = 1;
+  int UNUSED_PRESERVE = 2;
+}
+
 class VOP_SDWAe<VOPProfile P> : Enc64 {
   bits<8> src0;
   bits<3> src0_sel;
@@ -217,37 +233,74 @@ class VOP_SDWAe<VOPProfile P> : Enc64 {
   bits<2> dst_unused;
   bits<1> clamp;
 
-  bits<3> SDWA_DWORD = 6;
-  bits<2> SDWA_UNUSED_PRESERVE = 2;
-
   let Inst{39-32} = !if(P.HasSrc0, src0{7-0}, 0);
-  let Inst{42-40} = !if(P.EmitDst, dst_sel{2-0}, SDWA_DWORD{2-0});
-  let Inst{44-43} = !if(P.EmitDst, dst_unused{1-0}, SDWA_UNUSED_PRESERVE{1-0});
+  let Inst{42-40} = !if(P.EmitDst, dst_sel{2-0}, SDWA.DWORD);
+  let Inst{44-43} = !if(P.EmitDst, dst_unused{1-0}, SDWA.UNUSED_PRESERVE);
   let Inst{45}    = !if(P.HasSDWAClamp, clamp{0}, 0);
-  let Inst{50-48} = !if(P.HasSrc0, src0_sel{2-0}, SDWA_DWORD{2-0});
+  let Inst{50-48} = !if(P.HasSrc0, src0_sel{2-0}, SDWA.DWORD);
   let Inst{53-52} = !if(P.HasSrc0FloatMods, src0_modifiers{1-0}, 0);
   let Inst{51}    = !if(P.HasSrc0IntMods, src0_modifiers{0}, 0);
-  let Inst{58-56} = !if(P.HasSrc1, src1_sel{2-0}, SDWA_DWORD{2-0});
+  let Inst{58-56} = !if(P.HasSrc1, src1_sel{2-0}, SDWA.DWORD);
   let Inst{61-60} = !if(P.HasSrc1FloatMods, src1_modifiers{1-0}, 0);
   let Inst{59}    = !if(P.HasSrc1IntMods, src1_modifiers{0}, 0);
 }
 
-class VOP_SDWA <string OpName, VOPProfile P> :
-  InstSI <P.OutsSDWA, P.InsSDWA, OpName#P.AsmSDWA, []>,
-  VOP_SDWAe<P> {
+class VOP_SDWA_Pseudo <string opName, VOPProfile P, list<dag> pattern=[]> :
+  InstSI <P.OutsSDWA, P.InsSDWA, "", pattern>,
+  VOP <opName>,
+  SIMCInstr <opName#"_sdwa", SIEncodingFamily.NONE>,
+  MnemonicAlias <opName#"_sdwa", opName> {
+  
+  let isPseudo = 1;
+  let isCodeGenOnly = 1;
+  let UseNamedOperandTable = 1;
+
+  string Mnemonic = opName;
+  string AsmOperands = P.AsmSDWA;
+
+  let Size = 8;
   let mayLoad = 0;
   let mayStore = 0;
-  let hasSideEffects = 0;
-  let UseNamedOperandTable = 1;
+  let hasSideEffects = 0;  
+
   let VALU = 1;
   let SDWA = 1;
-  let Size = 8;
-
+  let Uses = [EXEC];
+  
   let SubtargetPredicate = isVI;
   let AssemblerPredicate = !if(P.HasExt, isVI, DisableInst);
   let AsmVariantName = !if(P.HasExt, AMDGPUAsmVariants.SDWA,
                                      AMDGPUAsmVariants.Disable);
   let DecoderNamespace = "SDWA";
+
+  VOPProfile Pfl = P;
+}
+
+class VOP_SDWA_Real <VOP_SDWA_Pseudo ps> :
+  InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []>,
+  SIMCInstr <ps.PseudoInstr, SIEncodingFamily.VI> {
+
+  let isPseudo = 0;
+  let isCodeGenOnly = 0;
+
+  let Defs = ps.Defs;
+  let Uses = ps.Uses;
+  let SchedRW = ps.SchedRW;
+  let hasSideEffects = ps.hasSideEffects;
+
+  let Constraints     = ps.Constraints;
+  let DisableEncoding = ps.DisableEncoding;
+
+  // Copy relevant pseudo op flags
+  let SubtargetPredicate   = ps.SubtargetPredicate;
+  let AssemblerPredicate   = ps.AssemblerPredicate;
+  let AsmMatchConverter    = ps.AsmMatchConverter;
+  let AsmVariantName       = ps.AsmVariantName;
+  let UseNamedOperandTable = ps.UseNamedOperandTable;
+  let DecoderNamespace     = ps.DecoderNamespace;
+  let Constraints          = ps.Constraints;
+  let DisableEncoding      = ps.DisableEncoding;
+  let TSFlags              = ps.TSFlags;
 }
 
 class VOP_DPPe<VOPProfile P> : Enc64 {