[AMDGPU] Assembler: support SDWA and DPP for VOP2b instructions

[SamWot]

Rebases on master

[SamWot]

Updated with changes for SDWA pseudo instructions.

[SamWot]

Ping

[artem.tamazov]

LGTM! Pls fix formatting and go ahead.

lib/Target/AMDGPU/VOP2Instructions.td

251

extra space at the beginning

[vpykhtin]

LGTM.

lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

void AMDGPUAsmParser::cvtDPP(MCInst &Inst, const OperandVector &Operands) {
   const MCInstrDesc &Desc = MII.get(Inst.getOpcode());
   for (unsigned J = 0; J < Desc.getNumDefs(); ++J) {
     ((AMDGPUOperand &)*Operands[I++]).addRegOperands(Inst, 1);
   }
 
   for (unsigned E = Operands.size(); I != E; ++I) {
     AMDGPUOperand &Op = ((AMDGPUOperand &)*Operands[I]);
     // Add the register arguments
-    if (isRegOrImmWithInputMods(Desc, Inst.getNumOperands())) {
+    if (Op.isReg() && Op.Reg.RegNo == AMDGPU::VCC) {
+      // VOP2b (v_add_u32, v_sub_u32 ...) sdwa use "vcc" token.
+      // Skip it.
+      continue;
+    } if (isRegOrImmWithInputMods(Desc, Inst.getNumOperands())) {
       Op.addRegOrImmWithFPInputModsOperands(Inst, 2);
     } else if (Op.isDPPCtrl()) {
       Op.addImmOperands(Inst, 1);
     } else if (Op.isImm()) {
       // Handle optional arguments
       OptionalIdx[Op.getImmTy()] = I;
     } else {
       llvm_unreachable("Invalid operand type");
void AMDGPUAsmParser::cvtSDWA(MCInst &Inst, const OperandVector &Operands,
   const MCInstrDesc &Desc = MII.get(Inst.getOpcode());
   for (unsigned J = 0; J < Desc.getNumDefs(); ++J) {
     ((AMDGPUOperand &)*Operands[I++]).addRegOperands(Inst, 1);
   }
 
   for (unsigned E = Operands.size(); I != E; ++I) {
     AMDGPUOperand &Op = ((AMDGPUOperand &)*Operands[I]);
     // Add the register arguments
-    if (BasicInstType == SIInstrFlags::VOPC &&
+    if ((BasicInstType == SIInstrFlags::VOPC || 
+         BasicInstType == SIInstrFlags::VOP2)&&
         Op.isReg() &&
         Op.Reg.RegNo == AMDGPU::VCC) {
-      // VOPC sdwa use "vcc" token as dst. Skip it.
+      // VOPC and VOP2b (v_add_u32, v_sub_u32 ...) sdwa use "vcc" token as dst.
+      // Skip it.
       continue;
     } else if (isRegOrImmWithInputMods(Desc, Inst.getNumOperands())) {
       Op.addRegOrImmWithInputModsOperands(Inst, 2);
     } else if (Op.isImm()) {
       // Handle optional arguments
       OptionalIdx[Op.getImmTy()] = I;
     } else {
       llvm_unreachable("Invalid operand type");

lib/Target/AMDGPU/SIInstrInfo.td

dag ret = !if(!eq(NumSrcArgs, 0),
                // VOP1 without input operands (V_NOP)
                (ins),
             !if(!eq(NumSrcArgs, 1),
                (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
                     clampmod:$clamp, dst_sel:$dst_sel, dst_unused:$dst_unused,
                     src0_sel:$src0_sel),
             !if(!eq(NumSrcArgs, 2),
                !if(!eq(DstVT.Size, 1),
-                  // VOPC_SDWA with float modifiers
+                  // VOPC_SDWA with modifiers
                   (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
                        Src1Mod:$src1_modifiers, Src1RC:$src1,
                        clampmod:$clamp, src0_sel:$src0_sel, src1_sel:$src1_sel),
-                  // VOP2_SDWA or VOPC_SDWA with float modifiers
+                  // VOP2_SDWA or VOPC_SDWA with modifiers
                   (ins Src0Mod:$src0_modifiers, Src0RC:$src0,
                        Src1Mod:$src1_modifiers, Src1RC:$src1,
                        clampmod:$clamp, dst_sel:$dst_sel, dst_unused:$dst_unused,
                        src0_sel:$src0_sel, src1_sel:$src1_sel)),
             (ins)/* endif */)));
 }
 
 // Outs for DPP and SDWA

lib/Target/AMDGPU/VOP2Instructions.td

multiclass VOP2bInst <string opName,
                       SDPatternOperator node = null_frag,
                       string revOp = opName,
                       bit useSGPRInput = !eq(P.NumSrcArgs, 3)> {
 
   let SchedRW = [Write32Bit, WriteSALU] in {
     let Uses = !if(useSGPRInput, [VCC, EXEC], [EXEC]), Defs = [VCC] in {
       def _e32 : VOP2_Pseudo <opName, P>,
                  Commutable_REV<revOp#"_e32", !eq(revOp, opName)>;
+      
+      def _sdwa : VOP2_SDWA_Pseudo <opName, P>,
+              Commutable_REV<revOp#"_sdwa", !eq(revOp, opName)>;
     }
     def _e64 : VOP3_Pseudo <opName, P, getVOP2Pat64<node, P>.ret>,
                Commutable_REV<revOp#"_e64", !eq(revOp, opName)>;
   }
 }
 
 multiclass VOP2eInst <string opName,
                       VOPProfile P,
def VOP_MAC_F32 : VOP_MAC <f32> {
   // 'not a string initializer' error.
   let Asm64 = getAsm64<1, 2, HasModifiers, f32>.ret;
 }
 
 // Write out to vcc or arbitrary SGPR.
 def VOP2b_I32_I1_I32_I32 : VOPProfile<[i32, i32, i32, untyped]> {
   let Asm32 = "$vdst, vcc, $src0, $src1";
   let Asm64 = "$vdst, $sdst, $src0, $src1";
+  let AsmSDWA = "$vdst, vcc, $src0_modifiers, $src1_modifiers$clamp $dst_sel $dst_unused $src0_sel $src1_sel";
+  let AsmDPP = "$vdst, vcc, $src0, $src1 $dpp_ctrl$row_mask$bank_mask$bound_ctrl";
   let Outs32 = (outs DstRC:$vdst);
   let Outs64 = (outs DstRC:$vdst, SReg_64:$sdst);
 }
 
 // Write out to vcc or arbitrary SGPR and read in from vcc or
 // arbitrary SGPR.
 def VOP2b_I32_I1_I32_I32_I1 : VOPProfile<[i32, i32, i32, i1]> {
   // We use VCSrc_b32 to exclude literal constants, even though the
   // encoding normally allows them since the implicit VCC use means
   // using one would always violate the constant bus
   // restriction. SGPRs are still allowed because it should
   // technically be possible to use VCC again as src0.
   let Src0RC32 = VCSrc_b32;
   let Asm32 = "$vdst, vcc, $src0, $src1, vcc";
   let Asm64 = "$vdst, $sdst, $src0, $src1, $src2";
+  let AsmSDWA = "$vdst, vcc, $src0_modifiers, $src1_modifiers, vcc $clamp $dst_sel $dst_unused $src0_sel $src1_sel";
+  let AsmDPP = "$vdst, vcc, $src0, $src1, vcc $dpp_ctrl$row_mask$bank_mask$bound_ctrl";
   let Outs32 = (outs DstRC:$vdst);
   let Outs64 = (outs DstRC:$vdst, SReg_64:$sdst);
 
   // Suppress src2 implied by type since the 32-bit encoding uses an
   // implicit VCC use.
   let Ins32 = (ins Src0RC32:$src0, Src1RC32:$src1);
+
+  let InsSDWA = (ins Src0Mod:$src0_modifiers, Src0SDWA:$src0,
+                     Src1Mod:$src1_modifiers, Src1SDWA:$src1,
+                     clampmod:$clamp, dst_sel:$dst_sel, dst_unused:$dst_unused,
+                     src0_sel:$src0_sel, src1_sel:$src1_sel);
+
+   let InsDPP = (ins Src0Mod:$src0_modifiers, Src0DPP:$src0,
+                     Src1Mod:$src1_modifiers, Src1DPP:$src1,
+                     dpp_ctrl:$dpp_ctrl, row_mask:$row_mask,
+                     bank_mask:$bank_mask, bound_ctrl:$bound_ctrl);
+  let HasExt = 1;
 }
 
 // Read in from vcc or arbitrary SGPR
 def VOP2e_I32_I32_I32_I1 : VOPProfile<[i32, i32, i32, i1]> {
   let Src0RC32 = VCSrc_b32; // See comment in def VOP2b_I32_I1_I32_I32_I1 above.
   let Asm32 = "$vdst, $src0, $src1, vcc";
   let Asm64 = "$vdst, $src0, $src1, $src2";
   let Outs32 = (outs DstRC:$vdst);

 }
 
 multiclass VOP2_Real_e64_vi <bits<10> op> {
   def _e64_vi :
     VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
     VOP3e_vi <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
 }
 
-multiclass VOP2be_Real_e32e64_vi <bits<6> op> : VOP2_Real_e32_vi<op> {
+multiclass Base_VOP2be_Real_e32e64_vi <bits<6> op> : VOP2_Real_e32_vi<op> {
   def _e64_vi :
     VOP3_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
     VOP3be_vi <{0, 1, 0, 0, op{5-0}}, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl>;
 }
 
 multiclass Base_VOP2_Real_e32e64_vi <bits<6> op> :
   VOP2_Real_e32_vi<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 VOP2be_Real_e32e64_vi <bits<6> op> :
+  Base_VOP2be_Real_e32e64_vi<op>, VOP2_SDWA_Real<op> {
+  // For now left dpp only for asm/dasm
+  // TODO: add corresponding pseudo
+  def _dpp : VOP2_DPP<op, !cast<VOP2_Pseudo>(NAME#"_e32")>;
+}
+
 multiclass VOP2_Real_e32e64_vi <bits<6> op> :
   Base_VOP2_Real_e32e64_vi<op>, VOP2_SDWA_Real<op> {
   // For now left dpp only for asm/dasm
   // TODO: add corresponding pseudo
   def _dpp : VOP2_DPP<op, !cast<VOP2_Pseudo>(NAME#"_e32")>;
 }
 
 defm V_CNDMASK_B32        : Base_VOP2_Real_e32e64_vi <0x0>;

test/MC/AMDGPU/vop_dpp.s

 
 // NOSICI: error:
 // VI: v_min_i16_dpp v1, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0 ; encoding: [0xfa,0x06,0x02,0x64,0x02,0x01,0x09,0xa1]
 v_min_i16 v1, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
 
 // NOSICI: error:
 // VI: v_ldexp_f16_dpp v1, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0 ; encoding: [0xfa,0x06,0x02,0x66,0x02,0x01,0x09,0xa1]
 v_ldexp_f16 v1, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
+
+// NOSICI: error:
+// VI: v_add_i32_dpp v1, vcc, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0 ; encoding: [0xfa,0x06,0x02,0x32,0x02,0x01,0x09,0xa1]
+v_add_i32 v1, vcc, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
+
+// NOSICI: error:
+// VI: v_sub_i32_dpp v1, vcc, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0 ; encoding: [0xfa,0x06,0x02,0x34,0x02,0x01,0x09,0xa1]
+v_sub_i32 v1, vcc, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
+
+// NOSICI: error:
+// VI: v_subrev_i32_dpp v1, vcc, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0 ; encoding: [0xfa,0x06,0x02,0x36,0x02,0x01,0x09,0xa1]
+v_subrev_i32 v1, vcc, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
+
+// NOSICI: error:
+// VI: v_addc_u32_dpp v1, vcc, v2, v3, vcc row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0 ; encoding: [0xfa,0x06,0x02,0x38,0x02,0x01,0x09,0xa1]
+v_addc_u32 v1, vcc, v2, v3, vcc row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
+
+// NOSICI: error:
+// VI: v_subb_u32_dpp v1, vcc, v2, v3, vcc row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0 ; encoding: [0xfa,0x06,0x02,0x3a,0x02,0x01,0x09,0xa1]
+v_subb_u32 v1, vcc, v2, v3, vcc row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
+
+// NOSICI: error:
+// VI: v_subbrev_u32_dpp v1, vcc, v2, v3, vcc row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0 ; encoding: [0xfa,0x06,0x02,0x3c,0x02,0x01,0x09,0xa1]
+v_subbrev_u32 v1, vcc, v2, v3, vcc row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0

test/MC/AMDGPU/vop_sdwa.s

 // RUN: not llvm-mc -arch=amdgcn -mcpu=tonga -show-encoding %s | FileCheck %s --check-prefix=GCN --check-prefix=CIVI --check-prefix=VI
 
 // RUN: not llvm-mc -arch=amdgcn -show-encoding %s 2>&1 | FileCheck %s --check-prefix=NOSI --check-prefix=NOSICI
 // RUN: not llvm-mc -arch=amdgcn -mcpu=SI -show-encoding %s 2>&1 | FileCheck %s --check-prefix=NOSI --check-prefix=NOSICI
 // RUN: not llvm-mc -arch=amdgcn -mcpu=bonaire -show-encoding %s 2>&1 | FileCheck %s --check-prefix=NOSICI
 // RUN: not llvm-mc -arch=amdgcn -mcpu=tonga -show-encoding %s 2>&1 | FileCheck %s --check-prefix=NOVI
 
-// ToDo: VOP2b (see vop_dpp.s)
 // ToDo: intrinsics
 
 //---------------------------------------------------------------------------//
 // Check SDWA operands
 //---------------------------------------------------------------------------//
 
 // NOSICI: error:
 // VI: v_mov_b32_sdwa v1, v2 dst_sel:BYTE_0 dst_unused:UNUSED_PRESERVE src0_sel:DWORD ; encoding: [0xf9,0x02,0x02,0x7e,0x02,0x10,0x06,0x06]

 // NOSICI: error:
 // VI: v_min_i16_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2 ; encoding: [0xf9,0x06,0x02,0x64,0x02,0x06,0x05,0x02]
 v_min_i16 v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2
 
 // NOSICI: error:
 // VI: v_ldexp_f16_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2 ; encoding: [0xf9,0x06,0x02,0x66,0x02,0x06,0x05,0x02]
 v_ldexp_f16 v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2
 
+// NOSICI: error:
+// VI: v_add_i32_sdwa v1, vcc, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2 ; encoding: [0xf9,0x06,0x02,0x32,0x02,0x06,0x05,0x02]
+v_add_i32_sdwa v1, vcc, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2
+
+// NOSICI: error:
+// VI: v_sub_i32_sdwa v1, vcc, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2 ; encoding: [0xf9,0x06,0x02,0x34,0x02,0x06,0x05,0x02]
+v_sub_i32_sdwa v1, vcc, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2
+
+// NOSICI: error:
+// VI: v_subrev_i32_sdwa v1, vcc, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2 ; encoding: [0xf9,0x06,0x02,0x36,0x02,0x06,0x05,0x02]
+v_subrev_i32_sdwa v1, vcc, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2
+
+// NOSICI: error:
+// VI: v_addc_u32_sdwa v1, vcc, v2, v3, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2 ; encoding: [0xf9,0x06,0x02,0x38,0x02,0x06,0x05,0x02]
+v_addc_u32_sdwa v1, vcc, v2, v3, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2
+
+// NOSICI: error:
+// VI: v_subb_u32_sdwa v1, vcc, v2, v3, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2 ; encoding: [0xf9,0x06,0x02,0x3a,0x02,0x06,0x05,0x02]
+v_subb_u32_sdwa v1, vcc, v2, v3, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2
+
+// NOSICI: error:
+// VI: v_subbrev_u32_sdwa v1, vcc, v2, v3, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2 ; encoding: [0xf9,0x06,0x02,0x3c,0x02,0x06,0x05,0x02]
+v_subbrev_u32_sdwa v1, vcc, v2, v3, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:BYTE_2
+
 //===----------------------------------------------------------------------===//
 // Check VOPC opcodes
 //===----------------------------------------------------------------------===//
 
 // NOSICI: error:
 // VI: v_cmp_eq_f32 vcc, v1, v2 src0_sel:BYTE_2 src1_sel:WORD_0 ; encoding: [0xf9,0x04,0x84,0x7c,0x01,0x16,0x02,0x04]
 v_cmp_eq_f32 vcc, v1, v2 src0_sel:BYTE_2 src1_sel:WORD_0