: AMDGPUInst <outs, ins, asm, pattern> {
field bits<32> Inst;
- bit Trig = 0;
+ bit Trig = 0;
bit Op3 = 0;
+ bit isVector = 0;
+ bits<2> FlagOperandIdx = 0;
let Inst = inst;
- let Namespace = "AMDIL";
+ let Namespace = "AMDGPU";
let OutOperandList = outs;
let InOperandList = ins;
let AsmString = asm;
let TSFlags{4} = Trig;
let TSFlags{5} = Op3;
+
+ // Vector instructions are instructions that must fill all slots in an
+ // instruction group
+ let TSFlags{6} = isVector;
+ let TSFlags{8-7} = FlagOperandIdx;
}
class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
{
field bits<64> Inst;
- let Namespace = "AMDIL";
+ let Namespace = "AMDGPU";
+}
+
+def MEMxi : Operand<iPTR> {
+ let MIOperandInfo = (ops R600_TReg32_X:$ptr, i32imm:$index);
}
-def MEMri : Operand<iPTRAny> {
+def MEMrr : Operand<iPTR> {
let MIOperandInfo = (ops R600_Reg32:$ptr, R600_Reg32:$index);
}
def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
+def ADDRDWord : ComplexPattern<i32, 1, "SelectADDRDWord", [], []>;
+def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
class R600_ALU {
}
+def R600_Pred : PredicateOperand<i32, (ops R600_Predicate),
+ (ops PRED_SEL_OFF)>;
+
class R600_1OP <bits<32> inst, string opName, list<dag> pattern,
InstrItinClass itin = AnyALU> :
InstR600 <inst,
(outs R600_Reg32:$dst),
- (ins R600_Reg32:$src, variable_ops),
- !strconcat(opName, " $dst, $src"),
+ (ins R600_Reg32:$src, R600_Pred:$p, variable_ops),
+ !strconcat(opName, " $dst, $src ($p)"),
pattern,
itin
>;
InstrItinClass itin = AnyALU> :
InstR600 <inst,
(outs R600_Reg32:$dst),
- (ins R600_Reg32:$src0, R600_Reg32:$src1, variable_ops),
+ (ins R600_Reg32:$src0, R600_Reg32:$src1,R600_Pred:$p, variable_ops),
!strconcat(opName, " $dst, $src0, $src1"),
pattern,
itin
InstrItinClass itin = AnyALU> :
InstR600 <inst,
(outs R600_Reg32:$dst),
- (ins R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2, variable_ops),
+ (ins R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2,R600_Pred:$p, variable_ops),
!strconcat(opName, " $dst, $src0, $src1, $src2"),
pattern,
itin>{
let Op3 = 1;
}
+
+
+def PRED_X : InstR600 <0, (outs R600_Predicate_Bit:$dst),
+ (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
+ "PRED $dst, $src0, $src1",
+ [], NullALU>
+{
+ let DisableEncoding = "$src0";
+ field bits<32> Inst;
+ bits<32> src1;
+
+ let Inst = src1;
+ let FlagOperandIdx = 3;
+}
+
+let isTerminator = 1, isBranch = 1 in {
+def JUMP : InstR600 <0x10,
+ (outs),
+ (ins brtarget:$target, R600_Pred:$p),
+ "JUMP $target ($p)",
+ [], AnyALU
+ >;
+}
+
class R600_REDUCTION <bits<32> inst, dag ins, string asm, list<dag> pattern,
InstrItinClass itin = VecALU> :
InstR600 <inst,
}]
>;
-def COND_EQ : PatLeaf <
- (cond),
- [{switch(N->get()){{default: return false;
- case ISD::SETOEQ: case ISD::SETUEQ:
- case ISD::SETEQ: return true;}}}]
->;
-
-def COND_NE : PatLeaf <
- (cond),
- [{switch(N->get()){{default: return false;
- case ISD::SETONE: case ISD::SETUNE:
- case ISD::SETNE: return true;}}}]
->;
-def COND_GT : PatLeaf <
- (cond),
- [{switch(N->get()){{default: return false;
- case ISD::SETOGT: case ISD::SETUGT:
- case ISD::SETGT: return true;}}}]
->;
-
-def COND_GE : PatLeaf <
- (cond),
- [{switch(N->get()){{default: return false;
- case ISD::SETOGE: case ISD::SETUGE:
- case ISD::SETGE: return true;}}}]
->;
-
-def COND_LT : PatLeaf <
- (cond),
- [{switch(N->get()){{default: return false;
- case ISD::SETOLT: case ISD::SETULT:
- case ISD::SETLT: return true;}}}]
->;
-
-def COND_LE : PatLeaf <
- (cond),
- [{switch(N->get()){{default: return false;
- case ISD::SETOLE: case ISD::SETULE:
- case ISD::SETLE: return true;}}}]
->;
-
-class EG_CF_RAT <bits <8> cf_inst, bits <6> rat_inst, dag outs, dag ins,
- string asm> :
- InstR600ISA <outs, ins, asm, []>
+class EG_CF_RAT <bits <8> cf_inst, bits <6> rat_inst, bits<4> rat_id, dag outs,
+ dag ins, string asm, list<dag> pattern> :
+ InstR600ISA <outs, ins, asm, pattern>
{
bits<7> RW_GPR;
bits<7> INDEX_GPR;
- bits<4> RAT_ID;
bits<2> RIM;
bits<2> TYPE;
bits<1> BARRIER;
/* CF_ALLOC_EXPORT_WORD0_RAT */
- let Inst{3-0} = RAT_ID;
+ let Inst{3-0} = rat_id;
let Inst{9-4} = rat_inst;
let Inst{10} = 0; /* Reserved */
let Inst{12-11} = RIM;
def load_param : PatFrag<(ops node:$ptr),
(load node:$ptr),
[{
- return true;
const Value *Src = cast<LoadSDNode>(N)->getSrcValue();
if (Src) {
PointerType * PT = dyn_cast<PointerType>(Src->getType());
- return PT && PT->getAddressSpace() == AMDILAS::PARAM_I_ADDRESS;
+ return PT && PT->getAddressSpace() == AMDGPUAS::PARAM_I_ADDRESS;
}
return false;
}]>;
//}
*/
def isR600 : Predicate<"Subtarget.device()"
- "->getGeneration() == AMDILDeviceInfo::HD4XXX">;
+ "->getGeneration() == AMDGPUDeviceInfo::HD4XXX">;
+def isR700 : Predicate<"Subtarget.device()"
+ "->getGeneration() == AMDGPUDeviceInfo::HD4XXX &&"
+ "Subtarget.device()->getDeviceFlag()"
+ ">= OCL_DEVICE_RV710">;
def isEG : Predicate<"Subtarget.device()"
- "->getGeneration() >= AMDILDeviceInfo::HD5XXX && "
+ "->getGeneration() >= AMDGPUDeviceInfo::HD5XXX && "
"Subtarget.device()->getDeviceFlag() != OCL_DEVICE_CAYMAN">;
def isCayman : Predicate<"Subtarget.device()"
"->getDeviceFlag() == OCL_DEVICE_CAYMAN">;
def isEGorCayman : Predicate<"Subtarget.device()"
- "->getGeneration() == AMDILDeviceInfo::HD5XXX"
+ "->getGeneration() == AMDGPUDeviceInfo::HD5XXX"
"|| Subtarget.device()->getGeneration() =="
- "AMDILDeviceInfo::HD6XXX">;
+ "AMDGPUDeviceInfo::HD6XXX">;
def isR600toCayman : Predicate<
- "Subtarget.device()->getGeneration() <= AMDILDeviceInfo::HD6XXX">;
+ "Subtarget.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX">;
let Predicates = [isR600toCayman] in {
[(set R600_Reg32:$dst, (int_AMDGPU_floor R600_Reg32:$src))]
>;
-def MOV : R600_1OP <0x19, "MOV", []>;
+def MOV : InstR600 <0x19, (outs R600_Reg32:$dst),
+ (ins R600_Reg32:$src0, i32imm:$flags,
+ R600_Pred:$p),
+ "MOV $dst, $src0", [], AnyALU> {
+ let FlagOperandIdx = 2;
+}
-def KILLGT : R600_2OP <
- 0x2D, "KILLGT",
- []
+class MOV_IMM <ValueType vt, Operand immType> : InstR600 <0x19,
+ (outs R600_Reg32:$dst),
+ (ins R600_Reg32:$alu_literal, R600_Pred:$p, immType:$imm),
+ "MOV_IMM $dst, $imm",
+ [], AnyALU
+>;
+
+def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
+def : Pat <
+ (imm:$val),
+ (MOV_IMM_I32 (i32 ALU_LITERAL_X), imm:$val)
+>;
+
+def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
+def : Pat <
+ (fpimm:$val),
+ (MOV_IMM_F32 (i32 ALU_LITERAL_X), fpimm:$val)
>;
+def KILLGT : InstR600 <0x2D,
+ (outs R600_Reg32:$dst),
+ (ins R600_Reg32:$src0, R600_Reg32:$src1, i32imm:$flags, R600_Pred:$p,
+ variable_ops),
+ "KILLGT $dst, $src0, $src1, $flags ($p)",
+ [],
+ NullALU>{
+ let FlagOperandIdx = 3;
+}
+
def AND_INT : R600_2OP <
0x30, "AND_INT",
[(set R600_Reg32:$dst, (and R600_Reg32:$src0, R600_Reg32:$src1))]
def CNDE_INT : R600_3OP <
0x1C, "CNDE_INT",
[(set (i32 R600_Reg32:$dst),
- (IL_cmov_logical R600_Reg32:$src0, R600_Reg32:$src2, R600_Reg32:$src1))]
+ (select R600_Reg32:$src0, R600_Reg32:$src2, R600_Reg32:$src1))]
>;
/* Texture instructions */
[]
>;
-def KILP : Pat <
- (int_AMDGPU_kilp),
- (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
->;
-
-def KIL : Pat <
- (int_AMDGPU_kill R600_Reg32:$src0),
- (MASK_WRITE (KILLGT (f32 ZERO), (f32 R600_Reg32:$src0)))
->;
-
/* Helper classes for common instructions */
class MUL_LIT_Common <bits<32> inst> : R600_3OP <
class CNDE_Common <bits<32> inst> : R600_3OP <
inst, "CNDE",
[(set (f32 R600_Reg32:$dst),
- (IL_cmov_logical R600_Reg32:$src0, R600_Reg32:$src2, R600_Reg32:$src1))]
+ (select (i32 (fp_to_sint (fneg R600_Reg32:$src0))), (f32 R600_Reg32:$src2), (f32 R600_Reg32:$src1)))]
>;
class CNDGT_Common <bits<32> inst> : R600_3OP <
class DOT4_Common <bits<32> inst> : R600_REDUCTION <
inst,
- (ins R600_Reg128:$src0, R600_Reg128:$src1),
+ (ins R600_Reg128:$src0, R600_Reg128:$src1, i32imm:$flags),
"DOT4 $dst $src0, $src1",
- [(set R600_Reg32:$dst, (int_AMDGPU_dp4 R600_Reg128:$src0, R600_Reg128:$src1))]
->;
+ []
+ > {
+ let FlagOperandIdx = 3;
+}
-class CUBE_Common <bits<32> inst> : InstR600 <
- inst,
- (outs R600_Reg128:$dst),
- (ins R600_Reg128:$src),
- "CUBE $dst $src",
- [(set R600_Reg128:$dst, (int_AMDGPU_cube R600_Reg128:$src))],
- VecALU
+class DOT4_Pat <Instruction dot4> : Pat <
+ (int_AMDGPU_dp4 R600_Reg128:$src0, R600_Reg128:$src1),
+ (dot4 R600_Reg128:$src0, R600_Reg128:$src1, 0)
>;
+multiclass CUBE_Common <bits<32> inst> {
+
+ def _pseudo : InstR600 <
+ inst,
+ (outs R600_Reg128:$dst),
+ (ins R600_Reg128:$src),
+ "CUBE $dst $src",
+ [(set R600_Reg128:$dst, (int_AMDGPU_cube R600_Reg128:$src))],
+ VecALU
+ >;
+
+ def _real : InstR600 <
+ inst,
+ (outs R600_Reg32:$dst),
+ (ins R600_Reg32:$src0, R600_Reg32:$src1, i32imm:$flags),
+ "CUBE $dst, $src0, $src1",
+ [], VecALU
+ >{
+ let FlagOperandIdx = 3;
+ }
+}
+
class EXP_IEEE_Common <bits<32> inst> : R600_1OP <
inst, "EXP_IEEE",
[(set R600_Reg32:$dst, (fexp2 R600_Reg32:$src))]
>;
class SIN_Common <bits<32> inst> : R600_1OP <
- inst, "SIN",
- [(set R600_Reg32:$dst, (int_AMDIL_sin R600_Reg32:$src))]>{
+ inst, "SIN", []>{
let Trig = 1;
}
class COS_Common <bits<32> inst> : R600_1OP <
- inst, "COS",
- [(set R600_Reg32:$dst, (int_AMDIL_cos R600_Reg32:$src))]> {
+ inst, "COS", []> {
let Trig = 1;
}
def CNDGT_r600 : CNDGT_Common<0x19>;
def CNDGE_r600 : CNDGE_Common<0x1A>;
def DOT4_r600 : DOT4_Common<0x50>;
- def CUBE_r600 : CUBE_Common<0x52>;
+ def : DOT4_Pat <DOT4_r600>;
+ defm CUBE_r600 : CUBE_Common<0x52>;
def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
}
-/* ----------------- */
-/* R700+ Trig helper */
-/* ----------------- */
-
-/*
-class TRIG_HELPER_r700 <InstR600 trig_inst>: Pat <
- (trig_inst R600_Reg32:$src),
- (trig_inst (fmul R600_Reg32:$src, (PI))))
+// Helper pattern for normalizing inputs to triginomic instructions for R700+
+// cards.
+class TRIG_eg <InstR600 trig, Intrinsic intr> : Pat<
+ (intr R600_Reg32:$src),
+ (trig (MUL (MOV_IMM_I32 (i32 ALU_LITERAL_X), CONST.TWO_PI_INV), R600_Reg32:$src))
>;
-*/
-/* ---------------------- */
-/* Evergreen Instructions */
-/* ---------------------- */
+//===----------------------------------------------------------------------===//
+// R700 Only instructions
+//===----------------------------------------------------------------------===//
+let Predicates = [isR700] in {
+ def SIN_r700 : SIN_Common<0x6E>;
+ def COS_r700 : COS_Common<0x6F>;
+
+ // R700 normalizes inputs to SIN/COS the same as EG
+ def : TRIG_eg <SIN_r700, int_AMDGPU_sin>;
+ def : TRIG_eg <COS_r700, int_AMDGPU_cos>;
+}
+
+//===----------------------------------------------------------------------===//
+// Evergreen Only instructions
+//===----------------------------------------------------------------------===//
let Predicates = [isEG] in {
+
+def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
+
+def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
+def MULHI_INT_eg : MULHI_INT_Common<0x90>;
+def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
+def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
+def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
+
+} // End Predicates = [isEG]
+
+/* ------------------------------- */
+/* Evergreen / Cayman Instructions */
+/* ------------------------------- */
+
+let Predicates = [isEGorCayman] in {
+
+ // BFE_UINT - bit_extract, an optimization for mask and shift
+ // Src0 = Input
+ // Src1 = Offset
+ // Src2 = Width
+ //
+ // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
+ //
+ // Example Usage:
+ // (Offset, Width)
+ //
+ // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
+ // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
+ // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
+ // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
+ def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
+ [(set R600_Reg32:$dst, (int_AMDIL_bit_extract_u32 R600_Reg32:$src0,
+ R600_Reg32:$src1,
+ R600_Reg32:$src2))],
+ VecALU
+ >;
+
+ def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT",
+ [(set R600_Reg32:$dst, (AMDGPUbitalign R600_Reg32:$src0, R600_Reg32:$src1,
+ R600_Reg32:$src2))],
+ VecALU
+ >;
+
+ def MULADD_eg : MULADD_Common<0x14>;
+ def ASHR_eg : ASHR_Common<0x15>;
+ def LSHR_eg : LSHR_Common<0x16>;
+ def LSHL_eg : LSHL_Common<0x17>;
+ def CNDE_eg : CNDE_Common<0x19>;
+ def CNDGT_eg : CNDGT_Common<0x1A>;
+ def CNDGE_eg : CNDGE_Common<0x1B>;
+ def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
+ def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
+ def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
+ def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
+ def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
+ def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
+ def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
+ def SIN_eg : SIN_Common<0x8D>;
+ def COS_eg : COS_Common<0x8E>;
+ def DOT4_eg : DOT4_Common<0xBE>;
+ def : DOT4_Pat <DOT4_eg>;
+ defm CUBE_eg : CUBE_Common<0xC0>;
+
+ def DIV_eg : DIV_Common<RECIP_IEEE_eg>;
+ def POW_eg : POW_Common<LOG_IEEE_eg, EXP_IEEE_eg, MUL, GPRF32>;
+ def SSG_eg : SSG_Common<CNDGT_eg, CNDGE_eg>;
+ def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
+
+ def : TRIG_eg <SIN_eg, int_AMDGPU_sin>;
+ def : TRIG_eg <COS_eg, int_AMDGPU_cos>;
-def RAT_WRITE_CACHELESS_eg :
- EG_CF_RAT <0x57, 0x2, (outs), (ins R600_TReg32_X:$rw_gpr,
- R600_TReg32_X:$index_gpr, i32imm:$rat_id), "">
+ def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
+ let Pattern = [];
+ }
+
+ def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
+
+ def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
+ let Pattern = [];
+ }
+
+ def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
+
+ def : Pat<(fp_to_sint R600_Reg32:$src),
+ (FLT_TO_INT_eg (TRUNC R600_Reg32:$src))>;
+
+ def : Pat<(fp_to_uint R600_Reg32:$src),
+ (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src))>;
+
+//===----------------------------------------------------------------------===//
+// Memory read/write instructions
+//===----------------------------------------------------------------------===//
+
+let usesCustomInserter = 1 in {
+
+def RAT_WRITE_CACHELESS_eg : EG_CF_RAT <0x57, 0x2, 0, (outs),
+ (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr),
+ "RAT_WRITE_CACHELESS_eg $rw_gpr, $index_gpr",
+ [(global_store (i32 R600_TReg32_X:$rw_gpr), R600_TReg32_X:$index_gpr)]>
{
let RIM = 0;
/* XXX: Have a separate instruction for non-indexed writes. */
let BARRIER = 1;
}
-def VTX_READ_eg : InstR600ISA < (outs R600_TReg32_X:$dst),
- (ins R600_TReg32_X:$src, i32imm:$buffer_id),
- "VTX_READ_eg $dst, $src", []>
-{
-/*
- bits<7> DST_GPR;
- bits<7> SRC_GPR;
- bits<8> BUFFER_ID;
-*/
- /* If any of these field below need to be calculated at compile time, and
- * a ins operand for them and move them to the list of operands above. */
+} // End usesCustomInserter = 1
- /* XXX: This instruction is manual encoded, so none of these values are used.
- */
-/*
- bits<5> VC_INST = 0; //VC_INST_FETCH
- bits<2> FETCH_TYPE = 2;
- bits<1> FETCH_WHOLE_QUAD = 1;
- bits<1> SRC_REL = 0;
- bits<2> SRC_SEL_X = 0;
- bits<6> MEGA_FETCH_COUNT = 4;
-*/
-/*
+// Floating point global_store
+def : Pat <
+ (global_store (f32 R600_TReg32_X:$val), R600_TReg32_X:$ptr),
+ (RAT_WRITE_CACHELESS_eg R600_TReg32_X:$val, R600_TReg32_X:$ptr)
+>;
- bits<1> DST_REL = 0;
- bits<3> DST_SEL_X = 0;
- bits<3> DST_SEL_Y = 7; //Masked
- bits<3> DST_SEL_Z = 7; //Masked
- bits<3> DST_SEL_W = 7; //Masked
- bits<1> USE_CONST_FIELDS = 1; //Masked
- bits<6> DATA_FORMAT = 0;
- bits<2> NUM_FORMAT_ALL = 0;
- bits<1> FORMAT_COMP_ALL = 0;
- bits<1> SRF_MODE_ALL = 0;
-*/
+class VTX_READ_eg <bits<8> buffer_id, dag outs, list<dag> pattern>
+ : InstR600ISA <outs, (ins MEMxi:$ptr), "VTX_READ_eg $dst, $ptr", pattern> {
-/*
+ // Operands
+ bits<7> DST_GPR;
+ bits<7> SRC_GPR;
+
+ // Static fields
+ bits<5> VC_INST = 0;
+ bits<2> FETCH_TYPE = 2;
+ bits<1> FETCH_WHOLE_QUAD = 0;
+ bits<8> BUFFER_ID = buffer_id;
+ bits<1> SRC_REL = 0;
+ // XXX: We can infer this field based on the SRC_GPR. This would allow us
+ // to store vertex addresses in any channel, not just X.
+ bits<2> SRC_SEL_X = 0;
+ bits<6> MEGA_FETCH_COUNT;
+ bits<1> DST_REL = 0;
+ bits<3> DST_SEL_X;
+ bits<3> DST_SEL_Y;
+ bits<3> DST_SEL_Z;
+ bits<3> DST_SEL_W;
+ // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
+ // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
+ // however, based on my testing if USE_CONST_FIELDS is set, then all
+ // these fields need to be set to 0.
+ bits<1> USE_CONST_FIELDS = 0;
+ bits<6> DATA_FORMAT;
+ bits<2> NUM_FORMAT_ALL = 1;
+ bits<1> FORMAT_COMP_ALL = 0;
+ bits<1> SRF_MODE_ALL = 0;
+
+ // LLVM can only encode 64-bit instructions, so these fields are manually
+ // encoded in R600CodeEmitter
+ //
+ // bits<16> OFFSET;
+ // bits<2> ENDIAN_SWAP = 0;
+ // bits<1> CONST_BUF_NO_STRIDE = 0;
+ // bits<1> MEGA_FETCH = 0;
+ // bits<1> ALT_CONST = 0;
+ // bits<2> BUFFER_INDEX_MODE = 0;
+
+ // VTX_WORD0
let Inst{4-0} = VC_INST;
let Inst{6-5} = FETCH_TYPE;
let Inst{7} = FETCH_WHOLE_QUAD;
let Inst{23} = SRC_REL;
let Inst{25-24} = SRC_SEL_X;
let Inst{31-26} = MEGA_FETCH_COUNT;
-*/
- /* DST_GPR is OK to leave uncommented, because LLVM 3.0 only prevents you
- * from statically setting bits > 31. This field will be set by
- * getMachineValueOp which can set bits > 31.
- */
-// let Inst{32-38} = DST_GPR;
-
- /* XXX: Uncomment for LLVM 3.1 which supports 64-bit instructions */
-/*
+ // VTX_WORD1_GPR
+ let Inst{38-32} = DST_GPR;
let Inst{39} = DST_REL;
- let Inst{40} = 0; //Reserved
+ let Inst{40} = 0; // Reserved
let Inst{43-41} = DST_SEL_X;
let Inst{46-44} = DST_SEL_Y;
let Inst{49-47} = DST_SEL_Z;
let Inst{61-60} = NUM_FORMAT_ALL;
let Inst{62} = FORMAT_COMP_ALL;
let Inst{63} = SRF_MODE_ALL;
-*/
+
+ // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
+ // is done in R600CodeEmitter
+ //
+ // Inst{79-64} = OFFSET;
+ // Inst{81-80} = ENDIAN_SWAP;
+ // Inst{82} = CONST_BUF_NO_STRIDE;
+ // Inst{83} = MEGA_FETCH;
+ // Inst{84} = ALT_CONST;
+ // Inst{86-85} = BUFFER_INDEX_MODE;
+ // Inst{95-86} = 0; Reserved
+
+ // VTX_WORD3 (Padding)
+ //
+ // Inst{127-96} = 0;
}
-/* XXX: Need to convert PTR to rat_id */
-/*
-def : Pat <(store_global (f32 R600_Reg32:$value), node:$ptr),
- (RAT_WRITE_CACHELESS_eg (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)),
- (f32 R600_Reg32:$value),
- sel_x),
- (f32 ZERO), 0, R600_Reg32:$ptr)>;
-*/
+class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
+ : VTX_READ_eg <buffer_id, (outs R600_TReg32_X:$dst), pattern> {
+
+ let MEGA_FETCH_COUNT = 4;
+ let DST_SEL_X = 0;
+ let DST_SEL_Y = 7; // Masked
+ let DST_SEL_Z = 7; // Masked
+ let DST_SEL_W = 7; // Masked
+ let DATA_FORMAT = 0xD; // COLOR_32
+
+ // This is not really necessary, but there were some GPU hangs that appeared
+ // to be caused by ALU instructions in the next instruction group that wrote
+ // to the $ptr registers of the VTX_READ.
+ // e.g.
+ // %T3_X<def> = VTX_READ_PARAM_i32_eg %T2_X<kill>, 24
+ // %T2_X<def> = MOV %ZERO
+ //Adding this constraint prevents this from happening.
+ let Constraints = "$ptr.ptr = $dst";
+}
-class VTX_Param_Read_Pattern <ValueType vt> : Pat <
- (vt (load_param ADDRParam:$mem)),
- (VTX_READ_eg (i32 R600_Reg32:$mem), 0)>;
+class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
+ : VTX_READ_eg <buffer_id, (outs R600_Reg128:$dst), pattern> {
-def : VTX_Param_Read_Pattern <f32>;
-def : VTX_Param_Read_Pattern <i32>;
+ let MEGA_FETCH_COUNT = 16;
+ let DST_SEL_X = 0;
+ let DST_SEL_Y = 1;
+ let DST_SEL_Z = 2;
+ let DST_SEL_W = 3;
+ let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
-} // End isEG Predicate
+ // XXX: Need to force VTX_READ_128 instructions to write to the same register
+ // that holds its buffer address to avoid potential hangs. We can't use
+ // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
+ // registers are different sizes.
+}
-/* ------------------------------- */
-/* Evergreen / Cayman Instructions */
-/* ------------------------------- */
+//===----------------------------------------------------------------------===//
+// VTX Read from parameter memory space
+//===----------------------------------------------------------------------===//
-let Predicates = [isEGorCayman] in {
-
-class TRIG_eg <InstR600 trig, Intrinsic intr> : Pat<
- (intr R600_Reg32:$src),
- (trig (MUL (MOV (LOADCONST_i32 CONST.TWO_PI_INV)), R600_Reg32:$src))
+class VTX_READ_PARAM_32_eg <ValueType vt> : VTX_READ_32_eg <0,
+ [(set (vt R600_TReg32_X:$dst), (load_param ADDRVTX_READ:$ptr))]
>;
- def MULADD_eg : MULADD_Common<0x14>;
- def ASHR_eg : ASHR_Common<0x15>;
- def LSHR_eg : LSHR_Common<0x16>;
- def LSHL_eg : LSHL_Common<0x17>;
- def CNDE_eg : CNDE_Common<0x19>;
- def CNDGT_eg : CNDGT_Common<0x1A>;
- def CNDGE_eg : CNDGE_Common<0x1B>;
- def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
- def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
- def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
- def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
- def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
- def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
- def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
- def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
- def SIN_eg : SIN_Common<0x8D>;
- def COS_eg : COS_Common<0x8E>;
- def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
- def MULHI_INT_eg : MULHI_INT_Common<0x90>;
- def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
- def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
- def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
- def DOT4_eg : DOT4_Common<0xBE>;
- def CUBE_eg : CUBE_Common<0xC0>;
+def VTX_READ_PARAM_i32_eg : VTX_READ_PARAM_32_eg<i32>;
+def VTX_READ_PARAM_f32_eg : VTX_READ_PARAM_32_eg<f32>;
- def DIV_eg : DIV_Common<RECIP_IEEE_eg>;
- def POW_eg : POW_Common<LOG_IEEE_eg, EXP_IEEE_eg, MUL, GPRF32>;
- def SSG_eg : SSG_Common<CNDGT_eg, CNDGE_eg>;
- def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
- def : TRIG_eg <SIN_eg, int_AMDGPU_sin>;
- def : TRIG_eg <COS_eg, int_AMDGPU_cos>;
+//===----------------------------------------------------------------------===//
+// VTX Read from global memory space
+//===----------------------------------------------------------------------===//
- def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
- let Pattern = [];
- }
+// 32-bit reads
- def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
+class VTX_READ_GLOBAL_eg <ValueType vt> : VTX_READ_32_eg <1,
+ [(set (vt R600_TReg32_X:$dst), (global_load ADDRVTX_READ:$ptr))]
+>;
- def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
- let Pattern = [];
- }
+def VTX_READ_GLOBAL_i32_eg : VTX_READ_GLOBAL_eg<i32>;
+def VTX_READ_GLOBAL_f32_eg : VTX_READ_GLOBAL_eg<f32>;
- def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
+// 128-bit reads
- def : Pat<(fp_to_sint R600_Reg32:$src),
- (FLT_TO_INT_eg (TRUNC R600_Reg32:$src))>;
+class VTX_READ_GLOBAL_128_eg <ValueType vt> : VTX_READ_128_eg <1,
+ [(set (vt R600_Reg128:$dst), (global_load ADDRVTX_READ:$ptr))]
+>;
+
+def VTX_READ_GLOBAL_v4i32_eg : VTX_READ_GLOBAL_128_eg<v4i32>;
+def VTX_READ_GLOBAL_v4f32_eg : VTX_READ_GLOBAL_128_eg<v4f32>;
- def : Pat<(fp_to_uint R600_Reg32:$src),
- (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src))>;
}
let Predicates = [isCayman] in {
- /* XXX: I'm not sure if this opcode is correct. */
- def RECIP_UINT_cm : RECIP_UINT_Common<0x77>;
+let isVector = 1 in {
+
+def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
+
+def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
+def MULHI_INT_cm : MULHI_INT_Common<0x90>;
+def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
+def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
+
+} // End isVector = 1
+
+// RECIP_UINT emulation for Cayman
+def : Pat <
+ (AMDGPUurecip R600_Reg32:$src0),
+ (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg R600_Reg32:$src0)),
+ (MOV_IMM_I32 (i32 ALU_LITERAL_X), 0x4f800000)))
+>;
} // End isCayman
[(set R600_TReg32:$dst, (intr))]
>;
-def TGID_X : R600PreloadInst <"TGID_X", int_r600_read_tgid_x>;
-def TGID_Y : R600PreloadInst <"TGID_Y", int_r600_read_tgid_y>;
-def TGID_Z : R600PreloadInst <"TGID_Z", int_r600_read_tgid_z>;
-
-def TIDIG_X : R600PreloadInst <"TIDIG_X", int_r600_read_tidig_x>;
-def TIDIG_Y : R600PreloadInst <"TIDIG_Y", int_r600_read_tidig_y>;
-def TIDIG_Z : R600PreloadInst <"TIDIG_Z", int_r600_read_tidig_z>;
-
-def NGROUPS_X : R600PreloadInst <"NGROUPS_X", int_r600_read_ngroups_x>;
-def NGROUPS_Y : R600PreloadInst <"NGROUPS_Y", int_r600_read_ngroups_y>;
-def NGROUPS_Z : R600PreloadInst <"NGROUPS_Z", int_r600_read_ngroups_z>;
-
-def GLOBAL_SIZE_X : R600PreloadInst <"GLOBAL_SIZE_X",
- int_r600_read_global_size_x>;
-def GLOBAL_SIZE_Y : R600PreloadInst <"GLOBAL_SIZE_Y",
- int_r600_read_global_size_y>;
-def GLOBAL_SIZE_Z : R600PreloadInst <"GLOBAL_SIZE_Z",
- int_r600_read_global_size_z>;
-
-def LOCAL_SIZE_X : R600PreloadInst <"LOCAL_SIZE_X",
- int_r600_read_local_size_x>;
-def LOCAL_SIZE_Y : R600PreloadInst <"LOCAL_SIZE_Y",
- int_r600_read_local_size_y>;
-def LOCAL_SIZE_Z : R600PreloadInst <"LOCAL_SIZE_Z",
- int_r600_read_local_size_z>;
-
def R600_LOAD_CONST : AMDGPUShaderInst <
(outs R600_Reg32:$dst),
(ins i32imm:$src0),
[(set R600_Reg32:$dst, (int_AMDGPU_load_const imm:$src0))]
>;
-def LOAD_INPUT : AMDGPUShaderInst <
- (outs R600_Reg32:$dst),
- (ins i32imm:$src),
- "LOAD_INPUT $dst, $src",
- [(set R600_Reg32:$dst, (int_R600_load_input imm:$src))]
->;
-
def RESERVE_REG : AMDGPUShaderInst <
(outs),
(ins i32imm:$src),
[(int_AMDGPU_reserve_reg imm:$src)]
>;
-def STORE_OUTPUT: AMDGPUShaderInst <
- (outs),
- (ins R600_Reg32:$src0, i32imm:$src1),
- "STORE_OUTPUT $src0, $src1",
- [(int_AMDGPU_store_output R600_Reg32:$src0, imm:$src1)]
->;
-
def TXD: AMDGPUShaderInst <
(outs R600_Reg128:$dst),
(ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$src3, i32imm:$src4),
def CLAMP_R600 : CLAMP <R600_Reg32>;
def FABS_R600 : FABS<R600_Reg32>;
+def FNEG_R600 : FNEG<R600_Reg32>;
-let isPseudo = 1 in {
+let usesCustomInserter = 1 in {
-def LOAD_VTX : AMDGPUShaderInst <
- (outs R600_Reg32:$dst),
- (ins MEMri:$mem),
- "LOAD_VTX",
- [(set (i32 R600_Reg32:$dst), (load_param ADDRParam:$mem))]
+def MASK_WRITE : AMDGPUShaderInst <
+ (outs),
+ (ins R600_Reg32:$src),
+ "MASK_WRITE $src",
+ []
>;
-
-} //End isPseudo
+} // End usesCustomInserter = 1
//===----------------------------------------------------------------------===//
// ISel Patterns
//===----------------------------------------------------------------------===//
+// KIL Patterns
+def KILP : Pat <
+ (int_AMDGPU_kilp),
+ (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO), 0))
+>;
+
+def KIL : Pat <
+ (int_AMDGPU_kill R600_Reg32:$src0),
+ (MASK_WRITE (KILLGT (f32 ZERO), (f32 R600_Reg32:$src0), 0))
+>;
+
// SGT Reverse args
def : Pat <
(selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LT),
def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 6, sel_z>;
def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 7, sel_w>;
+def : Vector_Build <v4f32, R600_Reg32>;
+def : Vector_Build <v4i32, R600_Reg32>;
+
+// bitconvert patterns
+
+def : BitConvert <i32, f32, R600_Reg32>;
+def : BitConvert <f32, i32, R600_Reg32>;
+def : BitConvert <v4f32, v4i32, R600_Reg128>;
+
} // End isR600toCayman Predicate
projects / mesa.git / blobdiff