//
//===----------------------------------------------------------------------===//
//
-// R600 Tablegen instruction definitions
+// TableGen definitions for instructions which are available on R600 family
+// GPUs.
//
//===----------------------------------------------------------------------===//
def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
def ADDRGA_CONST_OFFSET : ComplexPattern<i32, 1, "SelectGlobalValueConstantOffset", [], []>;
def ADDRGA_VAR_OFFSET : ComplexPattern<i32, 2, "SelectGlobalValueVariableOffset", [], []>;
-def ADDRIndirect : ComplexPattern<iPTR, 2, "SelectADDRIndirect", [], []>;
def R600_Pred : PredicateOperand<i32, (ops R600_Predicate),
let update_pred = 0;
let HasNativeOperands = 1;
let Op1 = 1;
+ let ALUInst = 1;
let DisableEncoding = "$literal";
+ let UseNamedOperandTable = 1;
let Inst{31-0} = Word0;
let Inst{63-32} = Word1;
class R600_1OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
InstrItinClass itin = AnyALU> :
R600_1OP <inst, opName,
- [(set R600_Reg32:$dst, (node R600_Reg32:$src0))]
+ [(set R600_Reg32:$dst, (node R600_Reg32:$src0))], itin
>;
-// If you add our change the operands for R600_2OP instructions, you must
+// If you add or change the operands for R600_2OP instructions, you must
// also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
// R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
let HasNativeOperands = 1;
let Op2 = 1;
+ let ALUInst = 1;
let DisableEncoding = "$literal";
+ let UseNamedOperandTable = 1;
let Inst{31-0} = Word0;
let Inst{63-32} = Word1;
}
class R600_2OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
- InstrItinClass itim = AnyALU> :
+ InstrItinClass itin = AnyALU> :
R600_2OP <inst, opName,
[(set R600_Reg32:$dst, (node R600_Reg32:$src0,
- R600_Reg32:$src1))]
+ R600_Reg32:$src1))], itin
>;
// If you add our change the operands for R600_3OP instructions, you must
let HasNativeOperands = 1;
let DisableEncoding = "$literal";
let Op3 = 1;
+ let UseNamedOperandTable = 1;
+ let ALUInst = 1;
let Inst{31-0} = Word0;
let Inst{63-32} = Word1;
def TEX_SHADOW : PatLeaf<
(imm),
[{uint32_t TType = (uint32_t)N->getZExtValue();
- return (TType >= 6 && TType <= 8) || (TType >= 11 && TType <= 13);
+ return (TType >= 6 && TType <= 8) || TType == 13;
}]
>;
def TEX_ARRAY : PatLeaf<
(imm),
[{uint32_t TType = (uint32_t)N->getZExtValue();
- return TType == 9 || TType == 10 || TType == 15 || TType == 16;
+ return TType == 9 || TType == 10 || TType == 16;
}]
>;
}]
>;
-class EG_CF_RAT <bits <8> cfinst, bits <6> ratinst, bits<4> ratid, dag outs,
- dag ins, string asm, list<dag> pattern> :
+def TEX_MSAA : PatLeaf<
+ (imm),
+ [{uint32_t TType = (uint32_t)N->getZExtValue();
+ return TType == 14;
+ }]
+>;
+
+def TEX_ARRAY_MSAA : PatLeaf<
+ (imm),
+ [{uint32_t TType = (uint32_t)N->getZExtValue();
+ return TType == 15;
+ }]
+>;
+
+class EG_CF_RAT <bits <8> cfinst, bits <6> ratinst, bits<4> ratid, bits<4> mask,
+ dag outs, dag ins, string asm, list<dag> pattern> :
InstR600ISA <outs, ins, asm, pattern>,
CF_ALLOC_EXPORT_WORD0_RAT, CF_ALLOC_EXPORT_WORD1_BUF {
- let cf_inst = cfinst;
- let rat_inst = ratinst;
let rat_id = ratid;
+ let rat_inst = ratinst;
+ let rim = 0;
+ // XXX: Have a separate instruction for non-indexed writes.
+ let type = 1;
+ let rw_rel = 0;
+ let elem_size = 0;
+
+ let array_size = 0;
+ let comp_mask = mask;
+ let burst_count = 0;
+ let vpm = 0;
+ let cf_inst = cfinst;
+ let mark = 0;
+ let barrier = 1;
let Inst{31-0} = Word0;
let Inst{63-32} = Word1;
+ let IsExport = 1;
+
+}
+
+class VTX_READ <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
+ : InstR600ISA <outs, (ins MEMxi:$src_gpr), name, pattern>,
+ VTX_WORD1_GPR {
+
+ // Static fields
+ let DST_REL = 0;
+ // 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.
+ let USE_CONST_FIELDS = 0;
+ let NUM_FORMAT_ALL = 1;
+ let FORMAT_COMP_ALL = 0;
+ let SRF_MODE_ALL = 0;
+
+ let Inst{63-32} = Word1;
+ // 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_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;
+
+ let VTXInst = 1;
}
class LoadParamFrag <PatFrag load_type> : PatFrag <
(ops node:$ptr), (load_type node:$ptr),
- [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
+ [{ return isConstantLoad(dyn_cast<LoadSDNode>(N), 0); }]
>;
def load_param : LoadParamFrag<load>;
-def load_param_zexti8 : LoadParamFrag<zextloadi8>;
-def load_param_zexti16 : LoadParamFrag<zextloadi16>;
+def load_param_exti8 : LoadParamFrag<az_extloadi8>;
+def load_param_exti16 : LoadParamFrag<az_extloadi16>;
def isR600 : Predicate<"Subtarget.getGeneration() <= AMDGPUSubtarget::R700">;
-def isR700 : Predicate<"Subtarget.getGeneration() == AMDGPUSubtarget::R700">;
-def isEG : Predicate<
- "Subtarget.getGeneration() >= AMDGPUSubtarget::EVERGREEN && "
- "Subtarget.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS && "
- "!Subtarget.hasCaymanISA()">;
-
-def isCayman : Predicate<"Subtarget.hasCaymanISA()">;
-def isEGorCayman : Predicate<"Subtarget.getGeneration() == "
- "AMDGPUSubtarget::EVERGREEN"
- "|| Subtarget.getGeneration() =="
- "AMDGPUSubtarget::NORTHERN_ISLANDS">;
def isR600toCayman : Predicate<
"Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS">;
[]
>;
+def COS_HW : SDNode<"AMDGPUISD::COS_HW",
+ SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisFP<1>]>
+>;
+
+def SIN_HW : SDNode<"AMDGPUISD::SIN_HW",
+ SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisFP<1>]>
+>;
+
def TEXTURE_FETCH_Type : SDTypeProfile<1, 19, [SDTCisFP<0>]>;
def TEXTURE_FETCH: SDNode<"AMDGPUISD::TEXTURE_FETCH", TEXTURE_FETCH_Type, []>;
(outs R600_Reg128:$dst),
(ins i32imm:$src0),
"INTERP_LOAD $src0 : $dst",
- []>;
+ [(set R600_Reg128:$dst, (int_R600_interp_const imm:$src0))]>;
def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
let bank_swizzle = 5;
class ExportSwzInst : InstR600ISA<(
outs),
(ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
- i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
+ RSel:$sw_x, RSel:$sw_y, RSel:$sw_z, RSel:$sw_w, i32imm:$inst,
i32imm:$eop),
- !strconcat("EXPORT", " $gpr"),
+ !strconcat("EXPORT", " $gpr.$sw_x$sw_y$sw_z$sw_w"),
[]>, ExportWord0, ExportSwzWord1 {
let elem_size = 3;
let Inst{31-0} = Word0;
let Inst{63-32} = Word1;
+ let IsExport = 1;
}
} // End usesCustomInserter = 1
let elem_size = 0;
let Inst{31-0} = Word0;
let Inst{63-32} = Word1;
+ let IsExport = 1;
}
//===----------------------------------------------------------------------===//
(ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1,
KCACHE:$KCACHE_MODE0, KCACHE:$KCACHE_MODE1,
i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1,
-i32imm:$COUNT),
+i32imm:$COUNT, i32imm:$Enabled),
!strconcat(OpName, " $COUNT, @$ADDR, "
"KC0[$KCACHE_MODE0], KC1[$KCACHE_MODE1]"),
[] >, CF_ALU_WORD0, CF_ALU_WORD1 {
let ALT_CONST = 0;
let WHOLE_QUAD_MODE = 0;
let BARRIER = 1;
+ let UseNamedOperandTable = 1;
let Inst{31-0} = Word0;
let Inst{63-32} = Word1;
class CF_CLAUSE_R600 <bits<7> inst, dag ins, string AsmPrint> : AMDGPUInst <(outs),
ins, AsmPrint, [] >, CF_WORD0_R600, CF_WORD1_R600 {
field bits<64> Inst;
+ bits<4> CNT;
let CF_INST = inst;
let BARRIER = 1;
let CF_CONST = 0;
let VALID_PIXEL_MODE = 0;
let COND = 0;
+ let COUNT = CNT{2-0};
let CALL_COUNT = 0;
- let COUNT_3 = 0;
+ let COUNT_3 = CNT{3};
let END_OF_PROGRAM = 0;
let WHOLE_QUAD_MODE = 0;
def CF_ALU : ALU_CLAUSE<8, "ALU">;
def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
+def CF_ALU_POP_AFTER : ALU_CLAUSE<10, "ALU_POP_AFTER">;
+def CF_ALU_CONTINUE : ALU_CLAUSE<13, "ALU_CONTINUE">;
+def CF_ALU_BREAK : ALU_CLAUSE<14, "ALU_BREAK">;
+def CF_ALU_ELSE_AFTER : ALU_CLAUSE<15, "ALU_ELSE_AFTER">;
def FETCH_CLAUSE : AMDGPUInst <(outs),
(ins i32imm:$addr), "Fetch clause starting at $addr:", [] > {
// XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
def SETE : R600_2OP <
0x08, "SETE",
- [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_EQ))]
+ [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_OEQ))]
>;
def SGT : R600_2OP <
0x09, "SETGT",
- [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GT))]
+ [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_OGT))]
>;
def SGE : R600_2OP <
0xA, "SETGE",
- [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_GE))]
+ [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_OGE))]
>;
def SNE : R600_2OP <
0xB, "SETNE",
- [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_NE))]
+ [(set f32:$dst, (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_UNE))]
>;
def SETE_DX10 : R600_2OP <
0xC, "SETE_DX10",
- [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_EQ))]
+ [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_OEQ))]
>;
def SETGT_DX10 : R600_2OP <
0xD, "SETGT_DX10",
- [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GT))]
+ [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_OGT))]
>;
def SETGE_DX10 : R600_2OP <
0xE, "SETGE_DX10",
- [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_GE))]
+ [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_OGE))]
>;
def SETNE_DX10 : R600_2OP <
0xF, "SETNE_DX10",
- [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_NE))]
+ [(set i32:$dst, (selectcc f32:$src0, f32:$src1, -1, 0, COND_UNE))]
>;
def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
-def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
+def TRUNC : R600_1OP_Helper <0x11, "TRUNC", ftrunc>;
def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
def CNDGE_INT : R600_3OP <
0x1E, "CNDGE_INT",
- [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GE))]
+ [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_SGE))]
>;
def CNDGT_INT : R600_3OP <
0x1D, "CNDGT_INT",
- [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_GT))]
+ [(set i32:$dst, (selectcc i32:$src0, 0, i32:$src1, i32:$src2, COND_SGT))]
>;
//===----------------------------------------------------------------------===//
def TEX_SAMPLE_LB : R600_TEX <0x12, "TEX_SAMPLE_LB">;
def TEX_SAMPLE_C_LB : R600_TEX <0x1A, "TEX_SAMPLE_C_LB">;
def TEX_LD : R600_TEX <0x03, "TEX_LD">;
+def TEX_LDPTR : R600_TEX <0x03, "TEX_LDPTR"> {
+ let INST_MOD = 1;
+}
def TEX_GET_TEXTURE_RESINFO : R600_TEX <0x04, "TEX_GET_TEXTURE_RESINFO">;
def TEX_GET_GRADIENTS_H : R600_TEX <0x07, "TEX_GET_GRADIENTS_H">;
def TEX_GET_GRADIENTS_V : R600_TEX <0x08, "TEX_GET_GRADIENTS_V">;
defm : TexPattern<7, TEX_GET_TEXTURE_RESINFO, v4i32>;
defm : TexPattern<8, TEX_GET_GRADIENTS_H>;
defm : TexPattern<9, TEX_GET_GRADIENTS_V>;
+defm : TexPattern<10, TEX_LDPTR, v4i32>;
//===----------------------------------------------------------------------===//
// Helper classes for common instructions
[(set f32:$dst, (fadd (fmul f32:$src0, f32:$src1), f32:$src2))]
>;
+class FMA_Common <bits<5> inst> : R600_3OP <
+ inst, "FMA",
+ [(set f32:$dst, (fma f32:$src0, f32:$src1, f32:$src2))], VecALU
+>;
+
class CNDE_Common <bits<5> inst> : R600_3OP <
inst, "CNDE",
- [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_EQ))]
+ [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_OEQ))]
>;
class CNDGT_Common <bits<5> inst> : R600_3OP <
inst, "CNDGT",
- [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GT))]
->;
+ [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_OGT))]
+> {
+ let Itinerary = VecALU;
+}
class CNDGE_Common <bits<5> inst> : R600_3OP <
inst, "CNDGE",
- [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_GE))]
->;
+ [(set f32:$dst, (selectcc f32:$src0, FP_ZERO, f32:$src1, f32:$src2, COND_OGE))]
+> {
+ let Itinerary = VecALU;
+}
let isCodeGenOnly = 1, isPseudo = 1, Namespace = "AMDGPU" in {
LITERAL:$literal0, LITERAL:$literal1),
"",
pattern,
- AnyALU> {}
+ AnyALU> {
+
+ let UseNamedOperandTable = 1;
+
+}
}
def DOT_4 : R600_VEC2OP<[(set R600_Reg32:$dst, (DOT4
def _pseudo : InstR600 <
(outs R600_Reg128:$dst),
- (ins R600_Reg128:$src),
- "CUBE $dst $src",
- [(set v4f32:$dst, (int_AMDGPU_cube v4f32:$src))],
+ (ins R600_Reg128:$src0),
+ "CUBE $dst $src0",
+ [(set v4f32:$dst, (int_AMDGPU_cube v4f32:$src0))],
VecALU
> {
let isPseudo = 1;
+ let UseNamedOperandTable = 1;
}
def _real : R600_2OP <inst, "CUBE", []>;
class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
inst, "EXP_IEEE", fexp2
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
inst, "FLT_TO_INT", fp_to_sint
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
inst, "INT_TO_FLT", sint_to_fp
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
inst, "FLT_TO_UINT", fp_to_uint
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
inst, "UINT_TO_FLT", uint_to_fp
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
inst, "LOG_IEEE", flog2
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
inst, "MULHI_INT", mulhs
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
inst, "MULHI", mulhu
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
inst, "MULLO_INT", mul
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
inst, "RECIP_CLAMPED", []
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
- inst, "RECIP_IEEE", [(set f32:$dst, (fdiv FP_ONE, f32:$src0))]
+ inst, "RECIP_IEEE", [(set f32:$dst, (AMDGPUrcp f32:$src0))]
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
inst, "RECIP_UINT", AMDGPUurecip
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
+// Clamped to maximum.
class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
- inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
+ inst, "RECIPSQRT_CLAMPED", AMDGPUrsq_clamped
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
-class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
- inst, "RECIPSQRT_IEEE", []
+class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
+ inst, "RECIPSQRT_IEEE", AMDGPUrsq_legacy
> {
- let TransOnly = 1;
let Itinerary = TransALU;
}
+// TODO: There is also RECIPSQRT_FF which clamps to zero.
+
class SIN_Common <bits<11> inst> : R600_1OP <
- inst, "SIN", []>{
+ inst, "SIN", [(set f32:$dst, (SIN_HW f32:$src0))]>{
let Trig = 1;
- let TransOnly = 1;
let Itinerary = TransALU;
}
class COS_Common <bits<11> inst> : R600_1OP <
- inst, "COS", []> {
+ inst, "COS", [(set f32:$dst, (COS_HW f32:$src0))]> {
let Trig = 1;
- let TransOnly = 1;
let Itinerary = TransALU;
}
+def CLAMP_R600 : CLAMP <R600_Reg32>;
+def FABS_R600 : FABS<R600_Reg32>;
+def FNEG_R600 : FNEG<R600_Reg32>;
+
//===----------------------------------------------------------------------===//
// Helper patterns for complex intrinsics
//===----------------------------------------------------------------------===//
+// FIXME: Should be predicated on unsafe fp math.
multiclass DIV_Common <InstR600 recip_ieee> {
def : Pat<
(int_AMDGPU_div f32:$src0, f32:$src1),
(fdiv f32:$src0, f32:$src1),
(MUL_IEEE $src0, (recip_ieee $src1))
>;
+
+def : RcpPat<recip_ieee, f32>;
}
class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee>
(exp_ieee (mul_lit (log_clamped (MAX $src_y, (f32 ZERO))), $src_w, $src_x))
>;
+// FROUND pattern
+class FROUNDPat<Instruction CNDGE, Instruction CNDGT> : Pat <
+ (AMDGPUround f32:$x),
+ (CNDGE $x,
+ (CNDGE (ADD (FNEG_R600 (f32 HALF)), (FRACT $x)), (CEIL $x), (FLOOR $x)),
+ (CNDGT (ADD (FNEG_R600 (f32 HALF)), (FRACT $x)), (CEIL $x), (FLOOR $x))
+ )
+>;
+
+
//===----------------------------------------------------------------------===//
// R600 / R700 Instructions
//===----------------------------------------------------------------------===//
def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_r600 $src))>;
+ defm : RsqPat<RECIPSQRT_IEEE_r600, f32>;
+
+ def : FROUNDPat <CNDGE_r600, CNDGT_r600>;
def R600_ExportSwz : ExportSwzInst {
let Word1{20-17} = 0; // BURST_COUNT
let Word1{21} = eop;
- let Word1{22} = 1; // VALID_PIXEL_MODE
+ let Word1{22} = 0; // VALID_PIXEL_MODE
let Word1{30-23} = inst;
let Word1{31} = 1; // BARRIER
}
def R600_ExportBuf : ExportBufInst {
let Word1{20-17} = 0; // BURST_COUNT
let Word1{21} = eop;
- let Word1{22} = 1; // VALID_PIXEL_MODE
+ let Word1{22} = 0; // VALID_PIXEL_MODE
let Word1{30-23} = inst;
let Word1{31} = 1; // BARRIER
}
defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
- def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$COUNT),
- "TEX $COUNT @$ADDR"> {
+ def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$ADDR, i32imm:$CNT),
+ "TEX $CNT @$ADDR"> {
let POP_COUNT = 0;
}
- def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$COUNT),
- "VTX $COUNT @$ADDR"> {
+ def CF_VC_R600 : CF_CLAUSE_R600<2, (ins i32imm:$ADDR, i32imm:$CNT),
+ "VTX $CNT @$ADDR"> {
let POP_COUNT = 0;
}
def WHILE_LOOP_R600 : CF_CLAUSE_R600<6, (ins i32imm:$ADDR),
"LOOP_START_DX10 @$ADDR"> {
let POP_COUNT = 0;
- let COUNT = 0;
+ let CNT = 0;
}
def END_LOOP_R600 : CF_CLAUSE_R600<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
let POP_COUNT = 0;
- let COUNT = 0;
+ let CNT = 0;
}
def LOOP_BREAK_R600 : CF_CLAUSE_R600<9, (ins i32imm:$ADDR),
"LOOP_BREAK @$ADDR"> {
let POP_COUNT = 0;
- let COUNT = 0;
+ let CNT = 0;
}
def CF_CONTINUE_R600 : CF_CLAUSE_R600<8, (ins i32imm:$ADDR),
"CONTINUE @$ADDR"> {
let POP_COUNT = 0;
- let COUNT = 0;
+ let CNT = 0;
}
def CF_JUMP_R600 : CF_CLAUSE_R600<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
"JUMP @$ADDR POP:$POP_COUNT"> {
- let COUNT = 0;
+ let CNT = 0;
+ }
+ def CF_PUSH_ELSE_R600 : CF_CLAUSE_R600<12, (ins i32imm:$ADDR),
+ "PUSH_ELSE @$ADDR"> {
+ let CNT = 0;
}
def CF_ELSE_R600 : CF_CLAUSE_R600<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
"ELSE @$ADDR POP:$POP_COUNT"> {
- let COUNT = 0;
+ let CNT = 0;
}
def CF_CALL_FS_R600 : CF_CLAUSE_R600<19, (ins), "CALL_FS"> {
let ADDR = 0;
- let COUNT = 0;
+ let CNT = 0;
let POP_COUNT = 0;
}
def POP_R600 : CF_CLAUSE_R600<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
"POP @$ADDR POP:$POP_COUNT"> {
- let COUNT = 0;
+ let CNT = 0;
}
def CF_END_R600 : CF_CLAUSE_R600<0, (ins), "CF_END"> {
- let COUNT = 0;
+ let CNT = 0;
let POP_COUNT = 0;
let ADDR = 0;
let END_OF_PROGRAM = 1;
}
-// Helper pattern for normalizing inputs to triginomic instructions for R700+
-// cards.
-class COS_PAT <InstR600 trig> : Pat<
- (fcos f32:$src),
- (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
->;
-
-class SIN_PAT <InstR600 trig> : Pat<
- (fsin f32:$src),
- (trig (MUL_IEEE (MOV_IMM_I32 CONST.TWO_PI_INV), $src))
->;
-
-//===----------------------------------------------------------------------===//
-// R700 Only instructions
-//===----------------------------------------------------------------------===//
-
-let Predicates = [isR700] in {
- def SIN_r700 : SIN_Common<0x6E>;
- def COS_r700 : COS_Common<0x6F>;
-
- // R700 normalizes inputs to SIN/COS the same as EG
- def : SIN_PAT <SIN_r700>;
- def : COS_PAT <COS_r700>;
-}
-
-//===----------------------------------------------------------------------===//
-// Evergreen Only instructions
-//===----------------------------------------------------------------------===//
-
-let Predicates = [isEG] in {
-
-def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
-defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
-
-def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
-def MULHI_INT_eg : MULHI_INT_Common<0x90>;
-def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
-def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
-def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
-def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
-def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
-def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
-def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
-def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
-def SIN_eg : SIN_Common<0x8D>;
-def COS_eg : COS_Common<0x8E>;
-
-def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
-def : SIN_PAT <SIN_eg>;
-def : COS_PAT <COS_eg>;
-def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>;
-} // End Predicates = [isEG]
-
-//===----------------------------------------------------------------------===//
-// Evergreen / Cayman Instructions
-//===----------------------------------------------------------------------===//
-
-let Predicates = [isEGorCayman] in {
-
- // BFE_UINT - bit_extract, an optimization for mask and shift
- // Src0 = Input
- // Src1 = Offset
- // Src2 = Width
- //
- // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
- //
- // Example Usage:
- // (Offset, Width)
- //
- // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0
- // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8
- // (16,8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16
- // (24,8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24
- def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
- [(set i32:$dst, (int_AMDIL_bit_extract_u32 i32:$src0, i32:$src1,
- i32:$src2))],
- VecALU
- >;
- def : BFEPattern <BFE_UINT_eg>;
-
- def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>;
- defm : BFIPatterns <BFI_INT_eg>;
-
- def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>;
- def : ROTRPattern <BIT_ALIGN_INT_eg>;
-
- def MULADD_eg : MULADD_Common<0x14>;
- def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
- def ASHR_eg : ASHR_Common<0x15>;
- def LSHR_eg : LSHR_Common<0x16>;
- def LSHL_eg : LSHL_Common<0x17>;
- def CNDE_eg : CNDE_Common<0x19>;
- def CNDGT_eg : CNDGT_Common<0x1A>;
- def CNDGE_eg : CNDGE_Common<0x1B>;
- def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
- def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
- def DOT4_eg : DOT4_Common<0xBE>;
- defm CUBE_eg : CUBE_Common<0xC0>;
-
-let hasSideEffects = 1 in {
- def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", []>;
-}
-
- def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
-
- def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
- let Pattern = [];
- }
-
- def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
-
- def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
- let Pattern = [];
- }
-
- def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
-
- // TRUNC is used for the FLT_TO_INT instructions to work around a
- // perceived problem where the rounding modes are applied differently
- // depending on the instruction and the slot they are in.
- // See:
- // https://bugs.freedesktop.org/show_bug.cgi?id=50232
- // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
- //
- // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
- // which do not need to be truncated since the fp values are 0.0f or 1.0f.
- // We should look into handling these cases separately.
- def : Pat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;
-
- def : Pat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;
-
- // SHA-256 Patterns
- def : SHA256MaPattern <BFI_INT_eg, XOR_INT>;
-
- def EG_ExportSwz : ExportSwzInst {
- let Word1{19-16} = 0; // BURST_COUNT
- let Word1{20} = 1; // VALID_PIXEL_MODE
- let Word1{21} = eop;
- let Word1{29-22} = inst;
- let Word1{30} = 0; // MARK
- let Word1{31} = 1; // BARRIER
- }
- defm : ExportPattern<EG_ExportSwz, 83>;
-
- def EG_ExportBuf : ExportBufInst {
- let Word1{19-16} = 0; // BURST_COUNT
- let Word1{20} = 1; // VALID_PIXEL_MODE
- let Word1{21} = eop;
- let Word1{29-22} = inst;
- let Word1{30} = 0; // MARK
- let Word1{31} = 1; // BARRIER
- }
- defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
-
- def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
- "TEX $COUNT @$ADDR"> {
- let POP_COUNT = 0;
- }
- def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
- "VTX $COUNT @$ADDR"> {
- let POP_COUNT = 0;
- }
- def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
- "LOOP_START_DX10 @$ADDR"> {
- let POP_COUNT = 0;
- let COUNT = 0;
- }
- def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
- let POP_COUNT = 0;
- let COUNT = 0;
- }
- def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
- "LOOP_BREAK @$ADDR"> {
- let POP_COUNT = 0;
- let COUNT = 0;
- }
- def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
- "CONTINUE @$ADDR"> {
- let POP_COUNT = 0;
- let COUNT = 0;
- }
- def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
- "JUMP @$ADDR POP:$POP_COUNT"> {
- let COUNT = 0;
- }
- def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
- "ELSE @$ADDR POP:$POP_COUNT"> {
- let COUNT = 0;
- }
- def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
- let ADDR = 0;
- let COUNT = 0;
- let POP_COUNT = 0;
- }
- def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
- "POP @$ADDR POP:$POP_COUNT"> {
- let COUNT = 0;
- }
- def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> {
- let COUNT = 0;
- let POP_COUNT = 0;
- let ADDR = 0;
- let END_OF_PROGRAM = 1;
- }
-
-//===----------------------------------------------------------------------===//
-// Memory read/write instructions
-//===----------------------------------------------------------------------===//
-let usesCustomInserter = 1 in {
-
-class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> mask, string name,
- list<dag> pattern>
- : EG_CF_RAT <0x57, 0x2, 0, (outs), ins, name, pattern> {
- let rim = 0;
- // XXX: Have a separate instruction for non-indexed writes.
- let type = 1;
- let rw_rel = 0;
- let elem_size = 0;
-
- let array_size = 0;
- let comp_mask = mask;
- let burst_count = 0;
- let vpm = 0;
- let mark = 0;
- let barrier = 1;
-}
-
-} // End usesCustomInserter = 1
-
-// 32-bit store
-def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
- (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
- 0x1, "RAT_WRITE_CACHELESS_32_eg $rw_gpr, $index_gpr, $eop",
- [(global_store i32:$rw_gpr, i32:$index_gpr)]
->;
-
-//128-bit store
-def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
- (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
- 0xf, "RAT_WRITE_CACHELESS_128 $rw_gpr.XYZW, $index_gpr, $eop",
- [(global_store v4i32:$rw_gpr, i32:$index_gpr)]
->;
-
-class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
- : InstR600ISA <outs, (ins MEMxi:$ptr), name, pattern>,
- VTX_WORD1_GPR, VTX_WORD0 {
-
- // Static fields
- let VC_INST = 0;
- let FETCH_TYPE = 2;
- let FETCH_WHOLE_QUAD = 0;
- let BUFFER_ID = buffer_id;
- let 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.
- let SRC_SEL_X = 0;
- let DST_REL = 0;
- // 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.
- let USE_CONST_FIELDS = 0;
- let NUM_FORMAT_ALL = 1;
- let FORMAT_COMP_ALL = 0;
- let SRF_MODE_ALL = 0;
-
- let Inst{31-0} = Word0;
- let Inst{63-32} = Word1;
- // 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_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;
-
- let VTXInst = 1;
-}
-
-class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
- : VTX_READ_eg <"VTX_READ_8 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
- pattern> {
-
- let MEGA_FETCH_COUNT = 1;
- 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 = 1; // FMT_8
-}
-
-class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
- : VTX_READ_eg <"VTX_READ_16 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$dst),
- pattern> {
- let MEGA_FETCH_COUNT = 2;
- 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 = 5; // FMT_16
-
-}
-
-class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
- : VTX_READ_eg <"VTX_READ_32 $dst, $ptr", buffer_id, (outs R600_TReg32_X:$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 $ptr registers of the VTX_READ.
- // e.g.
- // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
- // %T2_X<def> = MOV %ZERO
- //Adding this constraint prevents this from happening.
- let Constraints = "$ptr.ptr = $dst";
-}
-
-class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
- : VTX_READ_eg <"VTX_READ_128 $dst.XYZW, $ptr", buffer_id, (outs R600_Reg128:$dst),
- pattern> {
-
- 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
-
- // 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 $dst
- // registers are different sizes.
-}
-
-//===----------------------------------------------------------------------===//
-// VTX Read from parameter memory space
-//===----------------------------------------------------------------------===//
-
-def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
- [(set i32:$dst, (load_param_zexti8 ADDRVTX_READ:$ptr))]
->;
-
-def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
- [(set i32:$dst, (load_param_zexti16 ADDRVTX_READ:$ptr))]
->;
-
-def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
- [(set i32:$dst, (load_param ADDRVTX_READ:$ptr))]
->;
-
-def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
- [(set v4i32:$dst, (load_param ADDRVTX_READ:$ptr))]
->;
-
-//===----------------------------------------------------------------------===//
-// VTX Read from global memory space
-//===----------------------------------------------------------------------===//
-
-// 8-bit reads
-def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
- [(set i32:$dst, (zextloadi8_global ADDRVTX_READ:$ptr))]
->;
-
-// 32-bit reads
-def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
- [(set i32:$dst, (global_load ADDRVTX_READ:$ptr))]
->;
-
-// 128-bit reads
-def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
- [(set v4i32:$dst, (global_load ADDRVTX_READ:$ptr))]
->;
-
-//===----------------------------------------------------------------------===//
-// Constant Loads
-// XXX: We are currently storing all constants in the global address space.
-//===----------------------------------------------------------------------===//
-
-def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
- [(set i32:$dst, (constant_load ADDRVTX_READ:$ptr))]
->;
-
-}
//===----------------------------------------------------------------------===//
// Regist loads and stores - for indirect addressing
defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
-let Predicates = [isCayman] in {
-
-let isVector = 1 in {
-
-def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
-
-def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
-def MULHI_INT_cm : MULHI_INT_Common<0x90>;
-def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
-def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
-def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
-def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
-def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
-def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
-def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
-def SIN_cm : SIN_Common<0x8D>;
-def COS_cm : COS_Common<0x8E>;
-} // End isVector = 1
-
-def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
-def : SIN_PAT <SIN_cm>;
-def : COS_PAT <COS_cm>;
-
-defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
-
-// RECIP_UINT emulation for Cayman
-// The multiplication scales from [0,1] to the unsigned integer range
-def : Pat <
- (AMDGPUurecip i32:$src0),
- (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
- (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
->;
-
- def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
- let ADDR = 0;
- let POP_COUNT = 0;
- let COUNT = 0;
- }
-
-def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
-
-} // End isCayman
-
-//===----------------------------------------------------------------------===//
-// Branch Instructions
-//===----------------------------------------------------------------------===//
-
-
-def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src),
- "IF_PREDICATE_SET $src", []>;
-
-def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
- "PREDICATED_BREAK $src", []>;
//===----------------------------------------------------------------------===//
// Pseudo instructions
} // End isPseudo = 1
} // End usesCustomInserter = 1
-def CLAMP_R600 : CLAMP <R600_Reg32>;
-def FABS_R600 : FABS<R600_Reg32>;
-def FNEG_R600 : FNEG<R600_Reg32>;
-
-//===---------------------------------------------------------------------===//
-// Return instruction
-//===---------------------------------------------------------------------===//
-let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
- usesCustomInserter = 1 in {
- def RETURN : ILFormat<(outs), (ins variable_ops),
- "RETURN", [(IL_retflag)]>;
-}
-
//===----------------------------------------------------------------------===//
// Constant Buffer Addressing Support
def TEX_VTX_CONSTBUF :
InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "VTX_READ_eg $dst, $ptr",
[(set v4i32:$dst, (CONST_ADDRESS ADDRGA_VAR_OFFSET:$ptr, (i32 imm:$BUFFER_ID)))]>,
- VTX_WORD1_GPR, VTX_WORD0 {
+ VTX_WORD1_GPR, VTX_WORD0_eg {
let VC_INST = 0;
let FETCH_TYPE = 2;
def TEX_VTX_TEXBUF:
InstR600ISA <(outs R600_Reg128:$dst), (ins MEMxi:$ptr, i32imm:$BUFFER_ID), "TEX_VTX_EXPLICIT_READ $dst, $ptr",
[(set v4f32:$dst, (int_R600_load_texbuf ADDRGA_VAR_OFFSET:$ptr, imm:$BUFFER_ID))]>,
-VTX_WORD1_GPR, VTX_WORD0 {
+VTX_WORD1_GPR, VTX_WORD0_eg {
let VC_INST = 0;
let FETCH_TYPE = 2;
let VTXInst = 1;
}
+//===---------------------------------------------------------------------===//
+// Flow and Program control Instructions
+//===---------------------------------------------------------------------===//
+class ILFormat<dag outs, dag ins, string asmstr, list<dag> pattern>
+: Instruction {
+
+ let Namespace = "AMDGPU";
+ dag OutOperandList = outs;
+ dag InOperandList = ins;
+ let Pattern = pattern;
+ let AsmString = !strconcat(asmstr, "\n");
+ let isPseudo = 1;
+ let Itinerary = NullALU;
+ bit hasIEEEFlag = 0;
+ bit hasZeroOpFlag = 0;
+ let mayLoad = 0;
+ let mayStore = 0;
+ let hasSideEffects = 0;
+}
+
+multiclass BranchConditional<SDNode Op, RegisterClass rci, RegisterClass rcf> {
+ def _i32 : ILFormat<(outs),
+ (ins brtarget:$target, rci:$src0),
+ "; i32 Pseudo branch instruction",
+ [(Op bb:$target, (i32 rci:$src0))]>;
+ def _f32 : ILFormat<(outs),
+ (ins brtarget:$target, rcf:$src0),
+ "; f32 Pseudo branch instruction",
+ [(Op bb:$target, (f32 rcf:$src0))]>;
+}
+
+// Only scalar types should generate flow control
+multiclass BranchInstr<string name> {
+ def _i32 : ILFormat<(outs), (ins R600_Reg32:$src),
+ !strconcat(name, " $src"), []>;
+ def _f32 : ILFormat<(outs), (ins R600_Reg32:$src),
+ !strconcat(name, " $src"), []>;
+}
+// Only scalar types should generate flow control
+multiclass BranchInstr2<string name> {
+ def _i32 : ILFormat<(outs), (ins R600_Reg32:$src0, R600_Reg32:$src1),
+ !strconcat(name, " $src0, $src1"), []>;
+ def _f32 : ILFormat<(outs), (ins R600_Reg32:$src0, R600_Reg32:$src1),
+ !strconcat(name, " $src0, $src1"), []>;
+}
-
-//===--------------------------------------------------------------------===//
-// Instructions support
-//===--------------------------------------------------------------------===//
//===---------------------------------------------------------------------===//
// Custom Inserter for Branches and returns, this eventually will be a
-// seperate pass
+// separate pass
//===---------------------------------------------------------------------===//
let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
def BRANCH : ILFormat<(outs), (ins brtarget:$target),
"; Pseudo unconditional branch instruction",
[(br bb:$target)]>;
- defm BRANCH_COND : BranchConditional<IL_brcond>;
+ defm BRANCH_COND : BranchConditional<IL_brcond, R600_Reg32, R600_Reg32>;
}
//===---------------------------------------------------------------------===//
-// Flow and Program control Instructions
+// Return instruction
//===---------------------------------------------------------------------===//
+let isTerminator = 1, isReturn = 1, hasCtrlDep = 1,
+ usesCustomInserter = 1 in {
+ def RETURN : ILFormat<(outs), (ins variable_ops),
+ "RETURN", [(IL_retflag)]>;
+}
+
+//===----------------------------------------------------------------------===//
+// Branch Instructions
+//===----------------------------------------------------------------------===//
+
+def IF_PREDICATE_SET : ILFormat<(outs), (ins R600_Reg32:$src),
+ "IF_PREDICATE_SET $src", []>;
+
let isTerminator=1 in {
- def SWITCH : ILFormat< (outs), (ins GPRI32:$src),
- !strconcat("SWITCH", " $src"), []>;
- def CASE : ILFormat< (outs), (ins GPRI32:$src),
- !strconcat("CASE", " $src"), []>;
def BREAK : ILFormat< (outs), (ins),
"BREAK", []>;
def CONTINUE : ILFormat< (outs), (ins),
defm CONTINUEC : BranchInstr2<"CONTINUEC">;
}
+//===----------------------------------------------------------------------===//
+// Indirect addressing pseudo instructions
+//===----------------------------------------------------------------------===//
+
+let isPseudo = 1 in {
+
+class ExtractVertical <RegisterClass vec_rc> : InstR600 <
+ (outs R600_Reg32:$dst),
+ (ins vec_rc:$vec, R600_Reg32:$index), "",
+ [],
+ AnyALU
+>;
+
+let Constraints = "$dst = $vec" in {
+
+class InsertVertical <RegisterClass vec_rc> : InstR600 <
+ (outs vec_rc:$dst),
+ (ins vec_rc:$vec, R600_Reg32:$value, R600_Reg32:$index), "",
+ [],
+ AnyALU
+>;
+
+} // End Constraints = "$dst = $vec"
+
+} // End isPseudo = 1
+
+def R600_EXTRACT_ELT_V2 : ExtractVertical <R600_Reg64Vertical>;
+def R600_EXTRACT_ELT_V4 : ExtractVertical <R600_Reg128Vertical>;
+
+def R600_INSERT_ELT_V2 : InsertVertical <R600_Reg64Vertical>;
+def R600_INSERT_ELT_V4 : InsertVertical <R600_Reg128Vertical>;
+
+class ExtractVerticalPat <Instruction inst, ValueType vec_ty,
+ ValueType scalar_ty> : Pat <
+ (scalar_ty (extractelt vec_ty:$vec, i32:$index)),
+ (inst $vec, $index)
+>;
+
+def : ExtractVerticalPat <R600_EXTRACT_ELT_V2, v2i32, i32>;
+def : ExtractVerticalPat <R600_EXTRACT_ELT_V2, v2f32, f32>;
+def : ExtractVerticalPat <R600_EXTRACT_ELT_V4, v4i32, i32>;
+def : ExtractVerticalPat <R600_EXTRACT_ELT_V4, v4f32, f32>;
+
+class InsertVerticalPat <Instruction inst, ValueType vec_ty,
+ ValueType scalar_ty> : Pat <
+ (vec_ty (insertelt vec_ty:$vec, scalar_ty:$value, i32:$index)),
+ (inst $vec, $value, $index)
+>;
+
+def : InsertVerticalPat <R600_INSERT_ELT_V2, v2i32, i32>;
+def : InsertVerticalPat <R600_INSERT_ELT_V2, v2f32, f32>;
+def : InsertVerticalPat <R600_INSERT_ELT_V4, v4i32, i32>;
+def : InsertVerticalPat <R600_INSERT_ELT_V4, v4f32, f32>;
+
//===----------------------------------------------------------------------===//
// ISel Patterns
//===----------------------------------------------------------------------===//
//CNDGE_INT extra pattern
def : Pat <
- (selectcc i32:$src0, -1, i32:$src1, i32:$src2, COND_GT),
+ (selectcc i32:$src0, -1, i32:$src1, i32:$src2, COND_SGT),
(CNDGE_INT $src0, $src1, $src2)
>;
(MASK_WRITE (KILLGT (f32 ZERO), $src0))
>;
-// SGT Reverse args
-def : Pat <
- (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LT),
- (SGT $src1, $src0)
->;
-
-// SGE Reverse args
-def : Pat <
- (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, COND_LE),
- (SGE $src1, $src0)
->;
-
-// SETGT_DX10 reverse args
-def : Pat <
- (selectcc f32:$src0, f32:$src1, -1, 0, COND_LT),
- (SETGT_DX10 $src1, $src0)
->;
-
-// SETGE_DX10 reverse args
-def : Pat <
- (selectcc f32:$src0, f32:$src1, -1, 0, COND_LE),
- (SETGE_DX10 $src1, $src0)
->;
-
-// SETGT_INT reverse args
-def : Pat <
- (selectcc i32:$src0, i32:$src1, -1, 0, SETLT),
- (SETGT_INT $src1, $src0)
->;
-
-// SETGE_INT reverse args
-def : Pat <
- (selectcc i32:$src0, i32:$src1, -1, 0, SETLE),
- (SETGE_INT $src1, $src0)
->;
-
-// SETGT_UINT reverse args
-def : Pat <
- (selectcc i32:$src0, i32:$src1, -1, 0, SETULT),
- (SETGT_UINT $src1, $src0)
->;
-
-// SETGE_UINT reverse args
-def : Pat <
- (selectcc i32:$src0, i32:$src1, -1, 0, SETULE),
- (SETGE_UINT $src1, $src0)
->;
-
-// The next two patterns are special cases for handling 'true if ordered' and
-// 'true if unordered' conditionals. The assumption here is that the behavior of
-// SETE and SNE conforms to the Direct3D 10 rules for floating point values
-// described here:
-// http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
-// We assume that SETE returns false when one of the operands is NAN and
-// SNE returns true when on of the operands is NAN
-
-//SETE - 'true if ordered'
-def : Pat <
- (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETO),
- (SETE $src0, $src1)
->;
-
-//SETE_DX10 - 'true if ordered'
-def : Pat <
- (selectcc f32:$src0, f32:$src1, -1, 0, SETO),
- (SETE_DX10 $src0, $src1)
->;
-
-//SNE - 'true if unordered'
-def : Pat <
- (selectcc f32:$src0, f32:$src1, FP_ONE, FP_ZERO, SETUO),
- (SNE $src0, $src1)
->;
-
-//SETNE_DX10 - 'true if ordered'
-def : Pat <
- (selectcc f32:$src0, f32:$src1, -1, 0, SETUO),
- (SETNE_DX10 $src0, $src1)
->;
-
def : Extract_Element <f32, v4f32, 0, sub0>;
def : Extract_Element <f32, v4f32, 1, sub1>;
def : Extract_Element <f32, v4f32, 2, sub2>;
def : Insert_Element <i32, v4i32, 2, sub2>;
def : Insert_Element <i32, v4i32, 3, sub3>;
-def : Vector4_Build <v4f32, f32>;
-def : Vector4_Build <v4i32, i32>;
+def : Extract_Element <f32, v2f32, 0, sub0>;
+def : Extract_Element <f32, v2f32, 1, sub1>;
+
+def : Insert_Element <f32, v2f32, 0, sub0>;
+def : Insert_Element <f32, v2f32, 1, sub1>;
+
+def : Extract_Element <i32, v2i32, 0, sub0>;
+def : Extract_Element <i32, v2i32, 1, sub1>;
+
+def : Insert_Element <i32, v2i32, 0, sub0>;
+def : Insert_Element <i32, v2i32, 1, sub1>;
// bitconvert patterns
def : BitConvert <i32, f32, R600_Reg32>;
def : BitConvert <f32, i32, R600_Reg32>;
+def : BitConvert <v2f32, v2i32, R600_Reg64>;
+def : BitConvert <v2i32, v2f32, R600_Reg64>;
def : BitConvert <v4f32, v4i32, R600_Reg128>;
def : BitConvert <v4i32, v4f32, R600_Reg128>;
def : DwordAddrPat <i32, R600_Reg32>;
} // End isR600toCayman Predicate
+
+let Predicates = [isR600] in {
+// Intrinsic patterns
+defm : Expand24IBitOps<MULLO_INT_r600, ADD_INT>;
+defm : Expand24UBitOps<MULLO_UINT_r600, ADD_INT>;
+} // End isR600
+
+def getLDSNoRetOp : InstrMapping {
+ let FilterClass = "R600_LDS_1A1D";
+ let RowFields = ["BaseOp"];
+ let ColFields = ["DisableEncoding"];
+ let KeyCol = ["$dst"];
+ let ValueCols = [[""""]];
+}