let update_pred = 0;
let HasNativeOperands = 1;
let Op1 = 1;
+ let ALUInst = 1;
let DisableEncoding = "$literal";
+ let UseNamedOperandTable = 1;
let Inst{31-0} = Word0;
let Inst{63-32} = Word1;
[(set R600_Reg32:$dst, (node R600_Reg32:$src0))]
>;
-// If you add our change the operands for R600_2OP instructions, you must
+// If you add or change the operands for R600_2OP instructions, you must
// also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
// R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
let HasNativeOperands = 1;
let Op2 = 1;
+ let ALUInst = 1;
let DisableEncoding = "$literal";
+ let UseNamedOperandTable = 1;
let Inst{31-0} = Word0;
let Inst{63-32} = Word1;
let HasNativeOperands = 1;
let DisableEncoding = "$literal";
let Op3 = 1;
+ let UseNamedOperandTable = 1;
+ let ALUInst = 1;
let Inst{31-0} = Word0;
let Inst{63-32} = Word1;
}]
>;
-class EG_CF_RAT <bits <8> cf_inst, bits <6> rat_inst, bits<4> rat_id, dag outs,
+class EG_CF_RAT <bits <8> cfinst, bits <6> ratinst, bits<4> mask, dag outs,
dag ins, string asm, list<dag> pattern> :
- InstR600ISA <outs, ins, asm, pattern> {
- bits<7> RW_GPR;
- bits<7> INDEX_GPR;
-
- bits<2> RIM;
- bits<2> TYPE;
- bits<1> RW_REL;
- bits<2> ELEM_SIZE;
-
- bits<12> ARRAY_SIZE;
- bits<4> COMP_MASK;
- bits<4> BURST_COUNT;
- bits<1> VPM;
- bits<1> eop;
- bits<1> MARK;
- bits<1> BARRIER;
-
- // CF_ALLOC_EXPORT_WORD0_RAT
- let Inst{3-0} = rat_id;
- let Inst{9-4} = rat_inst;
- let Inst{10} = 0; // Reserved
- let Inst{12-11} = RIM;
- let Inst{14-13} = TYPE;
- let Inst{21-15} = RW_GPR;
- let Inst{22} = RW_REL;
- let Inst{29-23} = INDEX_GPR;
- let Inst{31-30} = ELEM_SIZE;
-
- // CF_ALLOC_EXPORT_WORD1_BUF
- let Inst{43-32} = ARRAY_SIZE;
- let Inst{47-44} = COMP_MASK;
- let Inst{51-48} = BURST_COUNT;
- let Inst{52} = VPM;
- let Inst{53} = eop;
- let Inst{61-54} = cf_inst;
- let Inst{62} = MARK;
- let Inst{63} = BARRIER;
+ InstR600ISA <outs, ins, asm, pattern>,
+ CF_ALLOC_EXPORT_WORD0_RAT, CF_ALLOC_EXPORT_WORD1_BUF {
+
+ let rat_id = 0;
+ let rat_inst = ratinst;
+ let rim = 0;
+ // XXX: Have a separate instruction for non-indexed writes.
+ let type = 1;
+ let rw_rel = 0;
+ let elem_size = 0;
+
+ let array_size = 0;
+ let comp_mask = mask;
+ let burst_count = 0;
+ let vpm = 0;
+ let cf_inst = cfinst;
+ let mark = 0;
+ let barrier = 1;
+
+ let Inst{31-0} = Word0;
+ let Inst{63-32} = Word1;
+
+}
+
+class VTX_READ <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
+ : InstR600ISA <outs, (ins MEMxi:$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 <
[]
>;
+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, []>;
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$sw_w"),
[]>, ExportWord0, ExportSwzWord1 {
let elem_size = 3;
let Inst{31-0} = Word0;
(ins i32imm:$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 {
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;
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:$src0))],
VecALU
> {
let isPseudo = 1;
+ let UseNamedOperandTable = 1;
}
def _real : R600_2OP <inst, "CUBE", []>;
}
class SIN_Common <bits<11> inst> : R600_1OP <
- inst, "SIN", []>{
+ inst, "SIN", [(set f32:$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:$src0))]> {
let Trig = 1;
let TransOnly = 1;
let Itinerary = TransALU;
}
defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
- def CF_TC_R600 : CF_CLAUSE_R600<1, (ins i32imm:$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_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), $src))
->;
-
//===----------------------------------------------------------------------===//
// R700 Only instructions
//===----------------------------------------------------------------------===//
let Predicates = [isR700] in {
def SIN_r700 : SIN_Common<0x6E>;
def COS_r700 : COS_Common<0x6F>;
-
- // R700 normalizes inputs to SIN/COS the same as EG
- def : SIN_PAT <SIN_r700>;
- def : COS_PAT <COS_r700>;
}
//===----------------------------------------------------------------------===//
def COS_eg : COS_Common<0x8E>;
def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
-def : SIN_PAT <SIN_eg>;
-def : COS_PAT <COS_eg>;
def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>;
+
+//===----------------------------------------------------------------------===//
+// Memory read/write instructions
+//===----------------------------------------------------------------------===//
+let usesCustomInserter = 1 in {
+
+class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> mask, string name,
+ list<dag> pattern>
+ : EG_CF_RAT <0x57, 0x2, mask, (outs), ins, name, pattern> {
+}
+
+} // End usesCustomInserter = 1
+
+// 32-bit store
+def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
+ (ins R600_TReg32_X:$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>
+ : VTX_WORD0_eg, VTX_READ<name, buffer_id, outs, pattern> {
+
+ // 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 Inst{31-0} = Word0;
+}
+
+class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
+ : VTX_READ_eg <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id,
+ (outs R600_TReg32_X:$dst_gpr), 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_gpr, $src_gpr", buffer_id,
+ (outs R600_TReg32_X:$dst_gpr), 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_gpr, $src_gpr", buffer_id,
+ (outs R600_TReg32_X:$dst_gpr), 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 $src_gpr registers of the VTX_READ.
+ // e.g.
+ // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
+ // %T2_X<def> = MOV %ZERO
+ //Adding this constraint prevents this from happening.
+ let Constraints = "$src_gpr.ptr = $dst_gpr";
+}
+
+class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
+ : VTX_READ_eg <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id,
+ (outs R600_Reg128:$dst_gpr), 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 $src_gpr.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_gpr, (load_param_zexti8 ADDRVTX_READ:$src_gpr))]
+>;
+
+def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
+ [(set i32:$dst_gpr, (load_param_zexti16 ADDRVTX_READ:$src_gpr))]
+>;
+
+def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
+ [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
+>;
+
+def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
+ [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
+>;
+
+//===----------------------------------------------------------------------===//
+// VTX Read from global memory space
+//===----------------------------------------------------------------------===//
+
+// 8-bit reads
+def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
+ [(set i32:$dst_gpr, (zextloadi8_global ADDRVTX_READ:$src_gpr))]
+>;
+
+// 32-bit reads
+def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
+ [(set i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
+>;
+
+// 128-bit reads
+def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
+ [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
+>;
+
+//===----------------------------------------------------------------------===//
+// Constant Loads
+// XXX: We are currently storing all constants in the global address space.
+//===----------------------------------------------------------------------===//
+
+def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
+ [(set i32:$dst_gpr, (constant_load ADDRVTX_READ:$src_gpr))]
+>;
+
+
} // End Predicates = [isEG]
//===----------------------------------------------------------------------===//
def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
let Pattern = [];
+ let TransOnly = 0;
+ let Itinerary = AnyALU;
}
def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
+def GROUP_BARRIER : InstR600 <
+ (outs), (ins), " GROUP_BARRIER", [(int_AMDGPU_barrier_local)], AnyALU>,
+ R600ALU_Word0,
+ R600ALU_Word1_OP2 <0x54> {
+
+ let dst = 0;
+ let dst_rel = 0;
+ let src0 = 0;
+ let src0_rel = 0;
+ let src0_neg = 0;
+ let src0_abs = 0;
+ let src1 = 0;
+ let src1_rel = 0;
+ let src1_neg = 0;
+ let src1_abs = 0;
+ let write = 0;
+ let omod = 0;
+ let clamp = 0;
+ let last = 1;
+ let bank_swizzle = 0;
+ let pred_sel = 0;
+ let update_exec_mask = 0;
+ let update_pred = 0;
+
+ let Inst{31-0} = Word0;
+ let Inst{63-32} = Word1;
+
+ let ALUInst = 1;
+}
+
+//===----------------------------------------------------------------------===//
+// LDS Instructions
+//===----------------------------------------------------------------------===//
+class R600_LDS <bits<6> op, dag outs, dag ins, string asm,
+ list<dag> pattern = []> :
+
+ InstR600 <outs, ins, asm, pattern, XALU>,
+ R600_ALU_LDS_Word0,
+ R600LDS_Word1 {
+
+ bits<6> offset = 0;
+ let lds_op = op;
+
+ let Word1{27} = offset{0};
+ let Word1{12} = offset{1};
+ let Word1{28} = offset{2};
+ let Word1{31} = offset{3};
+ let Word0{12} = offset{4};
+ let Word0{25} = offset{5};
+
+
+ let Inst{31-0} = Word0;
+ let Inst{63-32} = Word1;
+
+ let ALUInst = 1;
+ let HasNativeOperands = 1;
+ let UseNamedOperandTable = 1;
+}
+
+class R600_LDS_1A <bits<6> lds_op, string name, list<dag> pattern> : R600_LDS <
+ lds_op,
+ (outs R600_Reg32:$dst),
+ (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
+ LAST:$last, R600_Pred:$pred_sel,
+ BANK_SWIZZLE:$bank_swizzle),
+ " "#name#" $last OQAP, $src0$src0_rel $pred_sel",
+ pattern
+ > {
+
+ let src1 = 0;
+ let src1_rel = 0;
+ let src2 = 0;
+ let src2_rel = 0;
+
+ let Defs = [OQAP];
+ let usesCustomInserter = 1;
+ let LDS_1A = 1;
+ let DisableEncoding = "$dst";
+}
+
+class R600_LDS_1A1D <bits<6> lds_op, string name, list<dag> pattern> :
+ R600_LDS <
+ lds_op,
+ (outs),
+ (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
+ R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel,
+ LAST:$last, R600_Pred:$pred_sel,
+ BANK_SWIZZLE:$bank_swizzle),
+ " "#name#" $last $src0$src0_rel, $src1$src1_rel, $pred_sel",
+ pattern
+ > {
+
+ let src2 = 0;
+ let src2_rel = 0;
+ let LDS_1A1D = 1;
+}
+
+def LDS_READ_RET : R600_LDS_1A <0x32, "LDS_READ_RET",
+ [(set (i32 R600_Reg32:$dst), (local_load R600_Reg32:$src0))]
+>;
+
+def LDS_WRITE : R600_LDS_1A1D <0xD, "LDS_WRITE",
+ [(local_store (i32 R600_Reg32:$src1), R600_Reg32:$src0)]
+>;
+
// TRUNC is used for the FLT_TO_INT instructions to work around a
// perceived problem where the rounding modes are applied differently
// depending on the instruction and the slot they are in.
let END_OF_PROGRAM = 1;
}
+} // End Predicates = [isEGorCayman]
+
//===----------------------------------------------------------------------===//
-// Memory read/write instructions
+// Regist loads and stores - for indirect addressing
//===----------------------------------------------------------------------===//
-let usesCustomInserter = 1 in {
-class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> comp_mask, string name,
- list<dag> pattern>
- : EG_CF_RAT <0x57, 0x2, 0, (outs), ins, name, pattern> {
- let RIM = 0;
- // XXX: Have a separate instruction for non-indexed writes.
- let TYPE = 1;
- let RW_REL = 0;
- let ELEM_SIZE = 0;
+defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
- let ARRAY_SIZE = 0;
- let COMP_MASK = comp_mask;
- let BURST_COUNT = 0;
- let VPM = 0;
- let MARK = 0;
- let BARRIER = 1;
-}
+//===----------------------------------------------------------------------===//
+// Cayman Instructions
+//===----------------------------------------------------------------------===//
-} // End usesCustomInserter = 1
+let Predicates = [isCayman] in {
-// 32-bit store
-def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
- (ins R600_TReg32_X:$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:$index_gpr)]
->;
+let isVector = 1 in {
-//128-bit store
-def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
- (ins R600_Reg128:$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:$index_gpr)]
+def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
+
+def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
+def MULHI_INT_cm : MULHI_INT_Common<0x90>;
+def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
+def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
+def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
+def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
+def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
+def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
+def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
+def SIN_cm : SIN_Common<0x8D>;
+def COS_cm : COS_Common<0x8E>;
+} // End isVector = 1
+
+def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
+
+defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
+
+// RECIP_UINT emulation for Cayman
+// The multiplication scales from [0,1] to the unsigned integer range
+def : Pat <
+ (AMDGPUurecip i32:$src0),
+ (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
+ (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
>;
-class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
- : InstR600ISA <outs, (ins MEMxi:$ptr), name, pattern>,
- VTX_WORD1_GPR, VTX_WORD0 {
+ 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 $src))>;
+
+
+def RAT_STORE_DWORD_cm : EG_CF_RAT <
+ 0x57, 0x14, 0x1, (outs),
+ (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr),
+ "EXPORT_RAT_INST_STORE_DWORD $rw_gpr, $index_gpr",
+ [(global_store i32:$rw_gpr, i32:$index_gpr)]
+> {
+ let eop = 0; // This bit is not used on Cayman.
+}
+
+class VTX_READ_cm <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
+ : VTX_WORD0_cm, VTX_READ<name, buffer_id, outs, pattern> {
// Static fields
let VC_INST = 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 SRC_SEL_Y = 0;
+ let STRUCTURED_READ = 0;
+ let LDS_REQ = 0;
+ let COALESCED_READ = 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> {
+class VTX_READ_8_cm <bits<8> buffer_id, list<dag> pattern>
+ : VTX_READ_cm <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id,
+ (outs R600_TReg32_X:$dst_gpr), pattern> {
- let MEGA_FETCH_COUNT = 1;
let DST_SEL_X = 0;
let DST_SEL_Y = 7; // Masked
let DST_SEL_Z = 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;
+class VTX_READ_16_cm <bits<8> buffer_id, list<dag> pattern>
+ : VTX_READ_cm <"VTX_READ_16 $dst_gpr, $src_gpr", buffer_id,
+ (outs R600_TReg32_X:$dst_gpr), pattern> {
let DST_SEL_X = 0;
let DST_SEL_Y = 7; // Masked
let DST_SEL_Z = 7; // Masked
}
-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> {
+class VTX_READ_32_cm <bits<8> buffer_id, list<dag> pattern>
+ : VTX_READ_cm <"VTX_READ_32 $dst_gpr, $src_gpr", buffer_id,
+ (outs R600_TReg32_X:$dst_gpr), pattern> {
- let MEGA_FETCH_COUNT = 4;
let DST_SEL_X = 0;
let DST_SEL_Y = 7; // Masked
let DST_SEL_Z = 7; // Masked
// 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.
+ // to the $src_gpr registers of the VTX_READ.
// e.g.
// %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
// %T2_X<def> = MOV %ZERO
//Adding this constraint prevents this from happening.
- let Constraints = "$ptr.ptr = $dst";
+ let Constraints = "$src_gpr.ptr = $dst_gpr";
}
-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> {
+class VTX_READ_128_cm <bits<8> buffer_id, list<dag> pattern>
+ : VTX_READ_cm <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id,
+ (outs R600_Reg128:$dst_gpr), pattern> {
- let MEGA_FETCH_COUNT = 16;
let DST_SEL_X = 0;
let DST_SEL_Y = 1;
let DST_SEL_Z = 2;
// 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
+ // the same constraint as VTX_READ_32_eg, because the $src_gpr.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:$ptr))]
+def VTX_READ_PARAM_8_cm : VTX_READ_8_cm <0,
+ [(set i32:$dst_gpr, (load_param_zexti8 ADDRVTX_READ:$src_gpr))]
>;
-def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
- [(set i32:$dst, (load_param_zexti16 ADDRVTX_READ:$ptr))]
+def VTX_READ_PARAM_16_cm : VTX_READ_16_cm <0,
+ [(set i32:$dst_gpr, (load_param_zexti16 ADDRVTX_READ:$src_gpr))]
>;
-def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
- [(set i32:$dst, (load_param ADDRVTX_READ:$ptr))]
+def VTX_READ_PARAM_32_cm : VTX_READ_32_cm <0,
+ [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
>;
-def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0,
- [(set v4i32:$dst, (load_param ADDRVTX_READ:$ptr))]
+def VTX_READ_PARAM_128_cm : VTX_READ_128_cm <0,
+ [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
>;
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// 8-bit reads
-def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
- [(set i32:$dst, (zextloadi8_global ADDRVTX_READ:$ptr))]
+def VTX_READ_GLOBAL_8_cm : VTX_READ_8_cm <1,
+ [(set i32:$dst_gpr, (zextloadi8_global ADDRVTX_READ:$src_gpr))]
>;
// 32-bit reads
-def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
- [(set i32:$dst, (global_load ADDRVTX_READ:$ptr))]
+def VTX_READ_GLOBAL_32_cm : VTX_READ_32_cm <1,
+ [(set i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
>;
// 128-bit reads
-def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
- [(set v4i32:$dst, (global_load ADDRVTX_READ:$ptr))]
+def VTX_READ_GLOBAL_128_cm : VTX_READ_128_cm <1,
+ [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))]
>;
//===----------------------------------------------------------------------===//
// XXX: We are currently storing all constants in the global address space.
//===----------------------------------------------------------------------===//
-def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
- [(set i32:$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 $src0)),
- (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
+def CONSTANT_LOAD_cm : VTX_READ_32_cm <1,
+ [(set i32:$dst_gpr, (constant_load ADDRVTX_READ:$src_gpr))]
>;
- def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
- let ADDR = 0;
- let POP_COUNT = 0;
- let COUNT = 0;
- }
-
-def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
-
} // End isCayman
//===----------------------------------------------------------------------===//
def TEX_VTX_CONSTBUF :
InstR600ISA <(outs R600_Reg128:$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;
projects / oota-llvm.git / blobdiff