Reorder Buffer: register renaming and inorder completion Use of a reorder buffer Reorder buffer = circular queue with head and tail pointers At issue (renaming time), an instruction is assigned an entry at the tail of the reorder buffer (ROB) which becomes the name of (or a pointer to) the result register. Recall that instructions are issued in program order, thus the ROB stores instructions in program order At end of functional-unit computation, value is put in the instruction reorder buffers position When the instruction reaches the head of the buffer, its value is stored in the logical or physical (other reorder buffer entry) register. Need of a mapping table between logical registers and ROB entries Reg. ren & Mult. Issue CSE 471 1 Example Before: add r3,r3,4 after add rob6,r3,4 add r4,r7,r3 add rob7,r7,rob6 add r3, r2, r7 add rob8,r2,r7 Assume reorder buffer is initially at position 6 and has more than 8 slots The mapping table indicates the correspondence between ROB entries and logical registers Reg. ren & Mult. Issue CSE 471 2
Data dependencies with register renaming Register renaming does not get rid of RAW dependencies Still need for forwarding or for indicating whether a register has received its value Register renaming gets rid of WAW and WAR dependencies The reorder buffer, as its name implies, can be used for inorder completion Reg. ren & Mult. Issue CSE 471 3 More on reorder buffer Tomasulos scheme can be extended with the possibility of completing instructions in order Reorder buffer entry contains (this is not the only possible solution) Type of instruction (branch, store, ALU, or load) Destination (none, memory address, register including other ROB entry) Value and its presence/absence Reservation station tags and true register tags are now ids of entries in the reorder buffer Reg. ren & Mult. Issue CSE 471 4 Example machine revisited (Fig 2.14 (3.29)) From I-unit Reorder buffer From memory & CDB To memory Fp registers
Reservation stations To CDB (CDB not shown) F-p units Reg. ren & Mult. Issue CSE 471 5 Need for 4 stages In Tomasulos solution 3 stages: issue, execute, write Now 4 stages: issue, execute, write, commit Dispatch and Issue Check for structural hazards (reservation stations busy, reorder buffer full). If one exists, stall the instruction and those following If dispatch possible, send source operand values to reservation station if the values are available in either the registers or the reorder buffer. Otherwise send tag. Allocate an entry in the reorder buffer (rename result register) and send its number to the reservation station (to be used as a tag on CDB) When both operands are ready, issue to functional unit Reg. ren & Mult. Issue CSE 471 6 Need for 4 stages (ced) Execute Write Broadcast on common data bus the value and the tag (reorder buffer number). Reservation stations, if any match the tag, and reorder buffer (always) grab the value. Commit When instr. at head of the reorder buffer has its result in the buffer it stores it in the real register (for ALU) or memory (for store). The reorder buffer entry (and/or physical register) is freed. Reg. ren & Mult. Issue CSE 471
yes yes yes Mul F0, F2, F4 yes yes yes 4 Sub F8, F6, F2 yes yes yes 5 Div F10, F0, F6 yes yes 6 Add F6,F8,F2 Name Busy
Add 1 no Add2 Add3 Mul1 no no no Mul2 yes Fm Div yes yes Reservation Stations Vj Vk (#3) (#1) Qj yes Qk F2( )
F4 ( ) F6(#6 ) F8 (#4) Cycle after mul has written its result and committed Still waiting: only 1 commit per cycle Register status F0 () yes F10 (#5) Reg. ren & Mult. Issue CSE 471 F12... 12 Entry # Reorder buffer Issue Execute Instruction Write result Commit yes 1 Load F6, 34(r2)
Div F10, F0, F6 yes yes 4 yes5 6 Add F6,F8,F2 Name Busy Add 1 no Add2 Add3 Mul1 no no no Mul2 yes Fm Div yes yes
Reservation Stations Vj Vk (#3) (#1) Qj yes yes Qk The next interesting event is completion of div; then commit of #5, then commit of #6 Register status F0 () F2( ) F4 ( ) F6(#6 ) Still waiting for #4, #5 to commit F8 () F10 (#5) Reg. ren & Mult. Issue CSE 471 F12... 14
Register renaming Physical Register file Use a physical register file (as an alternative to reservation station or reorder buffer) larger than the ISA logical one When instruction is decoded Give a new name to result register from free list. The register is renamed The mapping table is updated Give source operands their physical names (from mapping table) Reg. ren & Mult. Issue CSE 471 15 Register renaming File of physical registers Extra set of registers organized as a free list At decode: Rename the result register (get from free list; update mapping table). If none available, we have a structural hazard Note that several physical registers can be mapped to the same logical register (corresponding to instructions at different times; avoids WAW hazards) When a physical register has been read for the last time, return it to the free list Have a counter associated with each physical register (+ when a source logical register is renamed to physical register; - when instruction uses physical register as operand; release when counter is 0) Simpler to wait till logical register has been assigned a new name by a later instruction and that later instruction has been committed Reg. ren & Mult. Issue CSE 471 16 Example Before: add r3,r3,4 add r4,r7,r3 add r3, r2, r7
after Free list r37,r38,r39 . r2, r3, r4, r7 not renamed yet add r37,r3,4 add r38,r7,r37 add r39,r2,r7 At this point r3 is remapped from r37 to r39 When r39 commits, r37 will be returned to the free list Reg. ren & Mult. Issue CSE 471 17 Conceptual execution on a processor which exploits ILP Instruction fetch and branch prediction Corresponds to IF in simple pipeline Complicated by multiple issue (see in a couple of slides) Instruction decode, dependence check, dispatch, issue Corresponds (many variations) to ID Although instructions are issued (i.e., assigned to functional units), they might not execute right away (cf. reservation stations) It is at this point that one distinguishes between in-order and out-of-order superscalars Instruction execution Corresponds to EX and/or MEM (with various latencies) Instruction commit (for OOO only) Corresponds to WB but more complex because of speculation and out-of-order completion Reg. ren & Mult. Issue CSE 471
18 Multiple Issue Alternatives Superscalar (hardware detects conflicts) Statically scheduled (in order dispatch and hence execution; cf. (DEC)Alpha 21164, Sun processor in Niagara, IBM Cell Synergetic Processor) Dynamically scheduled (in order issue, out of order dispatch and execution; cf. MIPS 10000, IBM Power 4 and 5, Intel Pentium P6 microarchitecture, AMD K5 et al, (DEC)Alpha 21264, Sun UltraSparc etc.) VLIW EPIC (Explicitly Parallel Instruction Computing) Compiler generates bundles of instructions that can be executed concurrently (cf. Intel Itanium, lot of DSPs) Reg. ren & Mult. Issue CSE 471 19 Multiple Issue for Static/Dynamic Scheduling Issue in order Otherwise bookkeeping is complex (the old data flow machines could issue any ready instruction in the whole program; see also new grid machines such as WaveScalar and Trip) Check for structural hazards; if any stall Dispatch for static scheduling Check for data dependencies; stall adequately Can take forwarding into account Dispatch for dynamic scheduling Dispatch out of order (reservation stations, instruction window) Rename registers Requires possibility of dispatching concurrently dependent instructions (otherwise little benefit over static scheduling)
Reg. ren & Mult. Issue CSE 471 20 Impact of Multiple Issue on IF IF: Need to fetch more than 1 instruction at a time Simpler if instructions are of fixed length In fact need to fetch as many instructions as the issue stage can handle in one cycle Simpler if restricted not to overlap I-cache lines But with branch prediction, this is not realistic hence introduction of (instruction) fetch buffers and trace caches Always attempt to keep at least as many instructions in the fetch buffer as can be issued in the next cycle (BTBs help for that) For example, have an 8 wide instruction buffer for a machine that can issue 4 instructions per cycle Reg. ren & Mult. Issue CSE 471 21 Stalls at the IF Stage Instruction cache miss Instruction buffer is full Most likely there are stalls in the stages downstream Branch misprediction Instructions are stored in several I-cache lines In one cycle one I-cache line can be brought into fetch buffer A basic block might start in the middle (or end) of an I-cache line Requires several cache lines to fill the buffer The ID (issue-dispatch) stage will stall if not enough instructions in the fetch buffer Reg. ren & Mult. Issue CSE 471
22 Sample of Old and Current Micros Two instruction issue: Alpha 21064, Sparc 2, Pentium, Cyrix Three instruction issue: Pentium P6 (but 5 uops from IF/ID to EX; Pentium 4 and AMD K7 have 4 uops, Intel Core has 6 uops) Four instruction issue: Alpha 21164, Alpha 21264, IBM Power4 and Power5 (but somewhat restricted), Sun UltraSparc, HP PA-8000, MIPS R10000 Many papers written in mid-90s predicted 16-way issue by 2000. We are still at 4 in 2007! Reg. ren & Mult. Issue CSE 471 23 The Decode Stage (simple case: dual issue and static scheduling) ID = Dispatch + Issue Some authors would call this Issue + Dispatch! Look for conflicts between the (say) 2 instructions If one integer op. and one f-p op., only check for structural hazard, i.e. the two instructions need the same f-u (easy to check with opcodes ) RAW dependencies resolved as in single pipelines Note that the load delay (assume 1 cycle) can now delay up to 3 instructions, i.e., 3 issue slots are lost Reg. ren & Mult. Issue CSE 471 24 Decode in Simple Multiple Issue Case If instructions i and i+1 are fetched together and: Instruction i stalls, instruction i+1 will stall Instruction i is dispatched but instruction i+1 stalls (e.g., because of structural hazard = need the same f-u), instruction i+2 will not
advance to the issue stage. It will have to wait till both i and i+1 have been dispatched Reg. ren & Mult. Issue CSE 471 25 Alpha 21164 (@1995) 4-wide Branch history Integer Unit L1 I-cache Fetch/ Decode ITB F-p Unit IB Add/ Mult Add/ Branch Add IS Mult/ Div Br.Pred Int. Reg.
27 Alpha 21164 Front-end IF S0: Access I-cache Prefetcher fetches 4 instructions (16 bytes) at a time in one of two instruction buffers (IB). Each instruction has been predecoded (5 bits) IF-S1 : Branch Prediction Prefetcher contains branch prediction logic tested at this stage: 4 entry return stack; 2 bit/instruction in the I-cache + static prediction BTFNT ID-S2: Slotting Initial decode yields 0, 1, 2, 3 or 4 instruction potential issue; align instructions depending on the functional unit there are headed for. ID-S3. Check for issue: WAW and WAR (my guess) so that all instructions after S3 can execute successfully w/o stalls Reg. ren & Mult. Issue CSE 471 28 Alpha 21164 Restrictions in front-end In integer programs, only 2 arithmetic instructions can pass from S2 to S3 (structural hazards) This percolates back . In S0, only instructions in the same cache line can be fetched in a given cycle Too bad if you branch in the middle of a cache line Target branch address computed in S1 So if predict taken, you have one bubble. Good chance it will be amortized by other effects downstream S3 uses the equivalent of a (simplified) scoreboard
Reg. ren & Mult. Issue CSE 471 29 Alpha 21164 - Back-end Load latency : 2 cycles If instruction i is a load issued (leave S3) at time t and inst. i+1 depends on it: real bubble since inst i+1 will leave S3 at time t+2 (If instead of inst i+1 it were inst i+2 that were dependent, could we still have a real bubble?) Scoreboard does not know if cache hit or miss Speculates hit (why?) If wrong, known at S5, instructions already in the back-end not dependent on the load can proceed (scoreboard knows that). Others are aborted On branch mispredict (and precise) exceptions Known at S5. All inst. in program order after the branch are aborted (how can we enforce precise exceptions on the integer and memory pipelines?) Other possible structural hazards due to store buffers etc. (see later) What happens on a D-TLB miss? Reg. ren & Mult. Issue CSE 471 30 Dynamic Scheduling: Reservation stations, register renaming and reorder buffer Decode means: Dispatch to either A centralized instruction window common to all functional units (Pentium Pro, Pentium III and Pentium 4) Reservation stations associated with functional units (MIPS 10000, AMD K5-7, IBM Power4 and Power5) Rename registers (either via ROB or physical file) Note the difficulty when renaming in the same cycle R1 <- R2 + R3; R4 <- R1 + R5
Set up entry at tail of reorder buffer (if supported by architecture) Issue operands, when ready, to functional unit Reg. ren & Mult. Issue CSE 471 31 Stalls in Decode (issue/dispatch) Stage If there are decentralized reservation stations, there can be several instructions ready to be dispatched in same cycle to same functional unit Possibility of not enough reservation stations If there is a centralized instruction window, there might not be enough bus/ports to forward values to the execution units that need them in the same cycle Both instances are instances of structural hazards Conflicts are resolved via a scheduling algorithm Try and define critical instructions Reg. ren & Mult. Issue CSE 471 32 The Execute Stage Use of forwarding Use of broadcast bus or cross-bar or other interconnection network Well talk at length about memory operations (load-store) in subsequent lecture and when we study memory hierarchies Reg. ren & Mult. Issue CSE 471 33 The Commit Step (in-order completion)
Recall: need of a mechanism (reorder buffer) to: Complete instructions in order. This commits the instruction. Since multiple issue machine, should be able to commit (retire) several instructions per cycle Know when an instruction has completed non-speculatively,i.e., what to do with branches Know whether the result of an instruction is correct, i.e., what to do with exceptions Reg. ren & Mult. Issue CSE 471 34 Impact on Branch Prediction and Completion When a conditional branch is decoded: Save the current physical-logical mapping Predict and proceed When branch is ready to commit (head of buffer) If prediction correct, discard the saved mapping If prediction incorrect Flush all instructions following mispredicted branch in reorder buffer Restore the mapping as it was before the branch as per the saved map Note that there have been proposals to execute both sides of a branch using register shadows limited to one extra set of registers Reg. ren & Mult. Issue CSE 471 35 Exceptions Instructions carry their exception status When instruction is ready to commit No exception: proceed normally Exception Flush (as in mispredicted branch) Restore mapping (more difficult than with branches because the mapping is not saved at every instruction; this method can also be
used for branches) Reg. ren & Mult. Issue CSE 471 36 Summary: OOO flow of instructions Front-end Back-end Step Resources read Resources written or utilized Fetch PC Branch Predictor I-cache PC Instruction Buffer Decode-rename Instruction Buffer Register map Decode Buffer Register map ROB Dispatch Decode Buffer Register map
Register file (logical and physical) Reservation stations ROB Issue Reservation stations Functional units D-cache Execute Functional Units D-cache Reservation stations ROB Physical register file Branch Predictor Store Buffer etc Commit ROB Physical register file Store buffer ROB Logical register file Register map D-cache Reg. ren & Mult. Issue CSE 471 37 Pentium Family (slightly more details in H&P
Sec 2.10 (3.10 in 3rd)) Fetch-Decode unit Transforms up to 3 instructions at a time into micro-operations (uops) and stores them in a global reservation table (instruction window). Does register renaming (RAT = register alias table) Dispatch (aka issue)-execution unit Issues uops to functional units that execute them and temporarily store the results Depending on the implementation from 3 to 6 uops can be issued concurrently Retire unit Commits the instructions in order (up to 3 commits/cycle) Reg. ren & Mult. Issue CSE 471 38 Fetch/Decode/ Dispath unit; 8 stages of the pipe Execute Unit: 5 to 10 different funct. Units. EX takes from 1 to 32 cycles Retire unit: 3 stages Instruction pool The 3 units of the Pentium P6 are independent and communicate through the instruction pool Reg. ren & Mult. Issue CSE 471 39
L2 cache Bus interface L1 I-cache Fetch/Decode unit ITLB L1 D-cache MOB Br. pred DTLB Exec/Dispatch unit Agu MIS Fpu Decoder Iu RS Reg. map (RAT) MMX ROB RF Instr. Pool & retire unit
Reg. ren & Mult. Issue CSE 471 40 A Few More Details: Front-end Instruction Fetch (not in Pentium 4) 4 (mini) stages for IF 1. 2. 3. 4. Access BTB-BPB combination (what if a miss?). If hit and predicted taken, a bubble is generated Initiates I-cache access at address given by BTB (what if a miss?) Continues I-cache access Completes I-cache access and transfer 16 bytes in Decode buffer Instruction Decode 3 (mini) stages 1 and 2. Find end of first 3 instructions and break then down in ops only one branch decoded Some CISC instructions require the leftmost decoder (MIS) 3. Detect branches; can correct some situations (undetected unconditional branch for example) Reg. ren & Mult. Issue CSE 471 41
Front-end (ctd) Register renaming Enter ops in reservation stations and ROB Reg. ren & Mult. Issue CSE 471 42 Back-end ops can get executed when Operands are available The Execution Unit for that op is available A result bus will be available at completion No more important op should be executed So it takes two cycle (pipe stages) to do all that. Then: ops are executed Well see about load-store later Commit (aka retire) All ops from the same instruction should be retired together (done by marking beg. And end of instructions when put in the ROB) Reg. ren & Mult. Issue CSE 471 43 Limits to Hardware-based ILP Inherent lack of parallelism in programs Partial remedy: loop unrolling and other compiler optimizations Branch prediction to allow earlier issue and dispatch Complexity in hardware
Needs large bandwidth for instruction fetch (might need to fetch from more than one I-cache line in one cycle) Requires large register bandwidth (multiported register files ) Forwarding/broadcast requires long wires (long wires are slow) as soon as there are many units. Reg. ren & Mult. Issue CSE 471 44 Limits to Hardware-based ILP (ced) Difficulties specific to the implementation More possibilities of structural hazards (need to encode some priorities in case of conflict in resource allocations) Parallel search in reservation stations, reorder buffer etc. Additional state savings for branches (mappings), more complex updating of BPTs and BTBs. Keeping precise exceptions is more complex Reg. ren & Mult. Issue CSE 471 45
Structural Gap FY14-FY19. Current Gap Analysis. Reflects projection of current structural gap between revenues and expenses if no change occurs. Current Enrollment Plan; 0% Tuition/Unified Fee; 0% Appropriation; ATB at CPI and 1% for Merit; Medical @ 7% Trend for...
As long as Service Provider only provides access in the express manner set forth in the previous sentence, Enterprise Customer shall not obtain, record, transmit, or distribute any information contained in the Service Provider Online Information Security Policy in a...
Supporting Implementation Of The Re-Designed Curriculum Through The Library Learning Commons. Description: This session will be a discussion of how the Library Learning Commons is supporting the implementation of the re-designed curriculum by using a co-planning, co-teaching and co-assessing model.
Erosion by Water. Moving water is one of the most powerful agent of erosion. Erosion by water can have destructive results. Rill erosion develops when running water cuts small channels into the side of a slope. Gully erosion is when...
La crisi de l'escolàstica i el nominalisme de Guillem d'Occam3.2 Guillem d'Occam (1290-1349) Pàgina 148. El problema de la realitat: el nominalisme. La concepció dels universals. La realitat és individual (tant en l'ordre . ontològic. com .
Grammar-based. Evolutionary. Learning systems. Hybrid systems. The 20th century brought new life and possibilities to generative music in the form of computers. I don't plan to talk about any of these in great detail right now, but we'll cover a...
Anti-Epileptics - Lamotrigine (Lamictal), Oxcarbazepine(Trileptal), Carbamazepine (Tegretol), Topirimate (Topamax), Divalproex Sodium (Depakote) Mood Stabilizers -Lithium, Gabapentin (Neurontin) Virtually every psychotropic medication has been tried for someone with ASD somewhere at sometime - few + studies