Out-of-order execution optimization control method based on RISC-V server processor

By employing a two-level cascaded launch queue and priority scheduling, burst transmission merging, and global timestamp verification, the out-of-order execution of the RISC-V server processor is optimized, resolving latency and hardware overhead issues under high concurrency and high memory access density, and improving processor performance and scalability.

CN123114549APending Publication Date: 2026-11-17SHANGHAI MUNOXIHE NETWORK TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611243486.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing RISC-V server processors suffer from issues such as increased critical path latency, low execution unit utilization, high memory access latency, and high hardware overhead and power consumption under high concurrency and high memory access density workloads, which limit the scalability of out-of-order cores.

Method used

It adopts a two-level cascaded launch queue and priority scheduling mechanism. By pre-decoding and calculating the criticality score of the instruction, high-priority and ordinary-level instructions are allocated to different launch queues. Combined with burst transmission, memory access and computation execution are merged. Globally incrementing timestamps are used to ensure memory access consistency, replacing the traditional fully associative dependency detection circuit.

Benefits of technology

Significantly reduces the issue latency of critical path instructions, improves execution port utilization, alleviates memory wall performance bottlenecks, reduces hardware overhead, and enhances the scalability of out-of-order kernels.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114549A_ABST
    Figure CN123114549A_ABST
Patent Text Reader

Abstract

This invention discloses an out-of-order execution optimization control method based on a RISC-V server processor, belonging to the field of microprocessor architecture design technology. It includes: S1, pre-decoding RISC-V instructions, calculating instruction criticality scores, and classifying them into high-critical and ordinary critical levels; S2, calling the built-in register renaming module to eliminate instruction name dependencies, and sending the graded instructions into two-level cascaded issue queues respectively; S3, performing deterministic issuance based on instruction priority and port constraints, and using bounded queues for backpressure control of pipeline flow; S4, coordinating memory access and computation timing, and achieving pipeline overlap through burst transfer merging; S5, ordering global timestamp submission and verifying memory access consistency. Using this method, critical instruction waiting latency can be reduced, execution port utilization can be improved, memory access latency can be masked, and verification hardware overhead can be reduced, thereby improving the throughput and architectural scalability of the server RISC processor.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of microprocessor architecture design technology, and in particular to an out-of-order execution optimization control method based on a RISC-V server processor. Background Technology

[0002] As the RISC-V open-source architecture penetrates into server computing scenarios, high-performance out-of-order execution kernels have become a core technological bottleneck in the industry. Currently, most mainstream open-source and commercial RISC-V out-of-order processor kernels adopt the classic architecture of unified issue queues + fully dynamic scheduling. For example, the NaxRiscv RV64GC out-of-order core is configured with 32 unified issue queues and 64 ROBs (Reorder Buffers). While this can meet performance requirements in embedded scenarios, its architectural shortcomings become increasingly apparent when facing server-level high-concurrency, high-memory-density workloads.

[0003] First, under a unified launch queue, critical path instructions and ordinary instructions compete for launch ports out of order. In low-latency sensitive workloads such as databases and microservices, critical memory access and branch instructions are easily blocked by a large number of computational instructions, leading to increased single-threaded critical path latency and persistently high pipeline vacuolation rates. Second, purely dynamic launch scheduling lacks resource budget control, resulting in frequent execution port conflicts, high overhead for instruction replay and pipeline rollback, and difficulty in maintaining high average utilization of execution units, thus limiting the throughput gain of superscalar architectures. Third, memory access instruction scheduling is independent of the memory controller, and memory access latency cannot be effectively masked by computational operations. In server-level scenarios with high memory access density and high memory access latency, the "memory wall" causes execution units to be idle for extended periods, resulting in significant waste of computing power. Fourth, instruction dependency detection and memory order guarantees rely on fully associative comparison circuits. As the launch width and ROB depth increase, hardware overhead and power consumption increase quadratically, limiting the expansion of out-of-order cores towards wider launch and deeper queues in server-level applications. Summary of the Invention

[0004] The purpose of this invention is to provide an out-of-order execution optimization control method based on a RISC-V server processor, thereby solving the aforementioned technical problems.

[0005] To achieve the above objectives, this invention provides an out-of-order execution optimization control method based on a RISC-V server processor, comprising the following steps: S1. Pre-decode the RISC-V instructions fetched from the instruction output, calculate the criticality score of each instruction and classify them into high critical level and ordinary level; S2. Call the processor's built-in registers to rename the hardware module, completing the mapping from logical registers to physical registers to eliminate instruction name conflicts; after renaming, allocate high-critical-level instructions to the high-priority issue queue and ordinary-level instructions to the ordinary-priority issue queue, forming a two-level cascaded issue queue scheduling structure. S3. Deterministic launch scheduling is performed based on instruction priority and execution port resource constraints, and pipeline traffic is controlled through a bounded queue backpressure mechanism; S4. The execution timing of memory access instructions and computation instructions is scheduled, and memory access is merged with burst transmission to achieve pipeline overlap between memory access latency and computation execution. S5. Assign a globally incrementing timestamp to each instruction, complete the instruction submission in the order of timestamps, and ensure memory access consistency through timestamp verification.

[0006] Preferably, in S1, the formula for calculating the criticality score of each instruction is: ; in, Score the criticality of the instruction. Instruction type weight; For data dependency depth; Memory access strength; , , Weighting coefficients pre-calibrated for server load; Set grading threshold If the instruction scores If the value is high, the corresponding instruction will be classified as a high-critical instruction; otherwise, it will be classified as a normal instruction.

[0007] Preferably, in S2, the scheduling rules for the two-level cascaded launch queue structure are as follows: Within each clock cycle, ready instructions from the high-priority transmit queue occupy the transmit ports first, and the remaining idle ports are filled by ready instructions from the normal-priority transmit queue. The total number of instructions issued per cycle shall not exceed the total number of processor issue ports, and the number of instructions issued per cycle in the high-priority queue under the dual-issue architecture shall not exceed 1.

[0008] Preferably, in S3, the deterministic launch scheduling specifically includes: The formula for calculating the launch priority of each ready command is as follows: ; in, To prioritize the overall launch of commands, The number of clock cycles that an instruction has been waiting in the launch queue. The launch scheduler selects commands to launch from high to low priority, while simultaneously satisfying the resource total constraint formula for various execution ports: ; in, The port type is used for execution, including integer arithmetic ports, floating-point arithmetic ports, and memory access ports; For the first The launch command is for the first Port usage; For the first Total number of available class ports; When the bounded queue at the execution phase entry point is filled to more than 80% of the total queue storage depth, a backpressure signal is sent to the preceding issue phase to suspend the issuance of the corresponding type of instruction.

[0009] Preferably, in S4, the pipeline overlap between memory access latency and computation execution is achieved as follows: After address resolution, the LSU (Load-Store Unit) merges consecutive address memory access requests to generate a burst transfer instruction, which is then sent to the FSM (Finite State Machine) and AXI (Advanced Extensible Interface) memory controllers. During the FSM AXI memory controller's processing cycle of burst transfers, it schedules computation instructions without data dependencies to fill idle time slots in the arithmetic unit. After the memory access data is returned, it is directly sent to the corresponding dependent instructions through the data bypass network without waiting for ROB to be committed.

[0010] Preferably, in S5, memory access consistency is ensured through timestamp verification, specifically as follows: When instructions are written to the ROB, a globally incrementing timestamp is assigned, and each timestamp corresponds to a unique original execution order of the program. The submitting unit submits the ready command in ascending order of time; For load instructions targeting the same memory address addr, the following conditions must be met. At that time, it is allowed to submit and read memory data, where, For the timestamp of the loading instruction, The timestamp of the latest committed store instruction at address addr.

[0011] Preferably, the RISC-V instructions include basic integer instructions, floating-point instructions, and RISC-V standard vector extension instructions, and all instruction types are included in the criticality classification and cascaded launch scheduling scope.

[0012] Preferably, the burst transmission parameters match the server's DDR (Double Data Rate) memory AXI bus protocol, the length of a single burst transmission is fixed at 8 data cycles, and the bus width is adapted to a 64-bit server memory channel.

[0013] Preferably, the total storage depth of the two-level cascaded launch queues is not less than 32 items, and the high-priority queue depth accounts for 25% to 35% of the total storage depth.

[0014] Preferably, the FSM AXI memory controller used to perform burst transmission scheduling uses FSM to build the AXI memory mapping interaction interface, which is compatible with multiple types of non-volatile memory chips such as MRAM (Magnetoresistive Random Access Memory), FRAM (Ferroelectric Random Access Memory), and ReRAM (Resistive Random Access Memory).

[0015] Therefore, the out-of-order execution optimization control method based on the RISC-V server processor described above has the following beneficial effects: 1. A two-level cascaded launch queue and priority scheduling mechanism are adopted to solve the resource contention defects of the traditional single launch queue. Based on the prototype test of the Zedboard FPGA server RISC-V kernel, the average launch latency of critical path instructions is reduced by more than 25%, which significantly improves the single-threaded processing performance of low-latency services such as databases and microservices.

[0016] 2. By introducing deterministic resource scheduling and bounded queue backpressure mechanisms, the defects of frequent port conflicts in pure dynamic scheduling are eliminated. FPGA synthesis and timing simulation results show that the average utilization of processor execution ports is increased by 15%~20%, pipeline cavitation rate is significantly reduced, and the overall instruction throughput is improved simultaneously.

[0017] 3. By using burst memory access merging and compute-memory parallel management strategies, idle memory access cycles are masked; hardware tests show that the compute-memory overlap rate is improved by more than 30%, effectively alleviating the memory wall performance bottleneck in server big data load scenarios.

[0018] 4. The memory sequence verification is completed using a globally incrementing timestamp, replacing the traditional fully associative detection circuit. According to the comprehensive data of FPGA resources, the hardware overhead of the memory consistency verification module is reduced by about 40%, the out-of-order kernel scalability is enhanced, and it can be adapted to high-bandwidth server RISC-V processor architectures such as 4-issue and 8-issue.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a five-stage pipeline architecture diagram of the out-of-order execution optimization control method based on the RISC-V server processor of the present invention. Figure 2 This is a block diagram illustrating the implementation of burst memory access merging, FSM AXI memory controller, and memory access computation overlap in this invention. Figure 3 This is a schematic diagram of the global timestamp ordered submission and memory consistency verification logic of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0022] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] like Figures 1-3 As shown, the out-of-order execution optimization control method based on RISC-V server processors provided by this invention is applicable to the design of general-purpose server kernels based on RISC-V architecture, such as cloud servers, database servers, and high-performance computing servers. It can also be used in FPGA prototype verification and open-source processor microarchitecture development scenarios.

[0025] In this embodiment, the hardware carrier is a dual-issue RV64GCV server-grade RISC-V superscalar out-of-order processor core. The overall design is based on the Xilinx Zedboard FPGA platform to complete hardware synthesis, timing simulation and prototype verification. The processor core is an industrial general out-of-order architecture, natively integrating a register renaming module, ROB, LSU, multiple execution ports and an AXI memory controller based on FSM, referred to as FSM AXI memory controller, without the need for additional dedicated hardware circuits.

[0026] Kernel hardware parameters: The total storage depth of the two-level cascaded transmit queue is 32 items, of which 10 items are high-priority queues (accounting for 31.25% of the total depth, falling within the 25%~35% parameter range) and 22 items are ordinary priority queues; it is configured with 2 integer operation ports, 1 floating-point operation port, and 1 memory access LSU port; it is equipped with a 64-bit wide AXI memory bus, and a single burst transmission is fixed at 8 data cycles; the FSM AXI memory controller is compatible with multiple types of non-volatile memory chips such as MRAM, FRAM, and ReRAM, and fully meets the hardware constraints of claims 8 and 10.

[0027] This method's pipeline consists of five progressive stages: pre-decoding instruction hierarchical processing, register renaming and queue allocation, resource deterministic issue scheduling, memory access-computation pipeline overlap, and timestamp-ordered submission. Each stage specifically includes: S1. Pre-decode the fetched RISC-V instructions, calculate the criticality score of each instruction and classify them into high-criticality levels, specifically including: Hardware support: The instruction fetch unit reads a continuous instruction stream from the on-chip 16KB instruction cache and sends it to the independent pre-decoder; the pre-decoder has a built-in PC history dependency table, which automatically calculates the data dependency depth D of a single instruction, and at the same time recognizes the full range of RISC-V instructions, covering basic integer, floating point, and RV64GCV vector extension instructions.

[0028] Keyness score calculation process: Pre-decoder parses instruction opcodes to determine instruction type and weight. The output of the statistical command depends on the link length. If it is an LSU memory access load / store instruction, then calculate the memory access intensity. If it is not a memory access instruction, then Substitute into the standard formula for calculating criticality score: ; in, Score the criticality of the instruction. Instruction type weight; For data dependency depth; Memory access strength; , , These are pre-defined weighting coefficients for server load; this embodiment uses fixed weights for cloud computing and database server load. =0.4, =0.4, =0.2; Instruction type weighting standardization: If it is a memory access instruction, then =3; if it is a branch instruction, then =2; if it is a floating-point / vector extension instruction. =2.5; if it is a normal integer instruction =1; Set the normalized ratio of expected memory access latency to 10 baseline clock cycles; set a global hierarchical threshold. =2.0.

[0029] Grading determination logic: If the instruction calculates a score, then the instruction score is... ≥ If the value is 2.0, the instruction is marked as a high-critical instruction; otherwise, it is marked as a normal instruction. The hierarchical identifier is transmitted to the back-end register renaming module synchronously with the instruction.

[0030] Example Calculation: A single load memory access instruction, data dependency depth Normalization intensity of visits Substituting into the formula, we get for: ; This load memory access instruction is then classified as a high-critical instruction. A dependency-free integer addition instruction, with a data dependency depth Normalization intensity of visits Substituting into the formula, we get for: ; This independent integer addition instruction is then classified as a normal instruction.

[0031] S2. Rename the hardware module by calling the processor's built-in registers, complete the logical-physical register mapping, and allocate a two-level issue queue. Specifically, this includes: Hardware basic description: The registers are renamed to the native standard hardware units of this out-of-order RISC-V kernel. The register alias table R and the physical register file are integrated. The core function is to eliminate the three types of register name conflicts: write-after-read, write-after-write, and read-after-write, so as to ensure that instructions can be independently scheduled out of order. This invention does not modify the renaming hardware itself, but only reuses its output instruction stream.

[0032] Instruction distribution: Instructions that complete register mapping and name removal carry hierarchical tags generated by S1. The distribution unit distributes instructions according to the tags: high-critical instructions are sent to a 10-item deep high-priority launch queue, and ordinary instructions are sent to a 22-item deep ordinary-priority launch queue. The two-level queues are physically isolated, forming a two-level cascaded launch scheduling structure with a total storage depth of 32 items.

[0033] Single-cycle scheduling rules: In each hardware clock cycle, the launch scheduler prioritizes scanning ready instructions in the high-priority queue; in the dual-launch architecture, the high-priority queue launches a maximum of 1 instruction per cycle, and the remaining idle launch port is allocated to ready instructions in the ordinary queue; the total number of launch instructions in a single cycle does not exceed the total number of processor launch ports, with a maximum of 2 instructions in this embodiment, balancing low latency for critical instructions with fairness in scheduling ordinary instructions, and avoiding starvation.

[0034] S3. Deterministic launch scheduling is performed based on instruction priority and port constraints, with bounded queue backpressure control of pipeline traffic, specifically including: Launch priority calculation: The launch scheduler reads the criticality score of ready instructions in the queue. Number of clock cycles already waited Substitute into the priority calculation formula: ; in, This formula assigns a comprehensive priority to commands; based on this formula, the higher the inherent criticality of a command and the shorter its waiting period, the higher its priority. The higher the value, the longer the low-priority instruction will wait. Increase the priority, and the underlying mechanism will prevent long-term blocking of instructions.

[0035] Perform port resource constraint verification: The send scheduler performs verification based on priority. Ready instructions are sorted from highest to lowest order, and corresponding execution ports are allocated sequentially. The resource total constraint determination formula is executed synchronously. ; in, The port type is used for execution, including integer arithmetic ports, floating-point arithmetic ports, and LSU memory access ports. For the first The launch command is for the first Port usage; For the first The total number of available ports; the hardware limit in this embodiment is [number missing]. At this time, integer ports = 2, floating-point ports = 1, and memory access ports = 1. The port occupancy in any period must not exceed the maximum hardware capacity.

[0036] Bounded queue backpressure mechanism: Each of the three types of execution unit entry points, namely integer, floating point, and LSU, is configured with an independent synchronous FIFO bounded queue with a single queue depth of 8 items; the hardware counts the queue filling entries in real time, and when the filling entries are ≥7 (filling degree reaches 80%), a hardware backpressure signal is sent to the front-level issue scheduling unit to suspend the issuance of the same type of instruction, prevent queue overflow from causing pipeline flushing, and reduce instruction replay overhead.

[0037] S4, scheduling memory access and computation execution timing, achieves memory access-computation pipeline overlap through burst transfer merging, specifically including: Underlying hardware support: The LSU receives memory access commands from the transmit port and interfaces with the FSM AXI memory controller at the back end; the controller divides the idle, read / write low 16 bits, read / write high 16 bits, and burst transmission into independent timing states, supports standard AXI burst transmission, has a bus width of 64 bits, and can simultaneously interface with DDR memory, MRAM / FRAM / ReRAM commercial non-volatile memory chips.

[0038] Burst merging logic: After the LSU completes memory access address resolution, it automatically identifies multiple load / store instructions with consecutive addresses and merges them into a single burst transfer request, which is then sent to the FSM AXI memory controller to reduce the protocol overhead caused by multiple address handshakes and bus interactions; a single burst transfer has a fixed 8 data cycles, which perfectly matches the bus parameters of claim 8.

[0039] Overlapping scheduling of memory access and computation: The FSM AXI memory controller has an access delay of dozens of clock cycles when processing a single burst transfer. During this idle time slot, the send scheduler continuously sends integer and floating-point computation instructions without data dependency to the computing port, fully filling the hardware idle cycle and masking the problem of idle computing power caused by the "memory wall".

[0040] Fast data bypass forwarding: The memory data read by burst is directly transmitted to the pending instructions that depend on the data through a dedicated data bypass network, without waiting for the instructions to be sent to ROB to complete the submission, thus shortening the overall latency of the data dependency link.

[0041] S5. Allocate globally incrementing timestamps, submit them in order, and ensure memory access consistency through timestamp verification. Specifically, this includes: Global timestamp allocation: When each instruction is sent from the issue port to the ROB, the kernel's global hardware counter automatically assigns a globally incrementing timestamp. The timestamp generation order corresponds strictly one-to-one with the original program's instruction fetching and execution order, with no duplicates or skips.

[0042] Ordered commit control: ROB has a built-in commit controller that outputs ready instructions marked "execution completed" only in ascending order of timestamps, ensuring that out-of-order execution does not destroy the original program semantics, while also supporting precise exception handling.

[0043] Memory consistency check rules: For load instructions at the same memory address `addr`, the timestamps of the instructions executed before commit are compared, and only those that meet the requirements are considered consistent. Only when this condition is met is a load instruction allowed to be submitted and memory data read; among which The timestamp of the current loading instruction. This is the timestamp of the last completed store instruction at this address; if the condition is not met, the submission is postponed until the preceding store instruction is completed, strictly following the RISC-VRVWMO memory model specification.

[0044] The test platform for this solution is a Xilinx Zedboard FPGA with a dual-issue RV64GCV RISC-V server core and a standard cloud database load. Compared with the traditional single-issue queue RISC-V out-of-order core, the actual measured performance is shown in Table 1.

[0045] Table 1: Measured Data Table

[0046] As shown in Table 1, under low-latency loads in databases and cloud services, the average waiting latency for high-critical memory access and branch instructions is reduced by 26%; the average hardware utilization of integer and floating-point execution ports is increased by 18%, and the pipeline cavitation rate is significantly reduced; the Burst memory access and computation overlap mechanism improves the memory access latency masking rate by 32%, effectively alleviating the memory wall bottleneck in server scenarios; the timestamp linear check replaces the traditional fully associative dependency detection circuit, reducing the memory order-related hardware LUT overhead by about 41%, and the architecture can be scaled up to 4-issue and 8-issue wide-range server processor cores.

[0047] Furthermore, the FSM AXI memory controller of this invention adopts a segmented FSM state machine architecture. After hardware synthesis, the LUT resource consumption is reduced by 30% compared with the Xilinx official native memory controller, which is consistent with the test results in the references. The controller hardware interface is compatible with multiple commercial NVM chips such as SPIReRAM, parallel FRAM / MRAM, etc. The RISC-V custom fast-store class extension instructions can be seamlessly integrated into the entire process of S1 to S3 graded launch scheduling without modifying the pipeline core control logic.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for optimizing and controlling out-of-order execution based on a RISC-V server processor, characterized in that, Includes the following steps: S1. Pre-decode the RISC-V instructions fetched from the instruction output, calculate the criticality score of each instruction and classify them into high critical level and ordinary level; S2. Call the processor's built-in registers to rename the hardware module, completing the mapping from logical registers to physical registers to eliminate instruction name conflicts; after renaming, allocate high-critical-level instructions to the high-priority issue queue and ordinary-level instructions to the ordinary-priority issue queue, forming a two-level cascaded issue queue scheduling structure. S3. Deterministic launch scheduling is performed based on instruction priority and execution port resource constraints, and pipeline traffic is controlled through a bounded queue backpressure mechanism; S4. The execution timing of memory access instructions and computation instructions is scheduled, and memory access is merged with burst transmission to achieve pipeline overlap between memory access latency and computation execution. S5. Assign a globally incrementing timestamp to each instruction, complete the instruction submission in the order of timestamps, and ensure memory access consistency through timestamp verification.

2. The out-of-order execution optimization control method based on a RISC-V server processor according to claim 1, characterized in that: In S1, the formula for calculating the criticality score of each instruction is as follows: ; in, Score the criticality of the instruction. Instruction type weight; For data dependency depth; Memory access strength; , , Weighting coefficients pre-calibrated for server load; Set grading threshold If the instruction scores If the value is high, the corresponding instruction will be classified as a high-critical instruction; otherwise, it will be classified as a normal instruction.

3. The out-of-order execution optimization control method based on a RISC-V server processor according to claim 1, characterized in that: In S2, the scheduling rules for the two-level cascaded launch queue structure are as follows: Within each clock cycle, ready instructions from the high-priority transmit queue occupy the transmit ports first, and the remaining idle ports are filled by ready instructions from the normal-priority transmit queue. The total number of instructions issued per cycle shall not exceed the total number of processor issue ports, and the number of instructions issued per cycle in the high-priority queue under the dual-issue architecture shall not exceed 1.

4. The out-of-order execution optimization control method based on a RISC-V server processor according to claim 1, characterized in that: In S3, deterministic launch scheduling specifically refers to: The formula for calculating the launch priority of each ready command is as follows: ; in, To prioritize the overall launch of commands, The number of clock cycles that an instruction has been waiting in the launch queue. The launch scheduler selects commands to launch from high to low priority, while simultaneously satisfying the resource total constraint formula for various execution ports: ; Among them, The port type is used for execution, including integer arithmetic ports, floating-point arithmetic ports, and memory access ports; For the first The launch command is for the first Port usage; For the first Total number of available class ports; When the bounded queue at the execution phase entry point is filled to more than 80% of the total queue storage depth, a backpressure signal is sent to the preceding issue phase to suspend the issuance of the corresponding type of instruction.

5. The out-of-order execution optimization control method based on a RISC-V server processor according to claim 1, characterized in that: In S4, the pipeline overlap between memory access latency and computation execution is achieved as follows: After address resolution, the LSU merges consecutive memory access requests into a burst transfer instruction and sends it to the FSMAXI memory controller. During the FSM AXI memory controller's processing cycle of burst transfers, it schedules computation instructions without data dependencies to fill idle time slots in the arithmetic unit. After the memory access data is returned, it is directly sent to the corresponding dependent instructions through the data bypass network without waiting for ROB to be committed.

6. The out-of-order execution optimization control method based on a RISC-V server processor according to claim 1, characterized in that: In S5, timestamp verification ensures memory access consistency, specifically as follows: When instructions are written to the ROB, a globally incrementing timestamp is assigned, and each timestamp corresponds to a unique original execution order of the program. The submitting unit submits the ready command in ascending order of time; For load instructions targeting the same memory address addr, the following conditions must be met. At that time, it is allowed to submit and read memory data, where, For the timestamp of the loading instruction, The timestamp of the latest committed store instruction at address addr.

7. The out-of-order execution optimization control method based on a RISC-V server processor according to claim 1, characterized in that: RISC-V instructions include basic integer instructions, floating-point instructions, and RISC-V standard vector extension instructions. All instruction types are included in the criticality classification and cascaded launch scheduling scope.

8. The out-of-order execution optimization control method based on a RISC-V server processor according to claim 5, characterized in that: The burst transmission parameters are matched with the server DDR memory AXI bus protocol, the length of a single burst transmission is fixed at 8 data cycles, and the bus width is adapted to the 64-bit server memory channel.

9. The out-of-order execution optimization control method based on a RISC-V server processor according to claim 1, characterized in that: The total storage depth of the two-level cascaded launch queues is no less than 32 items, and the high-priority queue depth accounts for 25% to 35% of the total storage depth.

10. The out-of-order execution optimization control method based on a RISC-V server processor according to claim 1, characterized in that: The FSM AXI memory controller used to perform burst transfer scheduling uses FSM to build the AXI memory mapping interaction interface, which is compatible with multiple types of non-volatile memory chips such as MRAM, FRAM, and ReRAM.