Patent Application
20160328237
1. Field of the Invention
The present invention relates in general to processing engines, and more specifically to a system and method of reducing the load-store collision penalty in a speculative out of order processing engine.
2. Description of the Related Art
A processing engine, such as a microprocessor or the like, executes the instructions of an instruction set architecture, such as the x86 instruction set architecture or the like. In many such engines, the instructions of the instruction set architecture, often referred to as macroinstructions, are first translated into microinstructions (or micro-operations or “μops”) that are issued to a reservation stations module that dispatches the instructions to the execution units. The microinstructions are more generally referred to herein simply as “instructions.” The instructions are also issued to a reorder buffer which ensures in-order retirement of the instructions.
An out-of-order (O-O-O) scheduler is widely used in processor design and provides an important distinction between high performance processor and others. In an O-O-O scheduler, each instruction is dispatched based on dependency, which is when the instructions use the same register as source and destination. Yet, the dependency of some instructions, such as load and store instructions, is difficult to recognize. This is because the dependency is not caused by the same register, but instead by the same address which is not known by the scheduler at the schedule stage. So, one common method is to speculatively assume that the load and store instructions do not have any collisions. When a collision is unfortunately detected afterwards, the result is incorrect, the pipeline is flushed, and the instructions are dispatched again. When the speculative dispatch of an instruction is incorrect, the flushing and re-dispatch of the instruction introduces a significant penalty.
A microprocessor according to one embodiment includes a load pipeline, a scheduler, an address generation unit, and a load-store queue. The load pipeline includes multiple stages which include at least one operand stage and two or more execution stages. The scheduler dispatches load instructions to the at least one operand stage for execution by the execution stages. The load instructions include a speculatively dispatched load instruction. The address generation unit provides a load instruction virtual address for the speculatively dispatched load instruction before the speculatively dispatched load instruction has progressed to the execution stages. The load-store queue asserts a clear signal to invalidate the speculatively dispatched load instruction when a match occurs between the load instruction virtual address and a store instruction virtual address of at least one previously dispatched store instruction in which corresponding store data has not yet been determined.
The clear signal invalidates the speculatively dispatched load instruction in the event of a collision, such as when a match occurs between the load instruction virtual address and a store instruction virtual address of at least one previously dispatched store instruction. In this manner, the scheduler, which is otherwise configured to schedule dispatch of instructions that are dependent upon the speculatively dispatched load instruction, may instead not prematurely dispatch the dependent instructions when the clear signal is asserted.
In one embodiment, the load pipeline is configured to assert a load valid signal when the speculatively dispatched load instruction has progressed to a selected execution stage. Kill logic is provided that prevents the load valid signal from being detected by the scheduler when the clear signal is asserted. Broadcast logic may be provided to receive and broadcast the load valid signal to the scheduler when asserted, except when the clear signal is asserted to invalidate the speculatively dispatched load instruction. The broadcast logic may include kill logic that prevents the broadcast logic from broadcasting the load valid signal when the clear signal is asserted.
The load-store queue may include multiple entries, each for comparing the load instruction virtual address of the speculatively dispatched load instruction with one or more store instruction virtual addresses. Valid logic and qualify logic may be provided for each entry to ensure a corresponding store instruction virtual address corresponds to a store instruction that is dispatched early than the speculatively dispatched load instruction and whose corresponding store data has not yet been determined. Each entry may assert a preliminary clear signal, and OR logic may be provided to assert the clear signal when any one or more of the preliminary clear signals are asserted.
A load-store collision detection system is disclosed for a speculative out of order processing engine. The processing engine includes a scheduler that dispatches instructions to multiple instruction pipelines, in which the instruction pipelines include a load pipeline that provides a load valid signal when a speculatively dispatched load instruction is executing. The load-store collision detection system includes comparator logic, broadcast logic, and kill logic. The comparator logic asserts a clear signal when a load instruction virtual address of the speculatively dispatched load instruction matches at least one store instruction virtual address of at least one previously dispatched store instruction whose corresponding store data is not ready yet. The broadcast logic broadcasts the load valid signal to the scheduler to enable dispatch of any instructions dependent upon the speculatively dispatched load instruction. The kill logic invalidates the load valid signal when the clear signal is asserted before or coincident with the load valid signal.
The kill logic may be incorporated within the broadcast logic or the scheduler or any suitable combination of both. The load-store collision detection system may include a memory for storing one or more store instruction virtual addresses.
A method of reducing load-store collisions in a speculative out of order processing engine includes providing a store instruction address for each of at least one previously dispatched store instruction whose corresponding data is not ready yet, speculatively dispatching a load instruction to a load pipeline, determining a load instruction address for the speculatively dispatched load instruction before said speculatively load instruction is executed, comparing the load instruction address with the store instruction address of each of the at least one previously dispatched store instruction, asserting a clear signal when the load instruction address matches the store instruction address of the at least one previously dispatched store instruction, asserting a load valid signal for the speculatively dispatched load instruction while being executed, and invalidating the load valid signal when the clear signal is also asserted.
The method may include broadcasting the load valid signal to each queue of a scheduler, and suppressing the broadcasting of the load valid signal when the clear signal is also asserted. The method may include determining the validity of the store instruction address of each of the at least one previously dispatched store instruction. The method may include validating and qualifying the store instruction address of each of the at least one previously dispatched store instruction. The method may include comparing the load instruction address with multiple store instruction address and asserting a corresponding one of multiple preliminary clear signals for each match, and asserting the clear signal when at least one of the preliminary clear signals is asserted. The method may include asserting a corresponding one of the preliminary clear signals only when a corresponding store instruction address is valid and qualified.
The benefits, features, and advantages of the present invention will become better understood with regard to the following description, and accompanying drawings where:
The inventors have recognized the penalty associated with a load-store collision caused by speculative dispatch of a load instruction within a processing engine. They have therefore developed a system and method of detecting load-store collisions before the load is dispatched for execution. The system and method further squashes, or otherwise suppresses, the dispatch valid of the load instruction to prevent the issuance of additional instructions which depend upon the speculatively dispatched load instruction. Since the potentially dependent instructions are not dispatched prematurely, the pipeline need not be flushed and the dependent instructions need not be replayed. In this manner, the penalty associated with the speculatively dispatched load instruction is reduced or otherwise minimized. A load-store (Lo-St) queue (LoStQ) structure is incorporated into the instruction pipeline which detects a collision between the load and any store instruction that is not ready to complete. A store instruction that is not ready to complete means that the address portion (STA) has been determined, but the data portion (STD) has not yet been determined, so that the store instruction is considered temporarily “LoSt.” The LoStQ detects this collision and issues a clear signal back to kill broadcast of the dispatch valid of the load instruction to scheduler queues holding instructions for dispatch. The clear signal suppresses the dispatch of additional instructions that depend upon the speculatively dispatched load to improve performance efficiency.
The execution units 112 may include one or more integer execution units, such as an integer arithmetic/logic unit (ALU) 114 or the like, one or more floating point execution units 116, such as including a single-instruction-multiple-data (SIMD) execution unit such as MMX and SSE units or the like, a memory order buffer (MOB) 118, etc. The MOB 118 generally handles memory type instructions to a system memory 120, such as including a load instruction execution pipe 117 and a similar store instruction execution pipe 119. The system memory 120 may be interfaced with the MOB 118 via a data cache (e.g., L2 data cache, not shown) and a bus interface unit (BIU, not shown). The execution units 112 provide their results to the ROB 110, which ensures in-order retirement of instructions.
In one embodiment, the reservation stations 108 includes multiple RS queues, in which each RS queue schedules and dispatches corresponding issued instructions to corresponding execution units 112 when the instructions are ready to be executed. In general, a separate RS queue is provided for each execution unit 112. For example, an RS Q1 122 is provided for the integer execution unit 114 and an RS Q2 124 is provided for the floating point execution unit 116. In one embodiment, a LD RS Q 126 provides load instructions to the load pipeline 117, and a separate ST RS Q 128 provides store instructions to the store pipeline 119. Each RS Q of the reservation stations 108 may alternatively be referred to as a scheduler that includes schedule logic or the like (not shown) that schedules dispatch of issued instructions to the corresponding execution unit 112.
The O-O-O instruction sequence 250 also begins with the STA micro-operation (μop) 202. In this case, however, the LD μop 208 is dispatched out of order (out of program order) and before the STD μop 206 since the data D operand is not yet known. The LD μop 208 loads the data X currently stored at the memory location 204 at address ADD into a selected storage location, in which data X is, for practical purposes, not the same as data D. Since the data D was not yet available, the LD μop 208 loads the incorrect data X. As shown by arrow 210, the speculative execution of the LD μop 208 is reported back to other dependent μops depending on the LD μop 208 at the reservation stations 108, and these dependent μops may be dispatched into the pipeline of a corresponding execution unit to retrieve operands prior to execution. Eventually as shown by arrow 212, the results of the load μop 208 are reported back to the ROB 110, which includes a “MISS WHY” routine or the like that detects the incorrect load result. In response, the ROB 110 flushes the load pipeline 117 to remove the LD μop 208, and also flushes any other execution pipeline processing any instructions dependent upon the LD μop 208 that have been dispatched. Also, as indicated by arrow 216, the LD μop 208 and the corresponding dependent instructions must ultimately be replayed after the STD μop 206 is executed to retrieve the correct data D. The flushing of the execution pipelines and the replay of the load instruction and any dependent instructions has a negative impact on performance of the microprocessor 100.
Stage D 302 is an issue stage which is common to each the pipelines of each of the execution units 112, in which instructions are issued from the RAT 106 to the scheduler RS Q's within the reservation stations 108. Stages Q, R and A(I) 304, 306 and 308 are the operand stages in which a selected load instruction is dispatched for execution and the operands for the selected instruction are determined prior to actual execution. The remaining stages B, C, E1 and E2 are the load execution stages for executing the dispatched load instruction. When a valid load instruction is dispatched from the LD RS Q 126 at stage Q, a dispatch valid signal DV(Q) is asserted. Each stage generates corresponding dispatch tags DT(Q), DT(R), DT(I), DT(B), DT(C) as the load instruction propagates through the load pipeline 117.
In stage A(I) 308, select logic 320 selects from among possible sources (e.g., the sources may be a register, one of several types of constants, a memory address, etc.) of the operands for the load instructions for determining both a first source SRCA and a second source SRCB, which are provided to respective inputs of an address generation unit (AGU) 322. As an example of the select logic 320, a first multiplexer (MUX) 319 selects from among the possible sources to provide the first source SRCA, and a second MUX 321 selects from among the possible sources to provide the second source SRCB, both provided to inputs of the AGU 322. The AGU 322 outputs a corresponding load instruction virtual address (LDVA) for accessing the system memory 120. It is understood that LDVA may be converted to a physical address for accessing the system memory 120. It should be noted that LDVA may be just part of the virtual address of the load instruction. For example, a 12-bit AGU 322 calculates [11:0] LDVA, which is identical with [11:0] of the physical address. However, if the [11:0] LDVA is the same with a store instruction virtual address (STVA) of previously dispatched store instruction whose store data has not yet been determined, the physical addresses of the LD instruction and ST instruction are probably the same, that means a load-store collision is detected. In other embodiments, more bits of the virtual address will be calculated and compared to improve the accuracy.
The LDVA is shown provided through the synchronous latches 318 to stage B 310 to initiate execution. In stage B 310, the latched version of LDVA is provided to an input of a load-store queue (LoStQ) 326, which develops a clear signal CLR to invalidate the speculatively dispatched load instruction when a match occurs between the LDVA and a store instruction virtual address (STVA) of at least one previously dispatched store instruction whose store data has not yet been determined. The detail of the LoStQ 326 will be further described herein. In stage B 310, load valid (LV) logic 325 asserts a load valid signal LD VALID is provided back to an input of broadcast (BC) logic 324 in stage D 302. The BC logic 324 is operative to forward or broadcast the LD VALID signal to one or more up to all of the scheduler RS Q's within the reservation stations 108. In this manner, any instruction that has been issued to the reservation stations 108 and that is dependent upon the LD instruction may be scheduled for dispatch into corresponding execution units. Generally, these dependent instructions are not dispatched until the LD VALID signal is provided to ensure proper execution. It is worth noting that the AGU 322 provides the LDVA of the load instruction as early as one of the operand stages (e.g., Stages A(I) 304), that is, before the load instruction has progressed to the execution stages (e.g., stages B) to produce the LD VALID signal.
The CLR signal from the LoStQ 326 is provided back to an input of KILL logic 328 shown within the BC logic 324. In general, the kill logic 328 is operative to prevent the LD VALID signal from being broadcasted by the BC logic 324 to the scheduler RS Q's of the reservation stations 108. In one embodiment, the KILL logic 328 is separately added to prevent broadcast of the LD VALID signal. In another embodiment, the BC logic 324 includes disable logic or the like incorporated within the BC logic 324, in which the BC logic 324 passes the LD VALID signal to the reservation stations 108 only if the CLR signal is not asserted to disable broadcast.
In one alternative embodiment, the KILL logic 328 may be provided external to the BC logic 324, in which the KILL logic 328 either passes or blocks the broadcast of the LD VALID signal based on the state of CLR. In another alternative embodiment, the KILL logic 328 is distributed among the scheduler RS Q's of the reservation stations 108. In this case, the CLR signal may instead be provided to each of the scheduler RS Q's within the reservation stations 108, in which case assertion of the CLR signal prevents the broadcasted LD VALID signal from changing operative signals, bits or values within each RS Q.
In general, when CLR is asserted before LD VALID or at least coincidentally asserted with LD VALID, then the CLR signal suppresses detection of assertion of the LD VALID signal. When the LD VALID signal is suppressed, then any instructions that are dependent upon the speculatively issued load instruction are not yet dispatched for execution. The dependent instructions may be executed at a later time after the load instruction is replayed.
The LoStQ 326 receives virtual addresses of one or more previously dispatched store instructions in which the corresponding data values to be stored have not yet been determined or provided. As understood by those skilled in the art, a store instruction is divided into a STA μop for determining the address of the system memory 120 and a corresponding STD μop for determining the corresponding data value. The STA and STD μop are processed within the store pipeline 119 of the MOB 118. In one embodiment, the LoStQ 326 only receives the virtual addresses determined by one or more previously dispatched STA μops in which the corresponding STD μop is yet to be processed by the store pipeline 119. In another embodiment, the LoStQ 326 receives all the virtual addresses from the store pipeline 119. For example, the LD RS Q 126 also includes 16 entries, which holds 16 store instructions that are then dispatched to the store pipeline 119. In such a case, the entries of the LoStQ 326 respectively correspond to the entries of the LD RS Q 126 and receive all the store instruction virtual addresses (STVA0-STVA15) from the store pipeline 119. As described further herein, the LoStQ 326 identifies a collision between the LDVA of a speculatively dispatched load instruction and any of the virtual addresses of store instructions. When a collision is detected, the LoStQ 326 asserts the CLR signal to suppress detection of the LD VALID signal by the reservation stations 108 to prevent premature dispatch of any instructions dependent upon the speculatively dispatched load instruction.
The first entry ENT00 includes a comparator 402 that compares the virtual address LDVA of the speculatively dispatched LD instruction with a store instruction virtual address STVA0 of the STA μop received by the first entry ENT00. If the virtual addresses are the same, then the comparator 402 asserts a match signal M0 to one input of AND logic 404 (in which “&” denotes a logical AND function). The first entry ENT00 also includes AND logic 406 that receives a qualify valid signal QV0 from the store pipeline 119 that indicates whether the virtual address STVA0 is valid. The first entry ENT00 also includes OR logic 408 that receives qualify condition signals QC0A and QC0B from the store pipeline 119, in which either qualify condition signal asserted true indicates validity of STVA0. The qualify condition signals QC0A and QC0B are logically OR'd together and then logically AND'd with QV0 to determine a store valid signal VE0 indicating the validity of STVA0. The validity means STVA0 corresponds to a store instruction that is dispatched early than the speculatively dispatched load instruction and the corresponding store data of the store instruction has not yet been determined. VE0 is provided to the other input of the AND logic 404, which outputs a first clear signal CL0 for the first entry ENT00. Each of the entries ENT00-ENT15 output a corresponding one of 16 clear signals CL1-CL15, which are provided to respective inputs of OR logic 410, which outputs the CLR signal.
In operation of the LoStQ 326, if the virtual address LDVA matches any valid one of up to 16 store virtual addresses STVA0-STVA15 provided to the entries ENT00-ENT15, respectively, then a collision is detected and CLR is asserted in the B stage 310. If CLR is asserted, then when the speculatively dispatched load instruction reaches the B stage 310 and then the LD VALID signal is asserted, then the KILL logic 324 suppresses or blocks assertion of the LD VALID signal from being detected by the reservation stations 108. In this manner, any instruction located at any one of the RS queues of the reservation stations 108 that is also dependent upon the load instruction is not dispatched prematurely for execution. Thus, the execution pipelines in which the dependent instructions have been prematurely dispatched to need not be flushed and the corresponding instructions need not be replayed. Eventually, the load instruction is replayed non-speculatively in program order avoiding the potential for collision. If CLR is not asserted before LD VALID or coincidentallywith LD VALID, then a collision is not detected and the speculatively dispatched load instruction is allowed to execute to completion.
In one embodiment, the store virtual addresses STVA0-STVA15 and corresponding qualify valid signals QV0-QV15 and qualify condition signals QC0A/B-QC15A/B may be provided directly from the store pipe 119. Alternatively, when timing or routing issues may prevent a direct interface, intermediate memory devices, such as registers or latches or the like, may be provided between the store pipe 119 and the LoStQ 326. Alternatively, as shown by exemplary entry ENnn in
It is understood that the LoStQ 326 may detect and indicate a collision that does not, in fact, exist. In this case, any benefit of speculatively dispatching the load instruction is lost and a slight performance drop may result by delaying the dependent instructions. Nonetheless, the statistical occurrence of false detections is relatively small so that the performance benefits of detecting actual collisions significantly outweighs the slight performance crop caused by false detections.
The foregoing description has been presented to enable one of ordinary skill in the art to make and use the present invention as provided within the context of a particular application and its requirements. Although the present invention has been described in considerable detail with reference to certain preferred versions thereof, other versions and variations are possible and contemplated. Various modifications to the preferred embodiments will be apparent to one skilled in the art, and the general principles defined herein may be applied to other embodiments. For example, the circuits described herein may be implemented in any suitable manner including logic devices or circuitry or the like.
Those skilled in the art should appreciate that they can readily use the disclosed conception and specific embodiments as a basis for designing or modifying other structures for carrying out the same purposes of the present invention without departing from the spirit and scope of the invention. Therefore, the present invention is not intended to be limited to the particular embodiments shown and described herein, but is to be accorded the widest scope consistent with the principles and novel features herein disclosed.
| Number | Date | Country | Kind |
|---|---|---|---|
| 201510229378.3 | May 2015 | CN | national |
