Patent Application
20200183689
Embodiments of the present invention relate in general to an out-of-order (OoO) processor and more specifically to handling effective address synonyms with a synonym detection table (SDT) in a load-store unit that operates without address translation.
In an OoO processor, an instruction sequencing unit (ISU) dispatches instructions to various issue queues, renames registers in support of OoO execution, issues instructions from the various issue queues to the execution pipelines, completes executed instructions, and handles exception conditions. Register renaming is typically performed by mapper logic in the ISU before the instructions are placed in their respective issue queues. The ISU includes one or more issue queues that contain dependency matrices for tracking dependencies between instructions. A dependency matrix typically includes one row and one column for each instruction in the issue queue.
In the domain of central processing unit (CPU) design, and particularly for OoO processors, hazards pose technical challenges with the instruction pipeline in the CPU microarchitectures when a next instruction cannot execute in the following clock cycle, because of potentially leading to incorrect computation results. Typical types of hazards include data hazards, structural hazards, and control flow hazards (branching hazards). Data hazards occur when instructions that exhibit data dependence modify data in different stages of a pipeline, for example read after write (RAW), write after read (WAR), and write after write (WAW). A structural hazard occurs when a part of the processor's hardware is needed by two or more instructions at the same time, for example a memory unit being accessed both in the fetch stage where an instruction is retrieved from memory, and the memory stage where data is written and/or read from memory. Further, branching hazards (also termed control hazards) occur with branches in the computer program being executed by the processor.
Embodiments of the present invention include methods, systems, and computer program products for implementing effective address based load store unit in out of order processors. A non-limiting example of a processing unit for executing one or more instructions includes a load-store unit (LSU) for transferring data between memory and registers. The LSU executes a plurality of instructions in an out-of-order (OoO) window. The execution includes, in response to determining a first effective address being used by a first instruction, the first effective address corresponding to a first real address, creating an effective real table (ERT) entry in an ERT, the ERT entry mapping the first effective address to the first real address. Further, the execution includes in response to determining an effective address synonym used by a second instruction, the effective address synonym being a second effective address that is also corresponding to said first real address: creating a synonym detection table (SDT) entry in an SDT, wherein the SDT entry maps the second effective address to the ERT entry, and relaunching the second instruction by replacing the second effective address in the second instruction with the first effective address.
According to one or more embodiments of the present invention, an example computer-implemented method for executing one or more out-of-order instructions by a processing unit includes issuing or executing, by a load-store unit (LSU), a plurality of instructions from an out-of-order (OoO) window. The issuing includes, in response to determining a first effective address being used by a first instruction, the first effective address corresponding to a first real address, creating an effective real table (ERT) entry in an ERT, the ERT entry mapping the first effective address to the first real address. Further, the execution includes in response to determining an effective address synonym used by a second instruction, the effective address synonym being a second effective address that is also corresponding to said first real address: creating a synonym detection table (SDT) entry in an SDT, wherein the SDT entry maps the second effective address to the ERT entry, and relaunching the second instruction by replacing the second effective address in the second instruction with the first effective address.
According to one or more embodiments a computer program product includes a computer readable storage medium having program instructions embodied therewith, where the program instructions executable by a processing unit to cause the processing unit to perform operations. The operations include issuing or executing, by a load-store unit (LSU), a plurality of instructions from an out-of-order (OoO) window. The issuing includes, in response to determining a first effective address being used by a first instruction, the first effective address corresponding to a first real address, creating an effective real table (ERT) entry in an ERT, the ERT entry mapping the first effective address to the first real address. Further, the execution includes in response to determining an effective address synonym used by a second instruction, the effective address synonym being a second effective address that is also corresponding to said first real address: creating a synonym detection table (SDT) entry in an SDT, wherein the SDT entry maps the second effective address to the ERT entry, and relaunching the second instruction by replacing the second effective address in the second instruction with the first effective address.
Additional features and advantages are realized through the techniques of the present invention. Other embodiments and aspects of the invention are described in detail herein and are considered a part of the claimed invention. For a better understanding of the invention with the advantages and the features, refer to the description and to the drawings.
The specifics of the exclusive rights described herein are particularly pointed out and distinctly claimed in the claims at the conclusion of the specification. The foregoing and other features and advantages of the embodiments of the invention are apparent from the following detailed description taken in conjunction with the accompanying drawings in which:
The diagrams depicted herein are illustrative. There can be many variations to the diagram or the operations described therein without departing from the spirit of the invention. For instance, the actions can be performed in a differing order or actions can be added, deleted or modified. Also, the term “coupled” and variations thereof describes having a communications path between two elements and does not imply a direct connection between the elements with no intervening elements/connections between them. All of these variations are considered a part of the specification.
One or more embodiments of the present invention described herein provide an effective address (EA) based load store unit (LSU) for an out-of-order (OoO) processor by dynamic removal of effective real address table entries in the OoO processor. The technical solutions described herein use an effective address directory (EAD) in conjunction with an effective real table (ERT) and a synonym detection table (SDT), among other components, to facilitate reduction in chip area and further to improve timing of OoO processors.
As mentioned above, due to the OoO nature of modern processors, younger load instructions may be chosen for execution ahead of older store instructions to a same real address (RA). When a younger load instruction executes ahead of an older store instruction to the same RA, the younger load instruction returns stale data and instructions in a processor pipeline must be flushed from the pipeline and re-fetched to restore program consistency. Typically, it is desirable to delay a younger load instruction until an older store instruction to a same RA executes so that other independent instructions can execute and not be unnecessarily flushed from a processor pipeline. A step in reducing processor pipeline flushes due to execution of a younger load instruction before an older store instruction to a same RA is to identify a load/store sequence that is to the same RA and will be executed OoO such that a processor pipeline flush is required. Alternatively, or in addition, the older load instruction from a load-reorder queue and all subsequent instruction therefrom are flushed from the load-reorder queue or from an instruction fetching unit (IFU). Flushing an instruction includes sending a flush message (that includes an appropriate identifier of the instruction) to the IFU.
Most modern computing devices provide support for virtual memory. Virtual memory is a technique by which application programs are given the impression that they have a contiguous working memory, or address space, when in fact the physical memory may be fragmented and may even overflow onto disk storage. Essentially, the application program is given a view of the memory of the computing device where the application accesses a seemingly contiguous memory using an EA, in the EA space visible to the application, which is then translated into a physical address of the actual physical memory or storage device(s) to actually perform the access operation. An EA is the value which is used to specify a memory location that is to be accessed by the operation from the perspective of the entity, e.g., application, process, thread, interrupt handler, kernel component, etc., issuing the operation.
That is, if a computing device does not support the concept of virtual memory, then the EA and the physical address are one and the same. However, if the computing device does support virtual memory, then the EA of the particular operation submitted by the application is translated by the computing device's memory mapping unit into a physical address which specifies the location in the physical memory or storage device(s) where the operation is to be performed.
Further, in modern computing devices, processors of the computing devices use processor instruction pipelines, comprising a series of data processing elements, to process instructions (operations) submitted by entities, e.g., applications, processes, etc. Instruction pipelining is a technique to increase instruction throughput by splitting the processing of computer instructions into a series of steps with storage at the end of each step. Instruction pipelining facilitates the computing device's control circuitry to issue instructions to the processor instruction pipeline at the processing rate of the slowest step which is much faster than the time needed to perform all steps at once. Processors with instruction pipelining, i.e. pipelined processors, are internally organized into stages which can semi-independently work on separate jobs. Each stage is organized and linked with a next stage in a series chain so that each stage's output is fed to another stage until the final stage of the pipeline.
Such pipelined processors may take the form of in-order or out-of-order pipelined processors. For in-order pipelined processors, instructions are executed in order such that if data is not available for the instruction to be processed at a particular stage of the pipeline, execution of instructions through the pipeline may be stalled until the data is available. OoO pipelined processors, on the other hand, allow the processor to avoid stalls that occur when the data needed to perform an operation are unavailable. The OoO processor instruction pipeline avoids these stalls by filling in “slots” in time with other instructions that are ready to be processed and then re-ordering the results at the end of the pipeline to make it appear that the instructions were processed in-order. The way the instructions are ordered in the original computer code is known as program order, whereas in the processor they are handled in data order, i.e. the order in which the data and operands become available in the processor's registers.
Modern processor instruction pipelines track an instruction's EA as the instruction flows through the instruction pipeline. It is important to track the instruction's EA because this EA is utilized whenever the processing of an instruction results in the taking of an exception, the instruction flushes to a prior state, the instruction branches to a new memory location relative to its current memory location, or the instruction completes its execution.
Tracking an instruction's EA is costly in terms of processor chip area, power consumption, and the like. This is because these EAs have large sizes (e.g., 64 bits) and modern processor instruction pipelines are deep, i.e. have many stages, causing the lifetime of an instruction from an instruction fetch stage of the processor instruction pipeline to a completion stage of the processor instruction pipeline to be very long. This cost may be further increased in highly multithreaded OoO processors, i.e. processors that execute instructions from multiple threads in an OoO manner, since a vast number of instructions from different address ranges can be processing, i.e. are “in flight,” at the same time.
In one or more examples, computing devices use a combination of pipeline latches, a branch information queue (BIQ), and a global completion table (GCT) to track an instruction's EA. The base EA for a group of instructions is transferred from the front-end of the pipeline using latches until it can be deposited and tracked in the GCT of the instruction sequencer unit (ISU). The number of latches needed to store this data is on the order of the number of pipeline stages between a Fetch stage and a Dispatch stage of the pipeline. This is wasteful, as the EA is typically not needed during these stages. Rather it is simply payload data that is “along for the ride” with the instruction group as it flows through the pipeline. In addition, this method leads to duplicate storage as branch instructions have their EAs in both the BIQ and the GCT.
Accordingly, computing devices have been developed, that remove these inefficiencies by tracking the EA solely in the GCT. For example, these new computing devices, an instruction sequencer unit creates an entry in the GCT at fetch time. The EA is loaded into the GCT at this time and then removed when the instruction completes. This eliminates many pipeline latches throughout the machine. Instead of a full EA that is as long as number of address lines, for example a 64-bit EA, a small tag is carried along with the instruction group through the pipeline. This tag points back to the entry in the GCT, which holds the base EA for this instruction group. Address storage in the BIQ is no longer needed as branches can retrieve their EA directly from the GCT when they issue. Such techniques improve area efficiency, but they are not applicable in an OoO processor. Further, they lack sufficient information to process address requests arriving out of program order. In addition, these techniques cannot support dispatch and completion bandwidth required for OoO execution because they lack the ability to track instruction groups that may have been formed from multiple disjoint address ranges. Historically, such mechanisms have only supported instruction groups from a single address range, which can significantly reduce the number of instructions available to execute OoO. Further, to lookup corresponding addresses, such as an RA corresponding to an EA (or vice versa) a Content Addressable Memory (CAM) is used. A CAM implements lookuptable function in a single clock cycle using dedicated comparison circuitry. The overall function of a CAM is to take a search word and return the matching memory location. However, such CAM takes chip area as well as consumes power for such lookups.
Further, such conversions of EA to corresponding RA are typically done at a second level of a memory nest associated with the processor. As used herein the term memory nest refers to the various types of storage that can be used by a processor to store data. In general, the memory nest includes a hierarchy of caches and physical memory. In general, as the level of the memory nest increases, the distance from the processor to the data increases and access latency for the processor to retrieve the data also increases. Thus, converting EA to RA slows the processor execution.
The illustrative embodiments of technical solutions described herein improve upon these techniques by providing an effective address directory (EAD), an effective real table (ERT), and a synonym detection table (SDT) that have the area efficiency of the GCT solution described above, but can also support a wide issue OoO pipeline while not inhibiting performance. The technical solutions described herein further facilitate the processors to run with only EAs, as long as the processor can avoid EA synonyms within an out of order (OoO) window. The OoO window is a set of instructions in an instruction pipeline of the processor. By avoiding EA synonyms in the OoO window, the processor reduces the chip area and power consumption for address translation, because the processor can avoid translation for the EA in the OoO window.
In other words, the technical solutions described herein address the technical problem by policing against EA aliasing within the OoO window, and thus reducing translation data structures and hardware for the load/store ports. Accordingly, the technical solutions described herein facilitate a reduction in chip area by tracking only one address, the EA. Further, the technical solutions facilitate the OoO processor to run in a 2 load and 2 store mode with partitioned load store queues, further reducing CAM ports that are typically used for the address translation.
For example, a processor that is capable of issuing and executing instructions OoO may permit load instructions to be executed ahead of store instructions. Assuming that a real address (RA) of a younger load instruction does not overlap with an RA of an older store instruction, OoO execution of the younger load instruction and the older store instruction may provide performance advantages. In a typical program, the likelihood that an RA of a younger load instruction overlaps with an RA of an older store instruction (that executes after the younger load instruction) is relatively low. As is known, a store violation condition (store-hit-load (SHL)) is indicated when an RA of a store instruction hits in a load reorder queue (LRQ) of a load store unit (LSU). That is, when an RA of a newly issued older store instruction matches an RA of a younger load instruction present in the LRQ of the LSU, an SHL is indicated. However, as detection of an SHL typically occurs late in an instruction execution pipeline, recovering from an SHL typically results in a relatively severe penalty on processor performance. For example, recovery from an SHL typically involves invalidating the younger load instruction that caused the SHL and reissuing the younger load instruction, as well as all instructions issued after the older store instruction.
Further, for example, if two load instructions to the same address are executed OoO, and the value of the data at that address is changed between the executions of the two load instructions (e.g., by another processor), the later (i.e., younger) load will obtain an earlier (i.e., old) value, and the earlier (i.e., older) load will obtain a later (i.e., new) value. This situation is termed a “load-load order violation” or a “load-hit-load hazard.” The requirement that if a younger load instruction obtains old data, an older load instruction to the same address must not obtain new data is termed “sequential load consistency.” In addition, if a later (i.e., younger) load instruction is executed before an earlier (i.e., older) store instruction to the same address (i.e., memory location) is completed, the load instruction will obtain an earlier (i.e., old) value. This situation is termed a “load-store order violation” or a “load-hit-store hazard.” (See, for example, “Power4 System Microarchitecture” by J. M. Tendler et al., IBM Journal of Research and Development, Volume 46, Number 1, January 2002, pp. 5-25.).
However, in case of the OoO processor that operates using EAs (and not RAs), technical challenges are posed for detecting the SHL and LHL conditions only based on EAs, and further recovering from these situations. Typically, processors have dedicated hardware to avoid load-load and load-store order violations, thereby helping to ensure program correctness. However, such hardware is typically complex and adds time delays. Further, such hardware relies on using RA to detect the hazard conditions, thus occupying larger chip area (for translation/CAM ports) and consuming more power for the EA-to-RA and/or RA-to-EA translation(s). The technical solutions described herein address such technical challenges by using the EA and using effective real translation table (ERT) indices stored with EAD entries. The technical solutions herein use a load-hit-load table to detect load-hit-load hazards and act upon such situations. The technical solutions thus facilitate the OoO processor have lesser chip area dedicated to storing and manipulating real addresses.
Turning now to
The mappers 110 shown in
The output from the setup block 108 is also input to a global completion table (GCT) 112 for tracking all of the instructions currently in the ISU. The output from the setup block 108 is also input to a dispatch unit 114 for dispatching the instructions to an issue queue. The embodiment of the ISU shown in
Instructions output from the dispatch logic and renamed registers from the LNK/CNT mapper, XER mapper, UMapper (GPR/VSR), ARCH Mapper (GPR/VSR), and FPSCR mapper are input to issue queue 102. As shown in
When the processor is executing in multi-threaded (MT) mode, ISQ01020 can be used to process N/2 instructions from a first thread and ISQ11021 is used to process N/2 instructions from a second thread ISQ11021.
As shown in
As shown in
In accordance with one or more embodiments of the present invention, when the processor is executing in MT mode, the first group of execution units 1040 execute instructions of the first thread issued from ISQ01020 and the second group of execution units 1041 execute instructions of the second thread issued from ISQ11021.
The number of entries in the issue queue 102 and sizes of other elements (e.g., bus widths, queue sizes) shown in
In one or more examples, the system 100, in accordance with the illustrative embodiments, is an OoO processor.
The instruction cache 202 receives instructions from the L2 cache 260 via the second level translation unit 262 and pre-decode unit 270. The second level translation unit 262 uses its associate segment look-aside buffer 264 and translation look-aside buffer 266 to translate addresses of the fetched instruction from effective addresses to system memory addresses. The pre-decode unit partially decodes instructions arriving from the L2 cache and augments them with unique identifying information that simplifies the work of the downstream instruction decoders.
The instructions fetched into the instruction fetch buffer 204 are also provided to the branch prediction unit 280 if the instruction is a branch instruction. The branch prediction unit 280 includes a branch history table 282, return stack 284, and count cache 286. These elements predict the next EA that should be fetched from the instruction cache. A branch instruction is a point in a computer program where flow of control is altered. It is the low-level machine instruction that is generated from control constructs in a computer program, such as if-then-else or do-while statements. A branch can be not taken, in which the flow of control is unchanged and the next instruction to be executed is the instruction immediately following it in memory, or it can be taken, in which the next instruction to be executed is an instruction at some other place in memory. If the branch is taken, a new EA needs to be presented to the instruction cache.
The EA and associated prediction information from the branch prediction unit are written into an effective address directory 290. This EA is later confirmed by the branch execution unit 222. If correct, the EA remains in the directory until all instructions from this address region have completed their execution. If incorrect, the branch execution unit flushes out the address and the corrected address is written in its place. The EAD 290 also includes a logic unit that facilitates using the directory as a CAM.
Instructions that read from or write to memory (such as load or store instructions) are issued to the LS/EX execution unit 238, 240. The LS/EX execution unit retrieves data from the data cache 250 using a memory address specified by the instruction. This address is an EA and needs to first be translated to a system memory address via the second level translation unit before being used. If an address is not found in the data cache, the load miss queue is used to manage the miss request to the L2 cache. In order to reduce the penalty for such cache misses, the advanced data prefetch engine predicts the addresses that are likely to be used by instructions in the near future. In this manner, data will likely already be in the data cache when an instruction needs it, thereby preventing a long latency miss request to the L2 cache.
The LS/EX execution unit 238, 240 executes instructions out of program order by tracking instruction ages and memory dependences in the load reorder queue 218 and store reorder queue 220. These queues are used to detect when OoO execution generated a result that is not consistent with an in-order execution of the same program. In such cases, the current program flow is flushed and performed again.
The processor architecture further includes the EAD 290 which maintains the effective address of a group of instructions in a centralized manner such that the EA is available when needed but is not required to be passed through the pipeline. Moreover, the EAD 290 includes circuitry and/or logic for supporting OoO processing.
Those of ordinary skill in the art will appreciate that the hardware in
Moreover, the data processing system 100 may take the form of any of a number of different data processing systems including client computing devices, server computing devices, a tablet computer, laptop computer, telephone or other communication device, a personal digital assistant (PDA), or the like. In some illustrative examples, data processing system 100 may be a portable computing device configured with flash memory to provide non-volatile memory for storing operating system files and/or user-generated data, for example. Essentially, data processing system 100 may be any known or later developed data processing system without architectural limitation.
As will be appreciated by one skilled in the art, the present invention may be embodied as a system, apparatus, or method. In one illustrative embodiment, the mechanisms are provided entirely in hardware, e.g., circuitry, hardware modules or units, etc. of a processor. However, in other illustrative embodiments, a combination of software and hardware may be utilized to provide or implement the features and mechanisms of the illustrative embodiments. The software may be provided, for example, in firmware, resident software, micro-code, or the like. The various flowcharts set forth hereafter provide an outline of operations that may be performed by this hardware and/or combination of hardware and software.
In illustrative embodiments in which the mechanisms of the illustrative embodiments are at least partially implemented in software, any combination of one or more computer usable or computer readable medium(s) that store this software may be utilized. The computer-usable or computer-readable medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), etc.
Typically, for every load and every store instruction, an EA is converted to corresponding RA. Such an EA to RA conversion is also performed for an instruction fetch (I-fetch). Such conversion typically required an effective to real address table (ERAT) for retrieval of instructions from lower order memory. In the technical solutions described herein, the EA to RA conversion is not performed for every load and store instruction, rather only in case of load-misses, I-Fetch misses, and all stores.
By using only EA for the operations, the technical solutions facilitate removal of RA bits (for example, bits 8:51) from one or more data structures, such as an EA directory (also referred to as L1 directory), LRQF entries, LMQ entries. Further, SRQ LHS RA compare logic is not executed if only the EA is being used. Removing such elements reduces chip area of the processor used, thus facilitating a reduction in chip area over typical processors.
Further, by only using EA, the technical solutions herein eliminate ERAT camming on every load and store address generation. The technical solutions further eliminate RA bus switching throughout the unit, and also avoids fast SRQ LHS RA cam. The technical solutions thus facilitate the processor to consume lesser power compared to typical processors by not performing the above operations.
Further yet, the technical solutions herein also facilitate improvements to L1 latency. For example, the technical solutions herein, by eliminating the address conversions are at least 1 cycle faster in determining “final dval” compared to typical processors that perform the EA to RA conversions. The latency is also improved because by using only the EA (without RA conversion) eliminates “bad dval” condition(s), such as setup table multi-hit, setup table hit/RA miss, and the like. In a similar manner, the technical solutions herein facilitate improvement to L2 latency.
The technical challenges of using only an EA-based LSU include that out of order execution of the instructions may lead to hazards (LHL, SHL, LHS), and such hazards are to be detected without using corresponding real addresses for an EA only implementation. The technical solutions described herein address such technical problem. The RA is not used for load-hit-store, store-hit-load or load-hit-load type of OoO execution hazard detection as it used typically. The RA calculation for stores happens before the store is completed, because after completion any interrupt for the store instructions are not processed (store can generate an address translation related interrupt, which needs to be processed before the store is completed). Here the RA calculation is done when the store is issued (and not when drained). The technical solutions described herein determine LHS (load-hit-store), SHL (store-hit-load) and LHL (load-hit-load) based on the EA and the ERT-Index stored with the EA directory entries, and also using an load reorder queue (LRQF) with structure and operation as described herein.
Referring again to the figures,
The LSU 104 includes a load-reorder-queue (LRQF) 218, where all load operations are tracked from dispatch to complete. The LSU 104 further includes a second load-reorder-queue LRQE 225. When a load is rejected (for cache miss, or translation miss, or previous instruction it depends on got rejected) the load is taken out of the issue queue and placed in a LRQE entry for it to be re-issued from there. The depicted LRQE 225 is partitioned into 2 instances, LRQE0, and LRQE1 for the two load mode, with 12 entries each (24 entries total). In ST mode, no threads/pipe based partition exists. In the MT mode, T0, T2 operations launched on pipe LD0; and T1, T3 operations launched on pipe LD1, for relaunch.
As depicted, the LRQF 218 is partitioned into 2 instances LRQF0 and LRQF1 for the two load mode, with 40 entries (each instance). The LRQF 218 is circular in order entry allocation, circular in order entry drain, and circular in order entry deallocation. Further, in MT mode, T0, T2 operations launched on pipes LD0, ST0; and T1, T3 operations launched on pipes LD1, ST1. In ST mode, the LRQF does not have any pipes/threads.
In one or more examples, the LRQF 218 (and other structures described herein) is partitioned for the SMT4 mode as T0:LRQF0[0:19] circular queue, T1: LRQF1[0:19] circular queue; and T2:LRQF0[20:39] circular queue, T3:LRQF1[20:39] circular queue.
In one or more examples, the LRQF 218 (and other structures described herein) is partitioned for the SMT2 mode as T0:LRQF0[0:39] circular queue, and T1:LRQF1[0:39] circular queue. Further, in one or more examples, for the ST mode, LRQF0[0:39] circular queue, with the LRQF1 being an LRQF0 copy. For other data structure, a similar partition pattern is used with the second instance being a copy of the first instance in ST mode.
In case of a cross invalidation flush (XI flush), for the LRQF, NTC+1 flush any thread that an XI or store drain from another thread hits so that explicit L/L ordering flushes on sync's is not performed by the LSU 104 in case of the XI flush.
All stores check against the LRQF 218 for SHL detection, upon which the LRQF 218 initiates a flush of any load, or everything (any instruction/operation) after the store. Further, DCB instructions check against the LRQF 218 for SHL cases, upon which the LRQF 218 causes a flush of the load, or everything after the DCB. Further, all loads check against the LRQF 218 for LHL detection (sequential load consistency), upon which the :RQF 218 causes a flush of younger load, or everything after the older load. In one or more examples, the LRQF 218 provides quad-word atomicity, and the LQ checks against the LRQF 218 for quad atomicity and flushes LQ if not atomic. Further yet, in case of LARX instructions, the LSU 104 checks against the LRQF 218 for larx-hit-larx cases, and in response flushes younger LARX, or everything after the older larx instruction.
Thus, the LRQF 218 facilitates tracking all load operations from issue to completion. Entries in the LRQF 218 are indexed with Real_Ltag (rltag), which is the physical location in the queue structure. The age of a load operation/entry in the LRQF 218 is determined with a Virtual_Ltag (vltag), which is in-order. The LRQF flushes a load using GMASK and partial group flush using GTAG and IMASK. The LRQF logic can flush from current iTag or iTag+1 or precise load iTag.
Further yet, the LRQF does not include an RA (8:51) field typically used, and instead is EA-based and includes an ERT ID (0:6), and EA(40:51) (saving of 24 bits). The LRQF page match on SHL, LHL is based on ERT ID match. Further, Each LRQ entry has a “Force Page Match” bit. When an ERT ID is invalidated that matches the LRQ Entry ERT ID the Force Page Match bit is set. The LRQ will detect LHL, SHL, and store ordering flushes involving any entry with Force Match Match=1.
The SRQ 220 of the LSU 104 has similar structure as the LRQF 218, with two instances SRQR0 and SRQR1 of 40 entries (each instance), which are circular in order entry allocation, circular in order entry drain, and circular in order entry deallocation. Further, the SRQ 220 is partitioned similar to the LRQ 218, for example T0, T2 ops launched on pipes LD0, ST0; T1, T3 ops launched on pipes LD1, ST1; and no pipe/thread partition in ST mode. In the ST mode, both copies have identical values, with the copies being different in the MT modes. In SMT4 mode, both instances are further partitioned, with each thread allocated 20 entries from the SRQ 220 (see example partition for LRQF described herein). In one or more examples, for store drain arbitration, an intra-SRQ read pointer multiplexing is performed in the SMT4 mode. Alternatively, or in addition, an inter SRQ0/1 multiplexing is performed in SMT2, and SMT4 modes. In the ST mode drain is performed only on SRQ0.
Each entry of the SRQ 220 contains a store TID(0:1), an ERT ID(0:6), EA(44:63), and RA(8:51). To detect LHS, the LSU uses the {Store Tid, EA(44:63)}, thus eliminating RA LHS alias check. The ERT ID is used to “catch” EA(44:63) partial match mis-speculation. The SRQ entry has the RA(8:51), which is translated at store agen, and is only used when sending store requests to the L2 (store instruction drained, not issued). Each SRQ entry also has a “Force Page Match” bit. The force page match bit is set when an ERT ID is invalidated that matches the SRQ entry ERT ID. The SRQ can detect LHS involving any entry with Force Page Match=1. For example, LHS against an entry with Force Page Match=1 causes a reject of the load instruction. Further, a store drain forces a miss in the L1 cache if Force Page Match=1 for the SRQ entry. This works in tandem with “Extended store hit reload” LMQ actions.
For example, for an LMQ, an LMQ Address Match={ERT ID, EA Page Offset(xx:51), EA(52:56)} match. Further, a “Force Page Match” bit of each LMQ entry is set (=1) when an ERT ID is invalidated that matches the LMQ Entry ERT ID. The LMQ rejects a load miss if a valid LMQ entry[x]ForcePageMatch=1 and Ld Miss EA[52:56]=LMQEntry[X]EA(52:56). Further, the LMQ has an extended store hit reload. For example, LMQ suppresses reload enable if Reload EA(52:56)=SRQEntry[X] EA(52:56) and SRQEntry[X]ForcePageMatch=1. Alternatively, or in addition, LMQ suppresses reload enable if LMQEntry[X]EA(52:56)=StDrain EA(52:56) and StDrainForcePageMatch=1.
The LSU 104 depicted collapses a Store Data Queue (SDQ) as part of the SRQ 220 itself to further save chip area. The operands are stored in an entry of the SRQ itself if the operand size is less than the SRQ entry size, for example 8 bytes. In case of wider operands, such as vector operands, for example are 16 bytes wide, the SRQ stores the operands using two consecutive entries in the SRQ 220 in MT mode. In ST mode, the wider operands are stored in the SRQ0 and SRQ1, for example 8 bytes each.
The SRQ 220 queues operations of type stores, barriers, DCB, ICBI or TLB type of operations. A single s-tag is used for both store_agen and store_data. The SRQ 220 handles load-hit-store (LHS) cases (same thread only). For example, all loads issued are checked by the SRQ 220 to ensure there is no older stores with a data conflict. For example, the data conflict is detected by comparing loads EA and data byte flags against older stores in the SRQ EA array.
SRQ entries are allocated at dispatch where the dispatched instruction tags (itags) are filled into the correct row. Further, SRQ entries are deallocated on store drain. In one or more examples, the itag arrays hold “overflow” dispatches. For example, information is written into the itag array at dispatch if the row in the SRQ that is desired, say SRQ entry x is still in use. When, the SRQ entry x is deallocated, its corresponding row in the SRQ overflow itag structure is read out and copied into the main SRQ itag array structure (read of the overflow itag structure gated by whether there are any valid entries in the overflow itag array for a given thread/region). The main SRQ 0/1 itag array is cammed (or ½ cammed in SMT4) to determine which physical row to write into upon store issue, so that the ISU issues stores based on the itag. The SRQ 220 sends to the ISU, the itag when a store drains & deallocates.
ISU assigns virtual sub-regions to store dispatches to avoid overlapping issues. For example, in ST mode, the ISU does not issue a virtual SRQ entry 40 until real SRQ entry 0 is deallocated by an entry 0 store drain. Further, in SMT4 mode, the ISU cannot issue Tx virtual SRQ entry 20 until real Tx SRQ entry 0 is drained and deallocated. The ISU subdivide each thread partition further into 4 regions.
For example, for ST mode subdivide the SRQ 220 into 4 subregions: Ping A:SRQ entries 0-9, Ping B:SRQ entries 10-19, Ping C:SRQ entries 20-29, Ping D:SRQ entries 30-39; and Pong A:SRQ entries 0-9, Pong B:SRQ entries 10-19, Pong C:SRQ entries 20-29, Pong D:SRQ entries 30-39. Initially, ISU issues Ping A,B,C,D itags. Further, the ISU does not issue Pong A itags until Ping A itags have been deallocated. Subsequently, after all Ping A itags are deallocated, ISU issues Pong A itags, but not issue Pong B itags until Ping B itags have been deallocated, similar to the case A. In one or more examples, the ISU holds 3 extra bits in the issue queue entry (1 wrap bit+2 bits to delineate which subregion) to create a pseudo issue dependency based on subregion.
As shown in
The base EA 310 is the starting EA of the group of instructions. Each instruction in the group of instructions has the same base EA and then an offset from it. For example, in one illustrative embodiment, the EA is a 64 bit address comprising bits 0:63. The base EA may comprise, in one illustrative embodiment, bits 0:56 of this EA with bits 57:61 representing the offset from the base EA for the specific instruction within the group of instructions. Bits 62 and 63 point to a specific byte of each instruction. In the illustrative embodiment, each address references an instruction that is 32 bits long (i.e. 4 bytes), where each byte in memory is addressable. An instruction cannot be further divided into addressable subcomponents, and thus an instruction address will always have bits 62 and 63 set to zero. Therefore, bits 62 and 63 do not need to be stored and can always be assumed to be zero by the EAD.
The first instruction identifier field 320 stores the EA offset bits, e.g., bits 57:61 of the EA for the first instruction in the group of instructions to which the EAD 290 entry corresponds. A combination of the base EA from field 310 and the EA offset bits in the first instruction identifier field 320 provides the EA for the first instruction in the group of instructions represented by the EAD 290 entry. This first field 320 may be used, as discussed hereafter, for recovering a refetch address and branch prediction information in the event that the pipeline is flushed, for example.
The last instruction identifier field 330 stores the EA offset bits, e.g., bits 57:61 of the EA, for the last instruction in the group of instructions to which the EAD 290 entry corresponds. EAD logic updates this field as additional instructions in the group of instructions represented by the EAD 290 entry are fetched. The EAD logic discontinues updating of this field 330 in the particular EAD 290 entry in response to the EAD 290 entry being closed when a cache line crossing or taken branch is found. This field will remain intact unless a pipeline flush occurs that clears out a portion of the EAD entry. In such cases, the EAD logic updates this field to store the EA offset bits of the instruction that is now the new last instruction in the entry as a result of the flush. This field is ultimately used for completion, as discussed hereafter, to release the entry in the EAD 290.
The closed identifier field 340 is used to indicate that the EAD 290 entry has been closed and no more instruction fetches will be made to fetch instructions for the instruction group corresponding to the EAD 290 entry. An EAD 290 entry may be closed for a variety of different reasons, including a cache line crossing, a branch being taken, or a flush of the pipeline. Any of these conditions may result in the value in the closed field 340 being set to indicate the EAD entry is closed, e.g., set to a value of “1.” This field 340 is used at completion to release an entry in the EAD 290, as discussed in greater detail hereafter.
The global history vector field 350 identifies the global history vector for the first instruction fetch group that created the entry in the EAD 290. The global history vector is used to identify a history of whether branches were taken or not taken, as discussed in greater detail hereafter. The global history vector is used for branch prediction purposes to help in determining, based on the recent history of branches being taken or not taken, whether a current branch is likely to be taken or not.
The link stack pointer field 360 identifies the link stack pointer for the first instruction fetch group that created the entry in the EAD 290. The link stack pointer is another branch prediction mechanism that will be described in greater detail hereafter.
The branch taken field 370 indicates whether the group of instructions corresponding to the EAD 290 entry had a branch instruction in which the branch was taken. The value in the branch taken field 370 is updated in response to a branch instruction of the instruction group represented by the EAD 290 entry being predicted as taken. In addition, once a branch in the instructions of the EAD 290 entry is taken, the EAD 290 entry is also closed by writing the appropriate value to the closed field 340. Since the branch taken field is written speculatively at prediction time, it may need to be replaced with the correct value when the branch is actually executed. For example, a branch could be predicted as not taken, in which case a “0” would be written into the branch taken field. However, later in execution, the branch could be found to be taken, in which case the field must be corrected by writing it to a value of “1”. The second write only occurs if the branch was mispredicted.
The branch information field 380 stores miscellaneous branch information that is used for updating branch prediction structures when a branch resolves, or architected EA state when a branch instruction completes.
The ERT_ID field 385 stores an index into the ERT table (described further), which identifies a corresponding ERT entry. When an ERT entry is invalidated, the associated ERT_ID is invalidated and it will also invalidate all associated entries in L1 cache and L1 D cache.
Entries in the EAD 290 are accessed using an effective address tag (eatag) that comprises at least two parts: base eatag and an eatag offset. In one illustrative embodiment, this eatag is a 10 bit value, which is relatively much smaller than the 64 bit EA. With a 10 bit eatag value, and a EAD 290 having a size of 14 entries, in one exemplary implementation, the eatag is comprised of a first 5 bits, referred to as the base eatag, for identifying an entry within the EAD 290 and a second 5 bits, referred to as the eatag offset, for providing the offset of the specific instruction within the group of instructions represented by the entry in the EAD 290. A first bit in the 5 bits identifying the entry within the EAD 290 may be used as a wrap bit to indicate whether a wrap occurred when going from the topmost entry to the bottom most entry of the EAD 290. This may be used for age detection. The second through fifth bits of the 5 bits identifying the entry within the EAD 290 may be used to index into the EAD to identify the base EA of the instruction, i.e. EA(0:56). The 5 bit offset value may be used to provide, for example, bits 57:61 of the particular instruction's EA. This example eatag is illustrated below:
row(1:4): Index into 14-entry EAD used to determine EA(0:56) of the instruction.
The ERT 255 includes a valid ERT entry, in general, exists for any page active in the L1 I-Cache or D-Cache directory (EAD 290) or an SRQ entry or an LRQF entry or an LMQ entry. In other words, ERT 255 is a table of all active RPN's in the LSU and IFU (L1 DC, SRQ, LRQE, LRQF, LMQ, IC). In one or more examples, if the processor 106 is operating in ST mode, all entries in the ERT 255 are used for the single thread that is being executed. Alternatively, in one or more examples, the entries in the ERT 255 are divided into sets, and in ST mode, each set has the same content. For example, if the ERT 255 has 128 total entries, and supports maximum two threads, when the processor operates in ST mode, the ERT 255 includes two sets of 64 entries each, and the two sets have the same content.
Alternatively, if the processor 106 is operating in the MT mode, the ERT entries are divided among the threads being executed. For example, in case of two threads, the ERT entries are divided into two equal sets, a first set of entries associated with a first thread, and a second set of entries associated with a second thread. For example, 1 copy for LD0 pipe L1 misses, ST0 pipe launches, T0/T2 I-Fetches: ERT0, which handles T0 in SMT2 mode and T0/T2 in SMT4 mode; and 1 copy for LD1 pipe L1 misses, ST1 pipe launches, T1/T3 I-Fetches: ERT1, which handles T1 in SMT2 mode and T1/T3 in SMT4 mode.
In one or more examples, each ERT entry includes at least the following fields ERT fields, ERT_ID (0:6), Tid_en (0:1), Page Size (0:1), EA (0:51), and RA (8:51). The ERT_ID field is a unique index for each ERT entry. For example, the ERT_ID may include a sequential number that identifies the ERT entry. The ERT_ID is stored in the ERT_ID field 285 of the EAD 290, and other data structures used by the LSU. The TID_en field indicates if the entry is enabled for being used in MT mode, and in one or more examples the thread identifier of the instruction that is using the ERT entry. Further, the Page Size indicates the memory page size to which the ERT entry refers. The RA includes a real address associated with the ERT entry.
The LSU refers to the ERT 255 only in cases where the RA is to be used for completing execution of an instruction. As described herein, the ERT 255 is consulted by the LSU for the following four functions, 1. Ifetch, Load or store missing the L1 cache; 2. stores from another thread within the core; 3. Snoop (XI) from another core; and 4. TLB and SLB invalidation.
In the first case of Ifetch, Load or store missing the L1 cache, the EA and thread_id are used to index into the ERT 255 and the RA from the corresponding ERT entry is sent to the L2 cache if a valid ERT entry exists. In case of an ERT miss, that is a valid ERT entry does not exist for the EA and thread_id, the SLB/TLB is used.
In the second case, where stores from another thread within the core, a store drained from the SRQ checks the ERT 255. If there is no hit from a different thread, then there is no load from another thread that is using the same RA. If there is a hit from a different thread using the same RA, the LSU checks the LRQ. Although, rare, in case a hit from another thread exists if the RA is used by another thread(s). Accordingly, the LSU looks up the ERT table 400 to find the relevant EA(s) for the common RA. The EA(s) are then used to look into the LRQ for a match (reject any store issue in that cycle). LRQ is partitioned per thread, so the LSU only looks into relevant thread's LRQ. If there is matching load(s) in the LRQ, the LSU flushes the oldest of the matching load(s).
In the third case of a snoop from another core of the processor, the LSU works similar to the second case, and checks for a hit from any of the other threads being executed. In case the TLB/SLB are invalidated, the ERT 255 is also invalidated.
The LRQF 218 in conjunction with the ERT table 400 is used for detecting and handling hazards, such as LHL, SHL, and LHS. For example, each store instruction is checked against the LRQF 218 for SHL detection, and in case a load instruction exists in the LRQF 218 for the same EA as the store instruction, the store instruction and other entries therefrom in the IFU are flushed, or the load instruction from the LRQF 218 is flushed. Further, in case of DCB instructions, the instructions are checked against the LRQF 218 for SHL cases and a load and everything after the DCB are flushed, or the load is flushed. For each data set that the system 100 wants to process, a corresponding data control block (DCB) and data definition (DD) statement or its dynamic allocation equivalent is to be configured.
Further, in one or more examples, as data are moved into and out of physical storage locations in system memory (e.g., in response to the invocation of a new process or a context switch), the entries in the TLB are updated to reflect the presence of the new data, and the TLB entries associated with data removed from system memory (e.g., paged out to nonvolatile mass storage) must be invalidated. Typically, the invalidation of TLB entries is the responsibility of software and is accomplished through the execution of an explicit TLB invalidate entry instruction (e.g., TLBIE in the POWER™ instruction set architecture (ISA)). The LRQF 218 provides TLBIE support by facilitating a TLBIE instruction to mark the entries in the LRQF 218, where the LRQF 218 indicates if any valid entries are marked.
Further, each load instruction is checked against the LRQF 218 for LHL detection (sequential load consistency), and in case of an LHL, the younger load instruction is flushed, or everything after the older load is flushed.
Further yet, each LARX instruction is checked against the LRQF 218 for larx-hit-larx cases, and in case the situation is detected, the younger LARX instruction is fluedh, or everything after the older LARX is flushed.
The technical solutions described herein thus facilitate hazard detection using only EA and without EA-to-RA translation for every load and store path, which is more expensive in terms of time as well as chip area (to store RA and CAM ports for translation). Further, the technical solutions facilitate an improved timing to detect SHL and suppress DVAL in time.
If the LSU experiences an EAD hit in the EAD 290, that is the EA of the instruction matches an entry in the EAD table 300, the LSU reads the contents of the matching EAD entry to determine a corresponding ERT entry, as shown at 520. Each EAD entry contains the ERT_ID (0:6) field 285. As described earlier, when an ERT entry is invalidated, the associated ERT_ID is invalidated, which also invalidates all associated entries in the EAD table 300. Accordingly, an EAD hit implies an ERT hit because using the ERT_ID field 285, an ERT entry can be found for the load/store instruction. Accordingly, in case of the EAD hit, after identifying the corresponding EAD entry, the LSU reads out ERT_ID from the EAD entry and sends to SRQ, LMQ, and/or LRQF, as shown at 530. The SRQ, LMQ, and/or LRQF use the EA from the EAD entry identified. In case of store instructions, which use RA, the RA from the ERT entry is read out for L2 access, as shown at 540 and 545. Thus, because the RA is not used anywhere else but the store instructions, core implementing the technical solutions herein is called EA-only core.
Now consider the case where the instruction misses the EAD 290, that is the EA of the instruction does not have a matching entry in the EAD table 300. The thread_id and EA are compared against each entry from the ERT 255, as shown at 550. If an ERT hit occurs, that is an ERT entry matches the parameters, the LSU reads out the RA (8:51) from the ERT entry, as shown at 555, and 530. For load requests the LSU sends the RA to the L2 cache for access, 530. For store instructions, the LSU stores the RA in the SRQ and then sends the RA to the L2 cache when the store drains to the L2 cache, as shown at 540-545.
If ERT miss occurs, the LSU initiates a reload for the ERT 255, as shown at 555 and 560. Further, ERT entry replacement is initiated. The ERT entry replacements are LRU based, and the LSU ensures to track synonyms within OoO window during this process.
Thus, by implementing the above method for a load, if there is an EA-hit in the EA-based L1 directory, then no address translation is performed. This improves on the typical processor where, the L1 directory is RA-based, which in case of a miss at the L1 directory for a load, causes the EA to be sent to an ERAT table for translation to get the RA that is sent to the L2 directory and beyond.
Further, for stores, with the method described herein the LSU has to go through the ERT table to determine the RA that is then stored in the SRQR to drain down to the caches (L1, L2, memory) when the store gets drained out of the SRQ. The SRQR holds all the RA for the stores. The RA is only stored for draining to the Nest (that is, L2, memory, and other units of the memory subsystem). The RA is not used for load-hit-store, store-hit-load or load-hit-load type of OoO execution hazard detection as it is done in the typical solutions. The RA calculation for stores happens before the store is completed, because after completion the LSU is unable to process any interrupt for the store (store can generate an address translation related interrupt, which is to be processed before the store is completed). Here the RA calculation is done when the store is issued (from the SRQR), thus preventing the LSU from having to perform the address translation. Thus, stores gets issued and executed OoO, and then get completed in-order, and subsequently the stores get drained from the SRQ in-order. Until a store is drained, no other thread or core knows about the store (only the current thread knows). After the store is drained from the SRQ, it is written into the L1 (if the line already exists in L1) and L2 caches (if caching is enabled) and at that point the store is known to all other threads and cores in the system 100.
For instruction fetches that miss the EA-based L1 I-Cache, the EA is translated to RA using the ERT 255 and the RA is sent to the Nest to fetch the I-Cache line. Here, the LHS (load-hit-store), SHL (store-hit-load) and LHL (load-hit-load) are all determined based on the EA and the ERT-Index stored with the directory entries in the EA-based L1 cache (EAD 290). All entries in the EAD table 300 have their translation valid in the ERT table 400, which can be used once LHS, SHL, and LHL are determined. If an ERT entry is invalidated, then the corresponding L1 cache entries are invalidated.
The LRQF, which is the load-reorder-queue, ensures that all load operations are tracked from dispatch to complete. When a load is rejected (for cache miss, or translation miss, or previous instruction it depends on got rejected) the load is taken out of the issue queue and placed in the LRQE for it to be re-issued from there.
If the existing entry(ies) in the ERT 255 is found with the same RA as the received RA (reloading RA), the ERT 255 compares page size(0:1) of the existing entry(ies) with that for the received RA, as shown at 620. If the page size of the existing entries is smaller than that for the reloading RA, the existing entries for that RA are removed from the ERT 255, and a new entry with a new ERT ID is added for the RA with the larger page size, as shown at 625. If the existing entry has a same or larger page size, and if the implementation uses an SDT, an entry is created in the SDT for the reloading RA, as shown at 627.
If the page size of the existing entries is the same size as the reloading RA, the ERT 255 checks if the existing entry is on the local ERT for the executing thread, as shown at 630. A local ERT in this case refers to the ERT being associated with the thread that is being executed, for example ERT0 for first thread and ERT1 for a second thread. If the RA hit is in the other ERT, that is the ERT that is not the local ERT, the ERT 255 creates a new entry in the local ERT with ERT_ID matching that in the non-local ERT, as shown at 632. For example, if the RA hit is in ERT1 for instruction executing by thread-0, an entry is created in the ERT0 with matching ERT_ID as the entry in the ERT1.
If the RA hit is on the local ERT instance, and if the EA also matches, because both EA and RA matche with an existing entry, but there was a ERT miss for this thread prompting the ERT reload, the ERT deems that this indicates that the two threads are sharing the same EA-RA mapping (with the same page size). Accordingly, the tid_en(0) or tid_en(1) bit in the existing matching entry for the bit corresponding to the reload thread is turned ON to indicate this case, as shown at 634.
If the RA hit is on local ERT instance, the EA does not match the existing entry, and if the existing entry is for the same thread as the reloading RA, the ERT identifies the aliasing case where two different EA maps to the same RA from the same thread, as shown at 636. If the processor is using an SDT-based implementation, a synonym entry is installed in the SDT that maps to the ERT ID, EA Offset(40:51) of the existing matching entry.
If the RA hit is on the local ERT instance, the EA does not match the existing entry, and if the existing entry is for a different thread, the ERT identifies the aliasing case where two EAs map to same RA from different threads, as shown at 638. If the processor is using an SDT-based implementation, a synonym entry is installed in the SDT that maps to the ERT ID, EA Offset(40:51) of the existing matching entry.
The above method facilitates that in SDT-based implementation, when two threads have the same RA but different EA, one of the translation uses the ERT entry and the other will use the SDT entry. Thus, the ERT entries facilitate the case where the same EA, and same RA is used across different threads by having a tid_en field in the ERT entry. For example, Tid_en(0:1)={tid 0 en, tid 1 en} on ERT0 instance; and Tid_en(0:1)={tid 1 en, tid 1 en} on ERT1 instance. Further, the ERT entry facilitates the case where same EA correspond to different RAs across different threads by having multiple entries in the ERT0 and ERT1 with their respective thread identifiers. The ERT entries also support the case with different EAs corresponding to the same RA (either same or different thread cases). Two cases are now described based on the implementation using SDT.
The LSU installs an entry in the SDT (synonym detection table) instead of the ERT 255 when a new instruction is detected with different EA corresponding to the same RA at ERT reload time. The SDT hits relaunch with the original (or earlier) ERT entry's EA. If new synonym page size is bigger than the page size in the existing ERT entry with matching RA, then the existing ERT entry is replaced by the new synonym (with the larger page size) instead of installing a synonym in the SDT. The old ERT entry is eventually reinstalled as synonym in the SDT.
Further, referring back to the ERT cases, consider the case where the LSU receives a snoop from another core from the processor 106. The snoop can be from a different core in the system (snoop indiecates another core or thread, changed the data at the same real address). The LSU also checks stores from a thread within the core as a potential snoop to the other threads within the core. All snoops (from other cores) or stores (from other threads within the core) comes with a RA. In such cases, the LSU reverse translates the RA to determine corresponding EA, ERT_ID, and page size based on the ERT 255. The LSU compares this information with the ERT_ID, PS, EA(40:56) stored in each of the following structures to detect a snoop hit and take the appropriate action. For example, if a snoop hit is detected in LRQF Entry, the LSU indicates a potential load-hit-load out of order hazard. If a snoop hit is detected in the EAD 290, the LSU initiates an L1 invalidate, if the snoop is from a different core. If the store is from another thread for a shared line, then the line automatically gets the new store and is updated.
If the LSU uses an SDT, and if a snoop hit exists in the LMQ, the LSU also updates the LMQ entry to not store in L1 Dcache, the SRQ Entry is not used for snoops in SRQ, only used for LHS EA miss RA hit style checking, and a new SDT entry is created for the snoop hit.
Thus, as described earlier, a technical challenge with the LSU that is based on EA-only with an execution flow described herein is that of an EA synonym on a thread. For example, same thread EA synonyms (that is, two different EA from a thread map to the same RA). Such a technical challenge may be across LHS, SHL, LHL, if the OoO window includes at least two L1 accesses such as:
Tid=w, EA(0:51)=x=>RA(8:51)=z; and
Tid=w, EA(0:51)=y =>RA(8:51)=z.
The technical solutions described herein address the technical challenge by using a SDT, which is another LSU subunit. In one or more examples, the SDT includes 16 entries for handling cases where different EAs have same RA. Such different EAs are referred to as synonyms, because they all are converted to the same RA. The SDT is a table of such synonyms in the LSU. The SDT is accessed at AGEN (address generation) on ERT misses. In one or more examples, the SDT may be accessed with restrictions, such as only if there is a valid SDT entry for the thread of the L/S operation being agen,
An entry is added into the SDT 800 during the ERT reload as described herein. For example, during ERT reload, the reload RA is compared against valid ERT entries. If an ERT entry with matching RA already exists and it is not an EA hit case where just an additional tid_en bit is being set in the original ERT entry, then the EA(32:51) from the existing ERT entry is read, and an entry is installed into the SDT 800 instead of adding an entry to the ERT 255.
Because SDT 800 has limited number of entries, the entries are replaced. In one or more examples, the entries are replaced based on least recently used (LRU) techniques, or any other order. In one or more examples, if an SDT entry is replaced, a subsequent launch using the secondary address re-triggers the SDT entry installation sequence. Further yet, CAM clears SDT entry(s) with ERT ID matching an ERT entry invalidate.
Now, if most subsequent accesses to RAz are with EAy, the LSU has to use the SDT more frequently than using the EAD itself. In one or more examples, the technical solutions to improve the efficiency of the LSU, by reducing such frequent trips to the SDT include providing an increment counter in each SDT entry. As depicted in
In one or more examples, the counter is a 4-bit field, implying a maximum value of 15. It should be understood that the field is of different length in other examples, and/or have different maximum value, which is used as a threshold. For example, after the instruction has been launched, the counter value is compared with the threshold, as shown at 845 and 850. If the counter is below the threshold, the LSU continues to operate as described. If the counter exceeds the threshold, or in some cases is equal to the threshold, the LSU invalidates the ERT entry corresponding to the SDT entry, as shown at 860. For example, the ERT entry with the ERT ID from the SDT entry is invalidated. The invalidation of the ERT entry causes the corresponding entries to be invalidated from the EA Directory, LRQF, LMQ, and SRQ.
Further, the LSU addresses a technical challenge of exceptions in a launched instruction requiring original EA to finish in the following manner. For example, consider the case where the launched instruction gets an SDT hit and wants to relaunch with replace address from the SDT entry instead of the original launch address, but an exception is taken that requires the original EA to finish. Such condition may occur in case of DAWR/SDAR, etc.
The LSU implementing the technical solutions described herein address such technical challenge by maintaining the original address in a queue in the LRQE. The LRQE also keeps an SDT hit flag (bit), an SDT Index(0:3) for each LRQE entry. When relaunching, the SDT index is read a cycle early to get the replace address. The LRQE further multiplexes between the LRQE entry address (original address) and the SDT replace address (read from the SDT) before relaunching. For exception cases, such as the above, where the original address is required to finish, the LRQE has an additional SDT hit override flag (bit) for each entry set on DAWR partial match, etc. The LRQE relaunches the case where there was an SDT hit that finishes with an exception and forces the original address to be launched. An SRQ relaunch is similar to the LRQE relaunch as described here, where the SDT hit override flag is used when it is determined before relaunch to finish with exception.
The technical solutions described herein thus facilitate using only the EA, providing technical advantages such that an ERAT (which was typically used in processors) is not referred to in a load/store path, and further that detecting SHL and suppressing DVAL in time do not cause timing problems. Further, the technical solutions described herein address technical problems with using only the EA, for example that LHS, SHL, LHL detection can miss when two different EA map to the same RA. The technical solutions described herein address such technical problem by either using a Synonym Detection Table (SDT) for the instructions in the OoO window. The technical solutions provide various technical advantages including reduction in chip area (by not storing RA), reduction in power consumption (by not translating EA-RA), and improvements in latency, among others.
Turning now to
In an exemplary embodiment, as shown in
The I/O devices 1047, 1045 may further include devices that communicate both inputs and outputs, for instance disk and tape storage, a network interface card (NIC) or modulator/demodulator (for accessing other files, devices, systems, or a network), a radio frequency (RF) or other transceiver, a telephonic interface, a bridge, a router, and the like.
The processor 1005 is a hardware device for executing hardware instructions or software, particularly those stored in memory 1012. The processor 1005 may be a custom made or commercially available processor, a central processing unit (CPU), an auxiliary processor among several processors associated with the computer system 1000, a semiconductor based microprocessor (in the form of a microchip or chip set), a microprocessor, or other device for executing instructions. The processor 1005 can include a cache such as, but not limited to, an instruction cache to speed up executable instruction fetch, a data cache to speed up data fetch and store, and a translation look-aside buffer (TLB) used to speed up virtual-to-physical address translation for both executable instructions and data. The cache may be organized as a hierarchy of more cache levels (L1, L2, etc.).
The memory 1012 may include one or combinations of volatile memory elements (e.g., random access memory, RAM, such as DRAM, SRAM, SDRAM, etc.) and nonvolatile memory elements (e.g., ROM, erasable programmable read only memory (EPROM), electronically erasable programmable read only memory (EEPROM), programmable read only memory (PROM), tape, compact disc read only memory (CD-ROM), disk, diskette, cartridge, cassette or the like, etc.). Moreover, the memory 1012 may incorporate electronic, magnetic, optical, or other types of storage media. Note that the memory 1012 may have a distributed architecture, where various components are situated remote from one another but may be accessed by the processor 1005.
The instructions in memory 1012 may include one or more separate programs, each of which comprises an ordered listing of executable instructions for implementing logical functions. In the example of
Additional data, including, for example, instructions for the processor 1005 or other retrievable information, may be stored in storage 1027, which may be a storage device such as a hard disk drive or solid state drive. The stored instructions in memory 1012 or in storage 1027 may include those enabling the processor 1005 to execute one or more aspects of the dispatch systems and methods of this disclosure.
The computer system 1000 may further include a display controller 1025 coupled to a display 1030. In an exemplary embodiment, the computer system 1000 may further include a network interface 1060 for coupling to a network 1065. The network 1065 may be an IP-based network for communication between the computer system 1000 and an external server, client and the like via a broadband connection. The network 1065 transmits and receives data between the computer system 1000 and external systems. In an exemplary embodiment, the network 1065 may be a managed IP network administered by a service provider. The network 1065 may be implemented in a wireless fashion, e.g., using wireless protocols and technologies, such as WiFi, WiMax, etc. The network 1065 may also be a packet-switched network such as a local area network, wide area network, metropolitan area network, the Internet, or other similar type of network environment. The network 1065 may be a fixed wireless network, a wireless local area network (LAN), a wireless wide area network (WAN) a personal area network (PAN), a virtual private network (VPN), intranet or other suitable network system and may include equipment for receiving and transmitting signals.
Systems and methods for providing address translation for sending real address to memory subsystem in effective address based load-store unit can be embodied, in whole or in part, in computer program products or in computer systems 1000, such as that illustrated in FIG.10.
Various embodiments of the invention are described herein with reference to the related drawings. Alternative embodiments of the invention can be devised without departing from the scope of this invention. Various connections and positional relationships (e.g., over, below, adjacent, etc.) are set forth between elements in the following description and in the drawings. These connections and/or positional relationships, unless specified otherwise, can be direct or indirect, and the present invention is not intended to be limiting in this respect. Accordingly, a coupling of entities can refer to either a direct or an indirect coupling, and a positional relationship between entities can be a direct or indirect positional relationship. Moreover, the various tasks and process steps described herein can be incorporated into a more comprehensive procedure or process having additional steps or functionality not described in detail herein.
The following definitions and abbreviations are to be used for the interpretation of the claims and the specification. As used herein, the terms “comprises,” “comprising,” “includes,” “including,” “has,” “having,” “contains” or “containing,” or any other variation thereof, are intended to cover a non-exclusive inclusion. For example, a composition, a mixture, process, method, article, or apparatus that comprises a list of elements is not necessarily limited to only those elements but can include other elements not expressly listed or inherent to such composition, mixture, process, method, article, or apparatus.
Additionally, the term “exemplary” is used herein to mean “serving as an example, instance or illustration.” Any embodiment or design described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments or designs. The terms “at least one” and “one or more” may be understood to include any integer number greater than or equal to one, i.e. one, two, three, four, etc. The terms “a plurality” may be understood to include any integer number greater than or equal to two, i.e. two, three, four, five, etc. The term “connection” may include both an indirect “connection” and a direct “connection.”
The terms “about,” “substantially,” “approximately,” and variations thereof, are intended to include the degree of error associated with measurement of the particular quantity based upon the equipment available at the time of filing the application. For example, “about” can include a range of ±8% or 5%, or 2% of a given value.
For the sake of brevity, conventional techniques related to making and using aspects of the invention may or may not be described in detail herein. In particular, various aspects of computing systems and specific computer programs to implement the various technical features described herein are well known. Accordingly, in the interest of brevity, many conventional implementation details are only mentioned briefly herein or are omitted entirely without providing the well-known system and/or process details.
The present invention may be a system, a method, and/or a computer program product. The computer program product may include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present invention.
The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium may be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.
Computer readable program instructions described herein can be downloaded to respective computing/processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and/or a wireless network. The network may comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and/or edge servers. A network adapter card or network interface in each computing/processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing/processing device.
Computer readable program instructions for carrying out operations of the present invention may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like, and conventional procedural programming languages, such as the “C” programming language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present invention.
Aspects of the present invention are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer readable program instructions.
These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and/or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function/act specified in the flowchart and/or block diagram block or blocks.
The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable apparatus, or other device implement the functions/acts specified in the flowchart and/or block diagram block or blocks.
The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts or carry out combinations of special purpose hardware and computer instructions.
The descriptions of the various embodiments of the present invention have been presented for purposes of illustration, but are not intended to be exhaustive or limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein was chosen to best explain the principles of the embodiments, the practical application or technical improvement over technologies found in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.
This application is a continuation of U.S. Non-Provisional application Ser. No. 15/825,494, entitled “HANDLING EFFECTIVE ADDRESS SYNONYMS IN A LOAD-STORE UNIT THAT OPERATES WITHOUT ADDRESS TRANSLATION”, filed Nov. 29, 2017, which is incorporated herein by reference in its entirety and is a continuation of U.S. Non-Provisional application Ser. No. 15/726,596, entitled “HANDLING EFFECTIVE ADDRESS SYNONYMS IN A LOAD-STORE UNIT THAT OPERATES WITHOUT ADDRESS TRANSLATION”, filed Oct. 6, 2017, which is also incorporated herein by reference in its entirety.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 15825494 | Nov 2017 | US |
| Child | 16793835 | US | |
| Parent | 15726596 | Oct 2017 | US |
| Child | 15825494 | US |
