Updated: 11 December 1998

VAX MACRO and Instruction Set Reference Manual

While MSYNC performs scalar/vector memory synchronization, it does more than that. MSYNC allows software to ensure that all previously issued memory instructions of the scalar/vector processor pair are complete and their results made visible before the scalar processor proceeds with the next instruction.

MSYNC is implemented through the nonprivileged MFVP instruction. Arithmetic and asynchronous memory management exceptions encountered by previous vector instructions can cause MSYNC to fault.

Once it issues MSYNC, the scalar processor executes no further instructions until MSYNC completes or faults.

MSYNC completes when the following events occur:

• All previously issued scalar and vector memory instructions have completed.
• All resultant memory write operations (scalar write operations and vector store operations) have been made visible to both the scalar and vector processor.
• No exception that should cause MSYNC to fault has occurred. (See the next paragraph.)

MSYNC faults when any unreported exception has occurred in the production or storage of any result (vector register element or vector control register bit) that MSYNC depends upon. Such results include all elements loaded or stored by a previously issued vector memory instruction as well as any element or control register bit that these elements depend upon.

It is UNPREDICTABLE whether MSYNC faults due to exceptions that occur in the production and storage of results (vector register elements and vector control register bits) that MSYNC does not depend upon. Software should not rely on such exceptions being reported by MSYNC for program correctness.

When MSYNC completes, a longword value (which is UNPREDICTABLE) is returned to the scalar processor, which writes it to the scalar destination of the MFVP. The scalar processor then proceeds to execute the next instruction. If the scalar destination is in memory, it is UNPREDICTABLE whether the new value of the destination becomes visible to the vector processor until another scalar/vector memory synchronization instruction is performed.

When MSYNC faults, it is not ensured that all previously issued scalar and vector memory instructions have finished. In this case, the scalar processor writes no longword value to the scalar destination of the MFVP. Depending on the exception encountered by the vector processor, the MSYNC takes a vector processor disabled fault or memory management fault. Note that it is UNPREDICTABLE whether the vector processor is idle when the fault is generated. After the fault has been serviced, the MSYNC may be returned to through an REI.

Section 10.5.3.3 gives the necessary rules and examples to determine what vector control register elements and vector control register bits MSYNC depends upon.

10.7.2.2 Memory Activity Completion Synchronization (VMAC)

Privileged software needs a way to ensure scalar/vector memory synchronization that will not result in any exceptions being reported. Reading the VMAC internal processor register (IPR) with the privileged MFPR instruction is provided for these situations. It is especially useful for context switching.

Once a MFPR from VMAC is issued by the scalar processor, the scalar processor executes no further instructions until VMAC completes, which it does when the following events occur:

• All vector and scalar memory activities have ceased.
• All resultant memory write operations have been made visible to both the scalar and vector processor.
• A longword value (which is UNPREDICTABLE) is returned to the scalar processor.

After writing the longword value to the scalar destination of the MFPR, the scalar processor then proceeds to execute the next instruction. If the scalar destination is in memory, it is UNPREDICTABLE whether the new value of the destination becomes visible to the vector processor until another scalar/vector memory synchronization operation is performed.

As stated in Section 10.7.2, Scalar/Vector Memory Synchronization, the ceasing of vector and scalar memory activities does not mean that previously issued vector memory instructions have completed. For example, consider a vector memory instruction that has suspended execution due to an asynchronous memory management exception or hardware error. Once it becomes suspended, the instruction will write no further elements and its memory activity will cease. As a result, a subsequently issued VMAC will complete as soon as those write operations that were made by the memory instruction before it was suspended are visible to both the scalar and vector processor. But, after the completion of the VMAC, the memory instruction is not completed and remains suspended.

Vector arithmetic and memory management exceptions of previous vector instructions never fault an MFPR-from-VMAC and never suspend its execution.

10.7.3 Other Synchronization Between the Scalar and Vector Processors

Synchronization between the scalar and vector processors also occurs in the following situations:

• In the absence of pending vector arithmetic exceptions, reading a vector control register using the MFVP instruction waits for all previous write operations to that register to complete. In addition, the scalar processor must wait for the MFVP result to be written before processing other instructions. An MFVP instruction that reads a vector control register must fault if there is any unreported exception that has occurred in the production of the value of the control register.
• Writing to VTBIA or VSAR with MTPR causes the new state of the changed vector IPR to affect the execution of all subsequently issued vector instructions.
• Reading from VPSR with MFPR after writing to VPSR with MTPR causes the new state of VPSR (and VAER if cleared by VPSR<RST>) to affect the execution of subsequently issued vector instructions.

The vector processor may concurrently execute a number of vector memory instructions through the use of multiple load/store paths to memory. When it is necessary to synchronize the accesses of multiple vector memory instructions the MSYNC instruction can be used; however, there are cases for which this instruction does more than is needed. If it is known that only synchronization between the memory accesses of vector instructions is required, the VSYNC instruction is more efficient.

VSYNC orders the conflicting memory accesses of vector-memory instructions issued after VSYNC with those of vector-memory instructions issued before VSYNC. Specifically, VSYNC forces the access of a memory location by any subsequent vector-memory instruction to wait for (depend upon) the completion of all prior conflicting accesses of that location by previous vector-memory instructions.

VSYNC does not have any synchronizing effect between scalar and vector memory access instructions. VSYNC also has no synchronizing effect between vector load instructions because multiple load accesses cannot conflict. It also does not ensure that previous vector memory management exceptions are reported to the scalar processor.

10.7.5 Required Use of Memory Synchronization Instructions

Table 10-15 shows for all possible pairs of vector or scalar read and write operations to a common memory location, whether one of the scalar/vector memory synchronization instructions or the VSYNC instruction must be issued after the first reference and before the second. Since the MSYNC instruction also includes the VSYNC function, it can always be used instead of VSYNC.

In general, these rules apply to any sequence of instructions that access a common memory location, no matter how many other vector or scalar instructions are issued between the first instruction that accesses the common location and the second instruction that accesses the same location. For example, the following code sequence depicts a vector load followed by a scalar write operation to the same memory location. Between these two instructions are other scalar/vector instructions that do not access the common memory location. A scalar/vector memory synchronization instruction (MSYNC or VMAC) must be executed sometime after the vector read operation and before the scalar write operation to the common location. (Here MSYNC is shown.)

VLDL A, #4, V0 . other scalar/vector instructions that do not access A . MSYNC Dst MOVL R0, A

In most cases, MSYNC is the preferred method for ensuring scalar/vector memory synchronization. However, there are special cases, usually encountered by an operating system, when VMAC is more appropriate.

Cases when scalar/vector memory synchronization is required are as follows:

• After a vector instruction that stores to memory and before a peripheral (I/O) data transfer of the stored location is initiated by an application program. This ensures that the value stored will be transferred to the output device. The application must ensure that this requirement is met by using MSYNC. Using VMAC in this case is not sufficient because unlike MSYNC, VMAC does not ensure that all previous vector memory instructions have successfully completed.
• After a vector instruction that stores to memory and before the associated scalar processor can execute a HALT instruction. This ensures that a read operation or modify operation by another processor will access the updated memory value. VMAC is the preferred method for this case.
• Before the vector processor state is saved as a result of power failure. A read or modify operation of the same memory must read the updated value (provided that the duration of the power failure does not exceed the maximum nonvolatile period of the main memory). Also, software is responsible for saving any pending vector processor exception status. VMAC is the preferred method for this case.
• Before a context switch. Software is responsible for ensuring that the vector processor has completed all its memory accesses before performing a context switch. Software is also responsible for saving any pending vector processor exception status. VMAC is the preferred method for this case.

The scalar/vector memory synchronization instructions are the only ones that guarantee that the memory operations of the vector and scalar processors are synchronized. Write operations to I/O space, changes in access mode, machine checks, interprocessor interrupts, execution of a HALT, REI, or interlocked instruction do not make the results of vector instructions that write to memory visible to the scalar processor, I/O subsystem, or other processors. Execution of a scalar/vector memory synchronization instruction must precede any of these mechanisms to ensure synchronization of all system components.

Table 10-15 Possible Pairs of Read and Write Operations When Scalar/Vector Memory Synchronization (M) or VSYNC (V) Is Required Between Instructions That Reference the Same Memory Location

First Reference Second Reference	Scalar Scalar	Scalar Vector	Vector Scalar	Vector Vector
Operation Sequence
Read, Read	No 1,2	No 1	No 1	No 1
Read, Write	No 2	No 3	M	V 5
Write, Read	No 2	M 4	M	V
Write, Write	No 2	M 4	M	V

There exist conditions when VSYNC is not required between conflicting vector memory accesses. A VSYNC is not required before a vector memory store instruction (VST/VSCAT) if, for each memory location to be accessed by the store, both of the following conditions are met:

• Each of the store's accesses to the location does not conflict with any access to the location by previously issued vector store instructions. Conflict is avoided in this case because one of the following events occurred:
  • The location is not shared.
  • All accesses to the location by previous store instructions were forced to complete by the issue of an MSYNC or VMAC.
• Each of the store's accesses to the location does not conflict with any access to the location by previously issued vector load (VLD/VGATH) instructions. Conflict is avoided in this case because one of the following events occurred:
  • The location is not shared.
  • All accesses to the location by previous load instructions were forced to complete by the issue of an MSYNC or VMAC.
  • Each of the store's accesses to the location depends on the completion (as seen by the vector processor) of all accesses to the location by previous LOAD instructions. (The examples immediately following demonstrate this concept.)

In all other cases of conflicting vector memory accesses, VSYNC is necessary to ensure correct results.

Examples Where VSYNC Is Not Required

In the following examples, VSYNC is not required because both of the previous conditions have been met for each location accessed by the store instruction:

VLDL       A, #4, V0           
VSTL       V0, A, #4               
      

VLDL       A, #4, V0 
VSSUBL     R0, V0, V1 
VSTL       V1, A, #4 
      

VLDL/0      A, #4 ,V0          
VSMULL/0    #3, V0, V0             
VLDL/1      A, #4 ,V1              
VVMULL/1    V1, V1, V1             
VVMERGE/1   V1, V0, V2             
VSTL        V2, A, #4              
      

VLDL        A, #4 ,V0          
VSGTRF     #0, V0 
VLDL/1     B, #4, V1 
VLDL/0     C, #4, V2 
VVMERGE/0  V2, V1, V3 
VSTL       V3, A, #4 
      

Examples Where VSYNC Is Required

In the following examples, VSYNC is required before the vector memory store instruction:

VLDL/1     A,#4,V0 
VSLSSL     #0,V1 
VSYNC 
VSTL/1     V1,A,#4 
      

If the VSYNC is not included, V0 could contain incorrect data at the end of the sequence since the vector processor is allowed to begin the VSTL before the VLDL is finished. This occurs because there is no dependence between the VMR value used by the VLDL and the VSTL.

VLDL       A, #4, V0 
VVMERGE/0  V0, V1, V1 
VSYNC 
VSTL       V1, A, #4 
      

Unless the programmer can ensure that the VMR mask being used by the VVMERGE will force the access of each location by the VSTL to depend on the access to that location by the VLDL, a VSYNC is required. Note that in general, when masked operations provide a conditional path of dependence between conflicting memory accesses, a VSYNC is usually necessary to ensure correct results.

VSTL       V1, A, #4 
MTVLR      #32 
VSYNC 
VLDL       A+128, #4, V2 
      

In this example, the VSTL writes locations A to A+255 and the VLDL reads locations A+128 to A+255. Without the VSYNC, the vector processor is allowed to start reading locations A+128 to A+255 for the VLDL before the vector processor completes (or even starts) writing locations A+128 to A+255 for the VSTL. Consequently, V2[0:31] will not contain V1[32:63], which is the intended result. Note that the rules on when VSYNC is not required (found in Section 10.7.5.1) only apply to waiving the use of VSYNC prior to VST/VSCAT instructions.

VGATHL      A, V2, V0    ; let at least two elements 
                         ; of V2 be equal 
VVMULL      V9, V0, V1 
VSYNC 
VSCATL      V1, A, V2 
      

The VSYNC is needed in this example because the VSCATL may store elements of V1 into a common location before the VGATHL has finished loading that location into all the appropriate elements of V0. As a result, elements of V0 fetched from the same location may be unequal. Suppose in the example that V2[0] = V2[63] = 0 and that the original value of location A before the sequence starts is X. Then it is possible without the VSYNC that V0[63] = X*V9[0] and that (A)= V1[63] = V9[63]*V9[0]*X after the sequence completes.

VLDL        A, #0, V0 
VVMULL      V9, V0, V1 
VSYNC 
VSTL        V1, A, #0 
      

The VSYNC is needed in this example because the VSTL may store elements of V1 into A before the VLDL has finished loading all elements of V0 from A. As a result, the elements of V0 may be unequal and so produce incorrect results.

The vector processor may include its own translation buffer and maintain its own copies of SBR, SLR, SPTEP, P0BR, P0LR, P1BR, and P1LR as a group, or may use the scalar processor's memory management unit. Hardware implementations must ensure that MTPR to these registers update the copy retained by the vector processor. Changes to P0BR, P0LR, P1BR, and P1LR due to a LDPCTX do not update the copies in the vector processor. Before software enables the vector processor again, explicit MTPRs to P0BR, P0LR, P1BR, and P1LR are required to guarantee correct operation.

An MTPR to TBIS must also invalidate the corresponding TB entry in the vector processor, and an MTPR to TBIA must also invalidate the entire TB in the vector processor. However, the vector TB is not invalidated by a LDPCTX instruction. Software can use an MTPR to the Vector TB Invalidate All (VTBIA) register to invalidate only the vector TB. An MTPR to VTBIA results in no operation on a processor that uses a common TB for the scalar and vector processors.

Updates to memory management registers and invalidates of translation buffer entries in the vector processor take place even when the vector processor is disabled (VPSR<VEN> is clear). However, the vector processor may load translation buffer entries only when the vector processor is executing a vector memory access instruction.

The vector processor implements the modify-fault option if its scalar processor implements the virtual-machine option.

Vector memory access instructions must not be used to read or write page tables. If a vector instruction is used to read or write page tables, the results are UNPREDICTABLE.

Vector instructions are not allowed to reference I/O space. If a vector instruction references I/O space, the results are UNPREDICTABLE.

Issuing vector instructions with memory management disabled causes the operation of the vector processor to be UNDEFINED. Disabling memory management when the vector processor is busy (VPSR<BSY> is set) also causes the operation of the vector processor to be UNDEFINED.

10.9 Hardware Errors

A vector processor implementation may experience error conditions (such as chip malfunctions, parity errors, or bus errors) that prevent it from executing and completing instructions and from which it cannot recover through its own means. Such errors are termed hardware errors and may occur at anytime, even when the vector processor is already disabled. Vector processor hardware errors do not normally halt the scalar processor.

At some point after the error condition occurs, the vector processor reports the error to the scalar processor. The reporting may be accomplished through a machine check; or by disabling the vector processor, setting VPSR<IMP>, and generating a vector processor disabled fault when the next vector instruction is issued. After the error is reported, the appropriate software handler will be invoked to diagnose the vector processor and to determine the severity of the hardware error and whether the vector processor can be restarted.

During execution, software may wish to force the reporting of hardware errors encountered by previous vector instructions before issuing further ones. This can be accomplished by reading the VMAC internal processor register (IPR) and by waiting for VPSR<BSY> to become clear.

An MFPR from VMAC ensures that all pending vector memory instructions have finished or are suspended by an asynchronous memory management exception, and that all vector-processor hardware errors encountered by these instructions are reported by the time the MFPR completes. Errors are handled as follows:

• If the errors are reported by machine check, then the exception is taken either upon the VMAC itself, or upon the instruction immediately following the VMAC.
• If the errors are reported through VPSR<IMP>, the vector processor sets VPSR<IMP> and disables itself by the time the scalar processor completes VMAC. Subsequently, a vector processor disabled fault will occur when the next vector instruction is issued. A read of VPSR immediately after the VMAC completes will find the vector processor disabled and VPSR<IMP> set.

Waiting for VPSR<BSY> to become clear before issuing further instructions ensures that all previous non-memory-access instructions have been finished or are suspended by an asynchronous memory management exception, and that all vector-processor hardware errors encountered by these instructions are reported by the time VPSR<BSY> becomes clear. Errors are handled as follows:

• If the errors are reported by machine check, then the exception is taken either upon the first instruction during which the new state of VPSR<BSY> becomes visible to the scalar processor or upon the instruction immediately thereafter.
• If the errors are reported through VPSR<IMP>, the vector processor sets VPSR<IMP> and disables itself by the time it clears VPSR<BSY>. Subsequently, a vector processor disabled fault will occur when the next vector instruction is issued. The first MFPR instruction that reads VPSR<BSY> as clear will also read VPSR<VEN> as clear and VPSR<IMP> as set.

VMAC does not ensure that hardware errors encountered by pending non-memory-access instructions will be reported. Waiting for VPSR<BSY> to become clear does not ensure that vector-processor hardware errors encountered by vector memory instructions are reported.

Software can force the reporting of hardware errors encountered during the execution of previous vector instructions (both memory and non-memory) by waiting for VPSR<BSY> to become clear and then by issuing an MFPR from VMAC. This technique can be used during scalar context switching to cause hardware errors resulting from the execution of vector instructions for the current process to be reported before that process is context-switched.

Copyright © Compaq Computer Corporation 1998. All rights reserved. Legal

4515PRO_034.HTML