RTEMS CPU Architecture Supplement (6.9e8141c).¶

2.1.1. CPU Model Name¶

The macro CPU_MODEL_NAME is a string which designates the name of this CPU model. For example, for the MC68020 processor model from the m68k architecture, this macro is set to the string “mc68020”.

2.1.2. Floating Point Unit¶

In most architectures, the presence of a floating point unit is an option. It does not matter whether the hardware floating point support is incorporated on-chip or is an external coprocessor as long as it appears an FPU per the ISA. However, if a hardware FPU is not present, it is possible that the floating point emulation library for this CPU is not reentrant and thus context switched by RTEMS.

RTEMS provides two feature macros to indicate the FPU configuration:

• CPU_HARDWARE_FP is set to TRUE to indicate that a hardware FPU is present.

• CPU_SOFTWARE_FP is set to TRUE to indicate that a hardware FPU is not present and that the FP software emulation will be context switched.

2.3.1. Calling Mechanism¶

In each of the architecture specific chapters, this subsection will describe the instruction(s) used to perform a normal subroutine invocation. All RTEMS directives are invoked as normal C language functions so it is important to the user application to understand the call and return mechanism.

2.3.2. Register Usage¶

In each of the architecture specific chapters, this subsection will detail the set of registers which are NOT preserved across subroutine invocations. The registers which are not preserved are assumed to be available for use as scratch registers. Therefore, the contents of these registers should not be assumed upon return from any RTEMS directive.

In some architectures, there may be a set of registers made available automatically as a side-effect of the subroutine invocation mechanism.

2.3.3. Parameter Passing¶

In each of the architecture specific chapters, this subsection will describe the mechanism by which the parameters or arguments are passed by the caller to a subroutine. In some architectures, all parameters are passed on the stack while in others some are passed in registers.

2.3.4. User-Provided Routines¶

All user-provided routines invoked by RTEMS, such as user extensions, device drivers, and MPCI routines, must also adhere to these calling conventions.

2.4.1. Flat Memory Model¶

Most RTEMS target processors can be initialized to support a flat address space. Although the size of addresses varies between architectures, on most RTEMS targets, an address is 32-bits wide which defines addresses ranging from 0x00000000 to 0xFFFFFFFF (4 gigabytes). Each address is represented by a 32-bit value and is byte addressable. The address may be used to reference a single byte, word (2-bytes), or long word (4 bytes). Memory accesses within this address space may be performed in little or big endian fashion.

On smaller CPU architectures supported by RTEMS, the address space may only be 20 or 24 bits wide.

If the CPU model has support for virtual memory or segmentation, it is the responsibility of the Board Support Package (BSP) to initialize the MMU hardware to perform address translations which correspond to flat memory model.

In each of the architecture specific chapters, this subsection will describe any architecture characteristics that differ from this general description.

2.5.1. Vectoring of an Interrupt Handler¶

In each of the architecture specific chapters, this subsection will describe the architecture specific details of the interrupt vectoring process. In particular, it should include a description of the Interrupt Stack Frame (ISF).

2.5.2. Interrupt Levels¶

In each of the architecture specific chapters, this subsection will describe how the interrupt levels available on this particular architecture are mapped onto the 255 reserved in the task mode. The interrupt level value of zero (0) should always mean that interrupts are enabled.

Any use of an interrupt level that is is not undefined on a particular architecture may result in behavior that is unpredictable.

2.5.3. Disabling of Interrupts by RTEMS¶

During the execution of directive calls, critical sections of code may be executed. When these sections are encountered, RTEMS disables all external interrupts before the execution of this section and restores them to the previous level upon completion of the section. RTEMS has been optimized to ensure that interrupts are disabled for the shortest number of instructions possible. Since the precise number of instructions and their execution time varies based upon target CPU family, CPU model, board memory speed, compiler version, and optimization level, it is not practical to provide the precise number for all possible RTEMS configurations.

Historically, the measurements were made by hand analyzing and counting the execution time of instruction sequences during interrupt disable critical sections. For reference purposes, on a 16 Mhz Motorola MC68020, the maximum interrupt disable period was typically approximately ten (10) to thirteen (13) microseconds. This architecture was memory bound and had a slow bit scan instruction. In contrast, during the same period a 14 Mhz SPARC would have a worst case disable time of approximately two (2) to three (3) microseconds because it had a single cycle bit scan instruction and used fewer cycles for memory accesses.

If you are interested in knowing the worst case execution time for a particular version of RTEMS, please contact OAR Corporation and we will be happy to product the results as a consulting service.

Non-maskable interrupts (NMI) cannot be disabled, and ISRs which execute at this level MUST NEVER issue RTEMS system calls. If a directive is invoked, unpredictable results may occur due to the inability of RTEMS to protect its critical sections. However, ISRs that make no system calls may safely execute as non-maskable interrupts.

2.10.1. System Reset¶

An RTEMS based application is initiated or re-initiated when the processor is reset or transfer is passed to it from a boot monitor or ROM monitor.

In each of the architecture specific chapters, this subsection describes the actions that the BSP must take assuming the application gets control when the microprocessor is reset.

3.1.1. CPU Model Name¶

The macro CPU_MODEL_NAME is a string which designates the architectural level of this CPU model. See in cpukit/score/cpu/aarch64/rtems/score/aarch64.h for the values.

3.1.2. Floating Point Unit and SIMD¶

The Advanced SIMD (NEON) and Floating-point instruction set extension is supported and expected to be present since all ARMv8-A CPUs are expected to support it as per the ARMv8-A Programmer’s Guide Chapter 7 introduction (https://developer.arm.com/docs/den0024/a/aarch64-floating-point-and-neon). As such, CPU_HARDWARE_FP will always be set to TRUE.

3.5.1. Interrupt Levels¶

There are exactly two interrupt levels on ARMv8 with respect to RTEMS. Level zero corresponds to interrupts enabled. Level one corresponds to interrupts disabled.

3.5.2. Interrupt Stack¶

The board support package must initialize the interrupt stack. The memory for the stacks is usually reserved in the linker script. The interrupt stack pointer is stored in the EL0 stack pointer and is accessed by switching to SP0 mode at the beginning of interrupt calls and back to SPx mode after completion of interrupt calls using the spsel instruction.

4.1.1. CPU Model Name¶

The macro CPU_MODEL_NAME is a string which designates the architectural level of this CPU model. See in cpukit/score/cpu/arm/rtems/score/arm.h for the values.

4.1.2. Count Leading Zeroes Instruction¶

The ARMv5 and later instruction sets have the count leading zeroes clz instruction which could be used to speed up the find first bit operation. The use of this instruction should significantly speed up the scheduling associated with a thread blocking. This is currently not used.

4.1.3. Floating Point Unit¶

The following floating point units are supported:

• VFPv2 (for example available on ARM926EJ-S processors)

• VFPv3-D32/NEON (for example available on Cortex-A processors)

• VFPv3-D16 (for example available on Cortex-R processors)

• FPv4-SP-D16 (for example available on Cortex-M processors)

• FPv5-D16 (for example available on Cortex-M7 processors)

4.5.1. Interrupt Levels¶

There are exactly two interrupt levels on ARM with respect to RTEMS. Level zero corresponds to interrupts enabled. Level one corresponds to interrupts disabled.

4.5.2. Interrupt Stack¶

The board support package must initialize the interrupt stack. The memory for the stacks is usually reserved in the linker script.

5.1.1. Count Leading Zeroes Instruction¶

The AVR CPU has the XXX instruction which could be used to speed up the find first bit operation. The use of this instruction should significantly speed up the scheduling associated with a thread blocking.

5.2.1. Processor Background¶

The AVR architecture supports a simple call and return mechanism. A subroutine is invoked via the call (call) instruction. This instruction saves the return address in the RETS register and transfers the execution to the given address.

It is the called funcions responsability to use the link instruction to reserve space on the stack for the local variables. Returning from a subroutine is done by using the RTS (RTS) instruction which loads the PC with the adress stored in RETS.

It is is important to note that the call instruction does not automatically save or restore any registers. It is the responsibility of the high-level language compiler to define the register preservation and usage convention.

5.2.2. Register Usage¶

A called function may clobber all registers, except RETS, R4-R7, P3-P5, FP and SP. It may also modify the first 12 bytes in the caller’s stack frame which is used as an argument area for the first three arguments (which are passed in R0…R3 but may be placed on the stack by the called function).

5.2.3. Parameter Passing¶

RTEMS assumes that the AVR GCC calling convention is followed. The first three parameters are stored in registers R0, R1, and R2. All other parameters are put pushed on the stack. The result is returned through register R0.

5.4.1. Vectoring of an Interrupt Handler¶

TBD

5.4.2. Disabling of Interrupts by RTEMS¶

During interrupt disable critical sections, RTEMS disables interrupts to level N (N) before the execution of this section and restores them to the previous level upon completion of the section. RTEMS uses the instructions CLI and STI to enable and disable Interrupts. Emulation, Reset, NMI and Exception Interrupts are never disabled.

5.4.3. Interrupt Stack¶

The AVR Architecture works with two different kind of stacks, User and Supervisor Stack. Since RTEMS and its Application run in supervisor mode, all interrupts will use the interrupted tasks stack for execution.

5.7.1. System Reset¶

TBD

6.1.1. Count Leading Zeroes Instruction¶

The Blackfin CPU has the BITTST instruction which could be used to speed up the find first bit operation. The use of this instruction should significantly speed up the scheduling associated with a thread blocking.

6.2.1. Processor Background¶

The Blackfin architecture supports a simple call and return mechanism. A subroutine is invoked via the call (call) instruction. This instruction saves the return address in the RETS register and transfers the execution to the given address.

It is the called funcions responsability to use the link instruction to reserve space on the stack for the local variables. Returning from a subroutine is done by using the RTS (RTS) instruction which loads the PC with the adress stored in RETS.

It is is important to note that the call instruction does not automatically save or restore any registers. It is the responsibility of the high-level language compiler to define the register preservation and usage convention.

6.2.2. Register Usage¶

A called function may clobber all registers, except RETS, R4-R7, P3-P5, FP and SP. It may also modify the first 12 bytes in the caller’s stack frame which is used as an argument area for the first three arguments (which are passed in R0…R3 but may be placed on the stack by the called function).

6.2.3. Parameter Passing¶

RTEMS assumes that the Blackfin GCC calling convention is followed. The first three parameters are stored in registers R0, R1, and R2. All other parameters are put pushed on the stack. The result is returned through register R0.

6.4.1. Vectoring of an Interrupt Handler¶

RTEMS maps levels 0 -15 directly to Blackfins event vectors EVT0 - EVT15. Since EVT0 - EVT6 are core events and it is suggested to use EVT15 and EVT15 for Software interrupts, 7 Interrupts (EVT7-EVT13) are left for periferical use.

When installing an RTEMS interrupt handler RTEMS installs a generic Interrupt Handler which saves some context and enables nested interrupt servicing and then vectors to the users interrupt handler.

6.4.2. Disabling of Interrupts by RTEMS¶

During interrupt disable critical sections, RTEMS disables interrupts to level four (4) before the execution of this section and restores them to the previous level upon completion of the section. RTEMS uses the instructions CLI and STI to enable and disable Interrupts. Emulation, Reset, NMI and Exception Interrupts are never disabled.

6.4.3. Interrupt Stack¶

The Blackfin Architecture works with two different kind of stacks, User and Supervisor Stack. Since RTEMS and its Application run in supervisor mode, all interrupts will use the interrupted tasks stack for execution.

6.7.1. System Reset¶

TBD

8.1.1. bswap Instruction¶

The macro I386_HAS_BSWAP is set to 1 to indicate that this CPU model has the bswap instruction which endian swaps a thirty-two bit quantity. This instruction appears to be present in all CPU models i486’s and above.

8.2.1. Processor Background¶

The i386 architecture supports a simple yet effective call and return mechanism. A subroutine is invoked via the call (call) instruction. This instruction pushes the return address on the stack. The return from subroutine (ret) instruction pops the return address off the current stack and transfers control to that instruction. It is is important to note that the i386 call and return mechanism does not automatically save or restore any registers. It is the responsibility of the high-level language compiler to define the register preservation and usage convention.

8.2.2. Calling Mechanism¶

All RTEMS directives are invoked using a call instruction and return to the user application via the ret instruction.

8.2.3. Register Usage¶

As discussed above, the call instruction does not automatically save any registers. RTEMS uses the registers EAX, ECX, and EDX as scratch registers. These registers are not preserved by RTEMS directives therefore, the contents of these registers should not be assumed upon return from any RTEMS directive.

8.2.4. Parameter Passing¶

RTEMS assumes that arguments are placed on the current stack before the directive is invoked via the call instruction. The first argument is assumed to be closest to the return address on the stack. This means that the first argument of the C calling sequence is pushed last. The following pseudo-code illustrates the typical sequence used to call a RTEMS directive with three (3) arguments:

push third argument
push second argument
push first argument
invoke directive
remove arguments from the stack

The arguments to RTEMS are typically pushed onto the stack using a push instruction. These arguments must be removed from the stack after control is returned to the caller. This removal is typically accomplished by adding the size of the argument list in bytes to the stack pointer.

8.3.1. Flat Memory Model¶

RTEMS supports the i386 protected mode, flat memory model with paging disabled. In this mode, the i386 automatically converts every address from a logical to a physical address each time it is used. The i386 uses information provided in the segment registers and the Global Descriptor Table to convert these addresses. RTEMS assumes the existence of the following segments:

• a single code segment at protection level (0) which contains all application and executive code.

• a single data segment at protection level zero (0) which contains all application and executive data.

The i386 segment registers and associated selectors must be initialized when the initialize_executive directive is invoked. RTEMS treats the segment registers as system registers and does not modify or context switch them.

This i386 memory model supports a flat 32-bit address space with addresses ranging from 0x00000000 to 0xFFFFFFFF (4 gigabytes). Each address is represented by a 32-bit value and is byte addressable. The address may be used to reference a single byte, half-word (2-bytes), or word (4 bytes).

8.4.1. Vectoring of Interrupt Handler¶

Although the i386 supports multiple privilege levels, RTEMS and all user software executes at privilege level 0. This decision was made by the RTEMS designers to enhance compatibility with processors which do not provide sophisticated protection facilities like those of the i386. This decision greatly simplifies the discussion of i386 processing, as one need only consider interrupts without privilege transitions.

Upon receipt of an interrupt the i386 automatically performs the following actions:

• pushes the EFLAGS register

• pushes the far address of the interrupted instruction

• vectors to the interrupt service routine (ISR).

A nested interrupt is processed similarly by the i386.

8.4.2. Interrupt Stack Frame¶

The structure of the Interrupt Stack Frame for the i386 which is placed on the interrupt stack by the processor in response to an interrupt is as follows:

8.4.3. Interrupt Levels¶

Although RTEMS supports 256 interrupt levels, the i386 only supports two - enabled and disabled. Interrupts are enabled when the interrupt-enable flag (IF) in the extended flags (EFLAGS) is set. Conversely, interrupt processing is inhibited when the IF is cleared. During a non-maskable interrupt, all other interrupts, including other non-maskable ones, are inhibited.

RTEMS interrupt levels 0 and 1 such that level zero (0) indicates that interrupts are fully enabled and level one that interrupts are disabled. All other RTEMS interrupt levels are undefined and their behavior is unpredictable.

8.4.4. Interrupt Stack¶

The i386 family does not support a dedicated hardware interrupt stack. On this processor, RTEMS allocates and manages a dedicated interrupt stack. As part of vectoring a non-nested interrupt service routine, RTEMS switches from the stack of the interrupted task to a dedicated interrupt stack. When a non-nested interrupt returns, RTEMS switches back to the stack of the interrupted stack. The current stack pointer is not altered by RTEMS on nested interrupt.

8.7.1. System Reset¶

An RTEMS based application is initiated when the i386 processor is reset. When the i386 is reset,

• The EAX register is set to indicate the results of the processor’s power-up self test. If the self-test was not executed, the contents of this register are undefined. Otherwise, a non-zero value indicates the processor is faulty and a zero value indicates a successful self-test.

• The DX register holds a component identifier and revision level. DH contains 3 to indicate an i386 component and DL contains a unique revision level indicator.

• Control register zero (CR0) is set such that the processor is in real mode with paging disabled. Other portions of CR0 are used to indicate the presence of a numeric coprocessor.

• All bits in the extended flags register (EFLAG) which are not permanently set are cleared. This inhibits all maskable interrupts.

• The Interrupt Descriptor Register (IDTR) is set to point at address zero.

• All segment registers are set to zero.

• The instruction pointer is set to 0x0000FFF0. The first instruction executed after a reset is actually at 0xFFFFFFF0 because the i386 asserts the upper twelve address until the first intersegment (FAR) JMP or CALL instruction. When a JMP or CALL is executed, the upper twelve address lines are lowered and the processor begins executing in the first megabyte of memory.

Typically, an intersegment JMP to the application’s initialization code is placed at address 0xFFFFFFF0.

8.7.2. Processor Initialization¶

This initialization code is responsible for initializing all data structures required by the i386 in protected mode and for actually entering protected mode. The i386 must be placed in protected mode and the segment registers and associated selectors must be initialized before the initialize_executive directive is invoked.

The initialization code is responsible for initializing the Global Descriptor Table such that the i386 is in the thirty-two bit flat memory model with paging disabled. In this mode, the i386 automatically converts every address from a logical to a physical address each time it is used. For more information on the memory model used by RTEMS, please refer to the Memory Model chapter in this document.

Since the processor is in real mode upon reset, the processor must be switched to protected mode before RTEMS can execute. Before switching to protected mode, at least one descriptor table and two descriptors must be created. Descriptors are needed for a code segment and a data segment. ( This will give you the flat memory model.) The stack can be placed in a normal read/write data segment, so no descriptor for the stack is needed. Before the GDT can be used, the base address and limit must be loaded into the GDTR register using an LGDT instruction.

If the hardware allows an NMI to be generated, you need to create the IDT and a gate for the NMI interrupt handler. Before the IDT can be used, the base address and limit for the idt must be loaded into the IDTR register using an LIDT instruction.

Protected mode is entered by setting thye PE bit in the CR0 register. Either a LMSW or MOV CR0 instruction may be used to set this bit. Because the processor overlaps the interpretation of several instructions, it is necessary to discard the instructions from the read-ahead cache. A JMP instruction immediately after the LMSW changes the flow and empties the processor if intructions which have been pre-fetched and/or decoded. At this point, the processor is in protected mode and begins to perform protected mode application initialization.

If the application requires that the IDTR be some value besides zero, then it should set it to the required value at this point. All tasks share the same i386 IDTR value. Because interrupts are enabled automatically by RTEMS as part of the initialize_executive directive, the IDTR MUST be set properly before this directive is invoked to insure correct interrupt vectoring. If processor caching is to be utilized, then it should be enabled during the reset application initialization code. The reset code which is executed before the call to initialize_executive has the following requirements:

For more information regarding the i386 data structures and their contents, refer to Intel’s 386 Programmer’s Reference Manual.

9.3.1. Calling Mechanism¶

A call instruction places the return address to register r29 and a return from subroutine (ret) is actually a branch to r29/ra.

9.3.2. Register Usage¶

A subroutine may freely use registers r1 to r10 which are not preserved across subroutine invocations.

9.3.3. Parameter Passing¶

When calling a C function the first eight arguments are stored in registers r1 to r8. Registers r1 and r2 hold the return value.

9.8.1. System Reset¶

An RTEMS based application is initiated or re-initiated when the processor is reset.

11.1.1. BFFFO Instruction¶

The macro M68K_HAS_BFFFO is set to 1 to indicate that this CPU model has the bfffo instruction.

11.1.2. Vector Base Register¶

The macro M68K_HAS_VBR is set to 1 to indicate that this CPU model has a vector base register (vbr).

11.1.3. Separate Stacks¶

The macro M68K_HAS_SEPARATE_STACKS is set to 1 to indicate that this CPU model has separate interrupt, user, and supervisor mode stacks.

11.1.4. Pre-Indexing Address Mode¶

The macro M68K_HAS_PREINDEXING is set to 1 to indicate that this CPU model has the pre-indexing address mode.

11.1.5. Extend Byte to Long Instruction¶

The macro M68K_HAS_EXTB_L is set to 1 to indicate that this CPU model has the extb.l instruction. This instruction is supposed to be available in all models based on the cpu32 core as well as mc68020 and up models.

11.2.1. Calling Mechanism¶

All RTEMS directives are invoked using either a bsr or jsr instruction and return to the user application via the rts instruction.

11.2.2. Register Usage¶

As discussed above, the bsr and jsr instructions do not automatically save any registers. RTEMS uses the registers D0, D1, A0, and A1 as scratch registers. These registers are not preserved by RTEMS directives therefore, the contents of these registers should not be assumed upon return from any RTEMS directive.

11.2.3. Parameter Passing¶

RTEMS assumes that arguments are placed on the current stack before the directive is invoked via the bsr or jsr instruction. The first argument is assumed to be closest to the return address on the stack. This means that the first argument of the C calling sequence is pushed last. The following pseudo-code illustrates the typical sequence used to call a RTEMS directive with three (3) arguments:

push third argument
push second argument
push first argument
invoke directive
remove arguments from the stack

The arguments to RTEMS are typically pushed onto the stack using a move instruction with a pre-decremented stack pointer as the destination. These arguments must be removed from the stack after control is returned to the caller. This removal is typically accomplished by adding the size of the argument list in bytes to the current stack pointer.

11.4.1. Vectoring of an Interrupt Handler¶

Depending on whether or not the particular CPU supports a separate interrupt stack, the MC68xxx family has two different interrupt handling models.

11.4.2. CPU Models Without VBR and RAM at 0¶

This is from a post by Zoltan Kocsi <zoltan@bendor.com.au> and is a nice trick in certain situations. In his words:

I think somebody on this list asked about the interupt vector handling w/o VBR and RAM at 0. The usual trick is to initialise the vector table (except the first 2 two entries, of course) to point to the same location BUT you also add the vector number times 0x1000000 to them. That is, bits 31-24 contain the vector number and 23-0 the address of the common handler. Since the PC is 32 bit wide but the actual address bus is only 24, the top byte will be in the PC but will be ignored when jumping onto your routine.

Then your common interrupt routine gets this info by loading the PC into some register and based on that info, you can jump to a vector in a vector table pointed by a virtual VBR:

//
//  Real vector table at 0
//
.long   initial_sp
.long   initial_pc
.long   myhandler+0x02000000
.long   myhandler+0x03000000
.long   myhandler+0x04000000
...
.long   myhandler+0xff000000
//
// This handler will jump to the interrupt routine   of which
// the address is stored at VBR[ vector_no ]
// The registers and stackframe will be intact, the interrupt
// routine will see exactly what it would see if it was called
// directly from the HW vector table at 0.
//
    .comm    VBR,4,2        // This defines the 'virtual' VBR
// From C: extern void *VBR;
myhandler:                  // At entry, PC contains the full vector
    move.l  %d0,-(%sp)      // Save d0
    move.l  %a0,-(%sp)      // Save a0
    lea     0(%pc),%a0      // Get the value of the PC
    move.l  %a0,%d0         // Copy it to a data reg, d0 is VV??????
    swap    %d0             // Now d0 is ????VV??
    and.w   #0xff00,%d0     // Now d0 is ????VV00 (1)
    lsr.w   #6,%d0          // Now d0.w contains the VBR table offset
    move.l  VBR,%a0         // Get the address from VBR to a0
    move.l  (%a0,%d0.w),%a0 // Fetch the vector
    move.l  4(%sp),%d0      // Restore d0
    move.l  %a0,4(%sp)      // Place target address to the stack
    move.l  (%sp)+,%a0      // Restore a0, target address is on TOS
    ret                     // This will jump to the handler and
// restore the stack

1. If ‘myhandler’ is guaranteed to be in the first 64K, e.g. just after the vector table then that insn is not needed.

There are probably shorter ways to do this, but it I believe is enough to illustrate the trick. Optimisation is left as an exercise to the reader :-)

11.4.3. Interrupt Levels¶

Eight levels (0-7) of interrupt priorities are supported by MC68xxx family members with level seven (7) being the highest priority. Level zero (0) indicates that interrupts are fully enabled. Interrupt requests for interrupts with priorities less than or equal to the current interrupt mask level are ignored.

Although RTEMS supports 256 interrupt levels, the MC68xxx family only supports eight. RTEMS interrupt levels 0 through 7 directly correspond to MC68xxx interrupt levels. All other RTEMS interrupt levels are undefined and their behavior is unpredictable.

11.7.1. System Reset¶

An RTEMS based application is initiated or re-initiated when the MC68020 processor is reset. When the MC68020 is reset, the processor performs the following actions:

• The tracing bits of the status register are cleared to disable tracing.

• The supervisor interrupt state is entered by setting the supervisor (S) bit and clearing the master/interrupt (M) bit of the status register.

• The interrupt mask of the status register is set to level 7 to effectively disable all maskable interrupts.

• The vector base register (VBR) is set to zero.

• The cache control register (CACR) is set to zero to disable and freeze the processor cache.

• The interrupt stack pointer (ISP) is set to the value stored at vector 0 (bytes 0-3) of the exception vector table (EVT).

• The program counter (PC) is set to the value stored at vector 1 (bytes 4-7) of the EVT.

• The processor begins execution at the address stored in the PC.

11.7.2. Processor Initialization¶

The address of the application’s initialization code should be stored in the first vector of the EVT which will allow the immediate vectoring to the application code. If the application requires that the VBR be some value besides zero, then it should be set to the required value at this point. All tasks share the same MC68020’s VBR value. Because interrupts are enabled automatically by RTEMS as part of the context switch to the first task, the VBR MUST be set by either RTEMS of the BSP before this occurs ensure correct interrupt vectoring. If processor caching is to be utilized, then it should be enabled during the reset application initialization code.

In addition to the requirements described in the Board Support Packages chapter of the Applications User’s Manual for the reset code which is executed before the call to initialize executive, the MC68020 version has the following specific requirements:

• Must leave the S bit of the status register set so that the MC68020 remains in the supervisor state.

• Must set the M bit of the status register to remove the MC68020 from the interrupt state.

• Must set the master stack pointer (MSP) such that a minimum stack size of MINIMUM_STACK_SIZE bytes is provided for the initialize executive directive.

• Must initialize the MC68020’s vector table.

12.3.1. Interrupt Levels¶

There are exactly two interrupt levels on MicroBlaze with respect to RTEMS. Level zero corresponds to interrupts disabled. Level one corresponds to interrupts enabled. This is the inverse of how most other architectures handle interrupt enable status.

12.3.2. Interrupt Stack¶

The memory region for the interrupt stack is defined by the BSP.

13.1.1. Another Optional Feature¶

The macro XXX

13.2.1. Processor Background¶

TBD

13.2.2. Calling Mechanism¶

TBD

13.2.3. Register Usage¶

TBD

13.2.4. Parameter Passing¶

TBD

13.3.1. Flat Memory Model¶

The MIPS family supports a flat 32-bit address space with addresses ranging from 0x00000000 to 0xFFFFFFFF (4 gigabytes). Each address is represented by a 32-bit value and is byte addressable. The address may be used to reference a single byte, word (2-bytes), or long word (4 bytes). Memory accesses within this address space are performed in big endian fashion by the processors in this family.

Some of the MIPS family members such as the support virtual memory and segmentation. RTEMS does not support virtual memory or segmentation on any of these family members.

13.4.1. Vectoring of an Interrupt Handler¶

Upon receipt of an interrupt the XXX family members with separate interrupt stacks automatically perform the following actions:

• TBD

A nested interrupt is processed similarly by these CPU models with the exception that only a single ISF is placed on the interrupt stack and the current stack need not be switched.

13.4.2. Interrupt Levels¶

TBD

13.7.1. System Reset¶

An RTEMS based application is initiated or re-initiated when the processor is reset. When the processor is reset, it performs the following actions:

• TBD

13.7.2. Processor Initialization¶

TBD

15.1.1. Floating Point Unit¶

A floating point unit is currently not supported.

15.3.1. Interrupt Levels¶

There are only two levels: interrupts enabled and interrupts disabled.

15.3.2. Interrupt Stack¶

The OpenRISC RTEMS port uses a dedicated software interrupt stack. The stack for interrupts is allocated during interrupt driver initialization. When an interrupt is entered, the _ISR_Handler routine is responsible for switching from the interrupted task stack to RTEMS software interrupt stack.

16.5.1. Interrupt Levels¶

There are exactly two interrupt levels on PowerPC with respect to RTEMS. Level zero corresponds to interrupts enabled. Level one corresponds to interrupts disabled.

16.5.2. Interrupt Stack¶

The interrupt stack size can be configured via the CONFIGURE_INTERRUPT_STACK_SIZE application configuration option.

17.3.1. Interrupt Levels¶

There are exactly two interrupt levels on RISC-V with respect to RTEMS. Level zero corresponds to machine interrupts enabled. Level one corresponds to machine interrupts disabled.

17.3.2. Interrupt Stack¶

The memory region for the interrupt stack is defined by the BSP.

18.1.1. Another Optional Feature¶

The macro XXX

18.2.1. Calling Mechanism¶

All RTEMS directives are invoked using a XXX instruction and return to the user application via the XXX instruction.

18.2.2. Register Usage¶

The SH1 has 16 general registers (r0..r15).

• r0..r3 used as general volatile registers

• r4..r7 used to pass up to 4 arguments to functions, arguments above 4 are passed via the stack)

• r8..13 caller saved registers (i.e. push them to the stack if you need them inside of a function)

• r14 frame pointer

• r15 stack pointer

18.2.3. Parameter Passing¶

XXX

18.3.1. Flat Memory Model¶

The SuperH family supports a flat 32-bit address space with addresses ranging from 0x00000000 to 0xFFFFFFFF (4 gigabytes). Each address is represented by a 32-bit value and is byte addressable. The address may be used to reference a single byte, word (2-bytes), or long word (4 bytes). Memory accesses within this address space are performed in big endian fashion by the processors in this family.

Some of the SuperH family members support virtual memory and segmentation. RTEMS does not support virtual memory or segmentation on any of the SuperH family members. It is the responsibility of the BSP to initialize the mapping for a flat memory model.

18.4.1. Vectoring of an Interrupt Handler¶

Upon receipt of an interrupt the XXX family members with separate interrupt stacks automatically perform the following actions:

• TBD

A nested interrupt is processed similarly by these CPU models with the exception that only a single ISF is placed on the interrupt stack and the current stack need not be switched.

18.4.2. Interrupt Levels¶

TBD

18.7.1. System Reset¶

An RTEMS based application is initiated or re-initiated when the processor is reset. When the processor is reset, it performs the following actions:

• TBD

18.7.2. Processor Initialization¶

TBD

19.1.1. CPU Model Feature Flags¶

Each processor family supported by RTEMS has a list of features which vary between CPU models within a family. For example, the most common model dependent feature regardless of CPU family is the presence or absence of a floating point unit or coprocessor. When defining the list of features present on a particular CPU model, one simply notes that floating point hardware is or is not present and defines a single constant appropriately. Conditional compilation is utilized to include the appropriate source code for this CPU model’s feature set. It is important to note that this means that RTEMS is thus compiled using the appropriate feature set and compilation flags optimal for this CPU model used. The alternative would be to generate a binary which would execute on all family members using only the features which were always present.

This section presents the set of features which vary across SPARC implementations and are of importance to RTEMS. The set of CPU model feature macros are defined in the file cpukit/score/cpu/sparc/sparc.h based upon the particular CPU model defined on the compilation command line.

while ( TRUE ) {
    enter low power mode
}

19.1.2. CPU Model Implementation Notes¶

The ERC32 is a custom SPARC V7 implementation based on the Cypress 601/602 chipset. This CPU has a number of on-board peripherals and was developed by the European Space Agency to target space applications. RTEMS currently provides support for the following peripherals:

• UART Channels A and B

• General Purpose Timer

• Real Time Clock

• Watchdog Timer (so it can be disabled)

• Control Register (so powerdown mode can be enabled)

• Memory Control Register

• Interrupt Control

The General Purpose Timer and Real Time Clock Timer provided with the ERC32 share the Timer Control Register. Because the Timer Control Register is write only, we must mirror it in software and insure that writes to one timer do not alter the current settings and status of the other timer. Routines are provided in erc32.h which promote the view that the two timers are completely independent. By exclusively using these routines to access the Timer Control Register, the application can view the system as having a General Purpose Timer Control Register and a Real Time Clock Timer Control Register rather than the single shared value.

The RTEMS Idle thread take advantage of the low power mode provided by the ERC32. Low power mode is entered during idle loops and is enabled at initialization time.

19.2.1. Programming Model¶

This section discusses the programming model for the SPARC architecture.

19.2.2. Register Windows¶

The SPARC architecture includes the concept of register windows. An overly simplistic way to think of these windows is to imagine them as being an infinite supply of “fresh” register sets available for each subroutine to use. In reality, they are much more complicated.

The save instruction is used to obtain a new register window. This instruction decrements the current window pointer, thus providing a new set of registers for use. This register set includes eight fresh local registers for use exclusively by this subroutine. When done with a register set, the restore instruction increments the current window pointer and the previous register set is once again available.

The two primary issues complicating the use of register windows are that (1) the set of register windows is finite, and (2) some registers are shared between adjacent registers windows.

Because the set of register windows is finite, it is possible to execute enough save instructions without corresponding restore’s to consume all of the register windows. This is easily accomplished in a high level language because each subroutine typically performs a save instruction upon entry. Thus having a subroutine call depth greater than the number of register windows will result in a window overflow condition. The window overflow condition generates a trap which must be handled in software. The window overflow trap handler is responsible for saving the contents of the oldest register window on the program stack.

Similarly, the subroutines will eventually complete and begin to perform restore’s. If the restore results in the need for a register window which has previously been written to memory as part of an overflow, then a window underflow condition results. Just like the window overflow, the window underflow condition must be handled in software by a trap handler. The window underflow trap handler is responsible for reloading the contents of the register window requested by the restore instruction from the program stack.

The Window Invalid Mask (WIM) and the Current Window Pointer (CWP) field in the PSR are used in conjunction to manage the finite set of register windows and detect the window overflow and underflow conditions. The CWP contains the index of the register window currently in use. The save instruction decrements the CWP modulo the number of register windows. Similarly, the restore instruction increments the CWP modulo the number of register windows. Each bit in the WIM represents represents whether a register window contains valid information. The value of 0 indicates the register window is valid and 1 indicates it is invalid. When a save instruction causes the CWP to point to a register window which is marked as invalid, a window overflow condition results. Conversely, the restore instruction may result in a window underflow condition.

Other than the assumption that a register window is always available for trap (i.e. interrupt) handlers, the SPARC architecture places no limits on the number of register windows simultaneously marked as invalid (i.e. number of bits set in the WIM). However, RTEMS assumes that only one register window is marked invalid at a time (i.e. only one bit set in the WIM). This makes the maximum possible number of register windows available to the user while still meeting the requirement that window overflow and underflow conditions can be detected.

The window overflow and window underflow trap handlers are a critical part of the run-time environment for a SPARC application. The SPARC architectural specification allows for the number of register windows to be any power of two less than or equal to 32. The most common choice for SPARC implementations appears to be 8 register windows. This results in the CWP ranging in value from 0 to 7 on most implementations.

The second complicating factor is the sharing of registers between adjacent register windows. While each register window has its own set of local registers, the input and output registers are shared between adjacent windows. The output registers for register window N are the same as the input registers for register window ((N - 1) modulo RW) where RW is the number of register windows. An alternative way to think of this is to remember how parameters are passed to a subroutine on the SPARC. The caller loads values into what are its output registers. Then after the callee executes a save instruction, those parameters are available in its input registers. This is a very efficient way to pass parameters as no data is actually moved by the save or restore instructions.

19.2.3. Call and Return Mechanism¶

The SPARC architecture supports a simple yet effective call and return mechanism. A subroutine is invoked via the call (call) instruction. This instruction places the return address in the caller’s output register 7 (o7). After the callee executes a save instruction, this value is available in input register 7 (i7) until the corresponding restore instruction is executed.

The callee returns to the caller via a jmp to the return address. There is a delay slot following this instruction which is commonly used to execute a restore instruction - if a register window was allocated by this subroutine.

It is important to note that the SPARC subroutine call and return mechanism does not automatically save and restore any registers. This is accomplished via the save and restore instructions which manage the set of registers windows.

In case a floating-point unit is supported, then floating-point return values appear in the floating-point registers. Single-precision values occupy %f0; double-precision values occupy %f0 and %f1. Otherwise, these are scratch registers. Due to this the hardware and software floating-point ABIs are incompatible.

19.2.4. Calling Mechanism¶

All RTEMS directives are invoked using the regular SPARC calling convention via the call instruction.

19.2.5. Register Usage¶

As discussed above, the call instruction does not automatically save any registers. The save and restore instructions are used to allocate and deallocate register windows. When a register window is allocated, the new set of local registers are available for the exclusive use of the subroutine which allocated this register set.

19.2.6. Parameter Passing¶

RTEMS assumes that arguments are placed in the caller’s output registers with the first argument in output register 0 (o0), the second argument in output register 1 (o1), and so forth. Until the callee executes a save instruction, the parameters are still visible in the output registers. After the callee executes a save instruction, the parameters are visible in the corresponding input registers. The following pseudo-code illustrates the typical sequence used to call a RTEMS directive with three (3) arguments:

load third argument into o2
load second argument into o1
load first argument into o0
invoke directive

19.2.7. User-Provided Routines¶

All user-provided routines invoked by RTEMS, such as user extensions, device drivers, and MPCI routines, must also adhere to these calling conventions.

In SPARC, the default behaviour is to execute instructions after a branch. As with the behaviour of most RISC (Reduced Instruction Set Computer) machines, SPARC uses a branch delay slot. This is because completing an instruction every clock cycle introduces the problem that a branch may not be resolved until the instruction has passed through the pipeline. By inserting stalls, this is prevented. In each cycle, if a stall is inserted, it is considered one branch delay slot.

For example, a regular branch instruction might look like so:

cmp %o4, %g4    /* if %o4 is equals to %g4 */
be 200fd06      /* then branch */
mov [%g4], %o4  /* instructions after the branch, this is a */
                /* branch delay slot it is executed regardless */
                /* of whether %o4 is equals to %g4 */

However, if marked with “,a”, the instructions after the branch will only be executed if the branch is taken. In other words, only if the condition before is true, then it would be executed. Otherwise if would be “annulled”.

cmp %o4, %g4    /* if %o4 is equals to %g4 */
be,a 200fd06    /* then branch */
mov [%g4], %o4  /* instruction after the branch */

The mov instruction is in a branch delay slot and is only executed if the branch is taken (e.g. if %o4 is equals to %g4).

This shows up in analysis of coverage reports when a single instruction is marked unexecuted when the instruction above and below it are executed.

19.3.1. Flat Memory Model¶

The SPARC architecture supports a flat 32-bit address space with addresses ranging from 0x00000000 to 0xFFFFFFFF (4 gigabytes). Each address is represented by a 32-bit value and is byte addressable. The address may be used to reference a single byte, half-word (2-bytes), word (4 bytes), or doubleword (8 bytes). Memory accesses within this address space are performed in big endian fashion by the SPARC. Memory accesses which are not properly aligned generate a “memory address not aligned” trap (type number 7). The following table lists the alignment requirements for a variety of data accesses:

Doubleword load and store operations must use a pair of registers as their source or destination. This pair of registers must be an adjacent pair of registers with the first of the pair being even numbered. For example, a valid destination for a doubleword load might be input registers 0 and 1 (i0 and i1). The pair i1 and i2 would be invalid. [NOTE: Some assemblers for the SPARC do not generate an error if an odd numbered register is specified as the beginning register of the pair. In this case, the assembler assumes that what the programmer meant was to use the even-odd pair which ends at the specified register. This may or may not have been a correct assumption.]

RTEMS does not support any SPARC Memory Management Units, therefore, virtual memory or segmentation systems involving the SPARC are not supported.

19.4.1. Synchronous Versus Asynchronous Traps¶

The SPARC architecture includes two classes of traps: synchronous and asynchronous. Asynchronous traps occur when an external event interrupts the processor. These traps are not associated with any instruction executed by the processor and logically occur between instructions. The instruction currently in the execute stage of the processor is allowed to complete although subsequent instructions are annulled. The return address reported by the processor for asynchronous traps is the pair of instructions following the current instruction.

Synchronous traps are caused by the actions of an instruction. The trap stimulus in this case either occurs internally to the processor or is from an external signal that was provoked by the instruction. These traps are taken immediately and the instruction that caused the trap is aborted before any state changes occur in the processor itself. The return address reported by the processor for synchronous traps is the instruction which caused the trap and the following instruction.

19.4.2. Trap Table¶

A SPARC processor uses a trap table to execute the trap handler associated with a trap. The trap table location is defined by the Trap Base Register (TBR). The trap table has 256 entries. Each entry has space for four instructions (16 bytes). RTEMS uses a statically initialized trap table. The start address of the trap table is associated with the trap_table global symbol. The first action of the system initialization (entry points _start and hard_reset) is to set the TBR to trap_table. The interrupt traps (trap numbers 16 to 31) are connected with the RTEMS interrupt handling. Some traps are connected to standard services defined by the SPARC architecture, for example the window overflow, underflow, and flush handling. Most traps are connected to a fatal error handler. The fatal error trap handler saves the processor context to an exception frame and starts the system termination procedure.

19.4.3. Vectoring of Interrupt Handler¶

Upon receipt of an interrupt a SPARC processor automatically performs the following actions:

• disables traps (sets the PSR.ET bit to 0 in the PSR),

• the PSR.S bit is copied into the Previous Supervisor Mode (PSR.PS) bit in the PSR,

• the CWP is decremented by one (modulo the number of register windows) to activate a trap window,

• the PC and nPC are loaded into local register 1 and 2 (%l0 and %l1),

• the trap type (tt) field of the Trap Base Register (TBR) is set to the appropriate value, and

• if the trap is not a reset, then the PC is written with the contents of the TBR and the nPC is written with TBR + 4. If the trap is a reset, then the PC is set to zero and the nPC is set to 4.

Trap processing on the SPARC has two features which are noticeably different than interrupt processing on other architectures. First, the value of PSR register in effect immediately before the trap occurred is not explicitly saved. Instead only reversible alterations are made to it. Second, the Processor Interrupt Level (PSR.PIL) is not set to correspond to that of the interrupt being processed. When a trap occurs, all subsequent traps are disabled. In order to safely invoke a subroutine during trap handling, traps must be enabled to allow for the possibility of register window overflow and underflow traps.

If the interrupt handler was installed as an RTEMS interrupt handler, then upon receipt of the interrupt, the processor passes control to the RTEMS interrupt handler which performs the following actions:

• saves the state of the interrupted task on it’s stack,

• insures that a register window is available for subsequent traps,

• if this is the outermost (i.e. non-nested) interrupt, then the RTEMS interrupt handler switches from the current stack to the interrupt stack,

• enables traps,

• invokes the vectors to a user interrupt service routine (ISR).

Asynchronous interrupts are ignored while traps are disabled. Synchronous traps which occur while traps are disabled result in the CPU being forced into an error mode.

A nested interrupt is processed similarly with the exception that the current stack need not be switched to the interrupt stack.

19.4.4. Traps and Register Windows¶

One of the register windows must be reserved at all times for trap processing. This is critical to the proper operation of the trap mechanism in the SPARC architecture. It is the responsibility of the trap handler to insure that there is a register window available for a subsequent trap before re-enabling traps. It is likely that any high level language routines invoked by the trap handler (such as a user-provided RTEMS interrupt handler) will allocate a new register window. The save operation could result in a window overflow trap. This trap cannot be correctly processed unless (1) traps are enabled and (2) a register window is reserved for traps. Thus, the RTEMS interrupt handler insures that a register window is available for subsequent traps before enabling traps and invoking the user’s interrupt handler.

19.4.5. Interrupt Levels¶

Sixteen levels (0-15) of interrupt priorities are supported by the SPARC architecture with level fifteen (15) being the highest priority. Level zero (0) indicates that interrupts are fully enabled. Interrupt requests for interrupts with priorities less than or equal to the current interrupt mask level are ignored. Level fifteen (15) is a non-maskable interrupt (NMI), which makes it unsuitable for standard usage since it can affect the real-time behaviour by interrupting critical sections and spinlocks. Disabling traps stops also the NMI interrupt from happening. It can however be used for power-down or other critical events.

Although RTEMS supports 256 interrupt levels, the SPARC only supports sixteen. RTEMS interrupt levels 0 through 15 directly correspond to SPARC processor interrupt levels. All other RTEMS interrupt levels are undefined and their behavior is unpredictable.

Many LEON SPARC v7/v8 systems features an extended interrupt controller which adds an extra step of interrupt decoding to allow handling of interrupt 16-31. When such an extended interrupt is generated the CPU traps into a specific interrupt trap level 1-14 and software reads out from the interrupt controller which extended interrupt source actually caused the interrupt.

19.4.6. Disabling of Interrupts by RTEMS¶

During the execution of directive calls, critical sections of code may be executed. When these sections are encountered, RTEMS disables interrupts to level fifteen (15) before the execution of the section and restores them to the previous level upon completion of the section. RTEMS has been optimized to ensure that interrupts are disabled for less than RTEMS_MAXIMUM_DISABLE_PERIOD microseconds on a RTEMS_MAXIMUM_DISABLE_PERIOD_MHZ Mhz ERC32 with zero wait states. These numbers will vary based the number of wait states and processor speed present on the target board. [NOTE: The maximum period with interrupts disabled is hand calculated. This calculation was last performed for Release RTEMS_RELEASE_FOR_MAXIMUM_DISABLE_PERIOD.]

[NOTE: It is thought that the length of time at which the processor interrupt level is elevated to fifteen by RTEMS is not anywhere near as long as the length of time ALL traps are disabled as part of the “flush all register windows” operation.]

Non-maskable interrupts (NMI) cannot be disabled, and ISRs which execute at this level MUST NEVER issue RTEMS system calls. If a directive is invoked, unpredictable results may occur due to the inability of RTEMS to protect its critical sections. However, ISRs that make no system calls may safely execute as non-maskable interrupts.

Interrupts are disabled or enabled by performing a system call to the Operating System reserved software traps 9 (SPARC_SWTRAP_IRQDIS) or 10 (SPARC_SWTRAP_IRQEN). The trap is generated by the software trap (Ticc) instruction or indirectly by calling sparc_disable_interrupts() or sparc_enable_interrupts() functions. Disabling interrupts return the previous interrupt level (on trap entry) in register G1 and sets PSR.PIL to 15 to disable all maskable interrupts. The interrupt level can be restored by trapping into the enable interrupt handler with G1 containing the new interrupt level.

19.4.7. Interrupt Stack¶

The SPARC architecture does not provide for a dedicated interrupt stack. Thus by default, trap handlers would execute on the stack of the RTEMS task which they interrupted. This artificially inflates the stack requirements for each task since every task stack would have to include enough space to account for the worst case interrupt stack requirements in addition to it’s own worst case usage. RTEMS addresses this problem on the SPARC by providing a dedicated interrupt stack managed by software.

The interrupt stack is statically allocated by RTEMS. There is one interrupt stack for each configured processor. The interrupt stack is used to initialize the system. The amount of memory allocated for the interrupt stack is determined by the CONFIGURE_INTERRUPT_STACK_SIZE application configuration option. As part of processing a non-nested interrupt, RTEMS will switch to the interrupt stack before invoking the installed handler.

19.7.1. System Reset¶

An RTEMS based application is initiated or re-initiated when the SPARC processor is reset. When the SPARC is reset, the processor performs the following actions:

• the enable trap (ET) of the PSR is set to 0 to disable traps,

• the supervisor bit (S) of the PSR is set to 1 to enter supervisor mode, and

• the PC is set 0 and the nPC is set to 4.

The processor then begins to execute the code at location 0. It is important to note that all fields in the PSR are not explicitly set by the above steps and all other registers retain their value from the previous execution mode. This is true even of the Trap Base Register (TBR) whose contents reflect the last trap which occurred before the reset.

19.7.2. Processor Initialization¶

It is the responsibility of the application’s initialization code to initialize the TBR and install trap handlers for at least the register window overflow and register window underflow conditions. Traps should be enabled before invoking any subroutines to allow for register window management. However, interrupts should be disabled by setting the Processor Interrupt Level (pil) field of the PSR to 15. RTEMS installs it’s own Trap Table as part of initialization which is initialized with the contents of the Trap Table in place when the rtems_initialize_executive directive was invoked. Upon completion of executive initialization, interrupts are enabled.

If this SPARC implementation supports on-chip caching and this is to be utilized, then it should be enabled during the reset application initialization code.

In addition to the requirements described in the Board Support Packages chapter of the C Applications Users Manual for the reset code which is executed before the call to``rtems_initialize_executive``, the SPARC version has the following specific requirements:

• Must leave the S bit of the status register set so that the SPARC remains in the supervisor state.

• Must set stack pointer (sp) such that a minimum stack size of MINIMUM_STACK_SIZE bytes is provided for the``rtems_initialize_executive`` directive.

• Must disable all external interrupts (i.e. set the pil to 15).

• Must enable traps so window overflow and underflow conditions can be properly handled.

• Must initialize the SPARC’s initial trap table with at least trap handlers for register window overflow and register window underflow.

19.8.1. General Structure¶

SPARC has 32 general purpose integer registers visible to the program at any given time. Of these, 8 registers are global registers and 24 registers are in a register window. A window consists of three groups of 8 registers, the out, local, and in registers. See table 1. A SPARC implementation can have from 2 to 32 windows, thus varying the number of registers from 40 to 520. Most implementations have 7 or 8 windows. The variable number of registers is the principal reason for the SPARC being “scalable”.

At any given time, only one window is visible, as determined by the current window pointer (CWP) which is part of the processor status register (PSR). This is a five bit value that can be decremented or incremented by the save and restore instructions, respectively. These instructions are generally executed on procedure call and return (respectively). The idea is that the in registers contain incoming parameters, the local register constitutes scratch registers, the out registers contain outgoing parameters, and the global registers contain values that vary little between executions. The register windows overlap partially, thus the out registers become renamed by save to become the in registers of the called procedure. Thus, the memory traffic is reduced when going up and down the procedure call. Since this is a frequent operation, performance is improved.

(That was the idea, anyway. The drawback is that upon interactions with the system the registers need to be flushed to the stack, necessitating a long sequence of writes to memory of data that is often mostly garbage. Register windows was a bad idea that was caused by simulation studies that considered only programs in isolation, as opposed to multitasking workloads, and by considering compilers with poor optimization. It also caused considerable problems in implementing high-end SPARC processors such as the SuperSPARC, although more recent implementations have dealt effectively with the obstacles. Register windows are now part of the compatibility legacy and not easily removed from the architecture.)

The overlap of the registers is illustrated in figure 1. The figure shows an implementation with 8 windows, numbered 0 to 7 (labeled w0 to w7 in the figure). Each window corresponds to 24 registers, 16 of which are shared with “neighboring” windows. The windows are arranged in a wrap-around manner, thus window number 0 borders window number 7. The common cause of changing the current window, as pointed to by CWP, is the restore and save instructions, shown in the middle. Less common is the supervisor rett instruction (return from trap) and the trap event (interrupt, exception, or trap instruction).

The “WIM” register is also indicated in the top left of Figure 1. The window invalid mask is a bit map of valid windows. It is generally used as a pointer, i.e. exactly one bit is set in the WIM register indicating which window is invalid (in the figure it’s window 7). Register windows are generally used to support procedure calls, so they can be viewed as a cache of the stack contents. The WIM “pointer” indicates how many procedure calls in a row can be taken without writing out data to memory. In the figure, the capacity of the register windows is fully utilized. An additional call will thus exceed capacity, triggering a window overflow trap. At the other end, a window underflow trap occurs when the register window “cache” if empty and more data needs to be fetched from memory.

19.8.2. Register Semantics¶

The SPARC Architecture includes recommended software semantics. These are described in the architecture manual, the SPARC ABI (application binary interface) standard, and, unfortunately, in various other locations as well (including header files and compiler documentation).

Figure 2 shows a summary of register contents at any given time.

            %g0  (r00)       always zero
            %g1  (r01)  [1]  temporary value
            %g2  (r02)  [2]  global 2
global      %g3  (r03)  [2]  global 3
            %g4  (r04)  [2]  global 4
            %g5  (r05)       reserved for SPARC ABI
            %g6  (r06)       reserved for SPARC ABI
            %g7  (r07)       reserved for SPARC ABI

            %o0  (r08)  [3]  outgoing parameter 0 / return value from callee
            %o1  (r09)  [1]  outgoing parameter 1
            %o2  (r10)  [1]  outgoing parameter 2
out         %o3  (r11)  [1]  outgoing parameter 3
            %o4  (r12)  [1]  outgoing parameter 4
            %o5  (r13)  [1]  outgoing parameter 5
       %sp, %o6  (r14)  [1]  stack pointer
            %o7  (r15)  [1]  temporary value / address of CALL instruction

            %l0  (r16)  [3]  local 0
            %l1  (r17)  [3]  local 1
            %l2  (r18)  [3]  local 2
local       %l3  (r19)  [3]  local 3
            %l4  (r20)  [3]  local 4
            %l5  (r21)  [3]  local 5
            %l6  (r22)  [3]  local 6
            %l7  (r23)  [3]  local 7

            %i0  (r24)  [3]  incoming parameter 0 / return value to caller
            %i1  (r25)  [3]  incoming parameter 1
            %i2  (r26)  [3]  incoming parameter 2
in          %i3  (r27)  [3]  incoming parameter 3
            %i4  (r28)  [3]  incoming parameter 4
            %i5  (r29)  [3]  incoming parameter 5
       %fp, %i6  (r30)  [3]  frame pointer
            %i7  (r31)  [3]  return address - 8

Figure 2 - SPARC register semantics

Particular compilers are likely to vary slightly.

Note that globals %g2-%g4 are reserved for the “application”, which includes libraries and compiler. Thus, for example, libraries may overwrite these registers unless they’ve been compiled with suitable flags. Also, the “reserved” registers are presumed to be allocated (in the future) bottom-up, i.e. %g7 is currently the “safest” to use.

Optimizing linkers and interpreters are examples that use global registers.

19.8.3. Register Windows and the Stack¶

The SPARC register windows are, naturally, intimately related to the stack. In particular, the stack pointer (%sp or %o6) must always point to a free block of 64 bytes. This area is used by the operating system (Solaris, SunOS, and Linux at least) to save the current local and in registers upon a system interrupt, exception, or trap instruction. (Note that this can occur at any time.)

Other aspects of register relations with memory are programming convention. The typical and recommended layout of the stack is shown in figure 3. The figure shows a stack frame.

Note that the top boxes of figure 3 are addressed via the stack pointer (%sp), as positive offsets (including zero), and the bottom boxes are accessed over the frame pointer using negative offsets (excluding zero), and that the frame pointer is the old stack pointer. This scheme allows the separation of information known at compile time (number and size of local parameters, etc) from run-time information (size of blocks allocated by alloca()).

“addressable scalar automatics” is a fancy name for local variables.

The clever nature of the stack and frame pointers is that they are always 16 registers apart in the register windows. Thus, a save instruction will make the current stack pointer into the frame pointer and, since the save instruction also doubles as an add, create a new stack pointer. Figure 4 illustrates what the top of a stack might look like during execution. (The listing is from the pwin command in the SimICS simulator.)

Note how the stack contents are not necessarily synchronized with the registers. Various events can cause the register windows to be “flushed” to memory, including most system calls. A programmer can force this update by using ST_FLUSH_WINDOWS trap, which also reduces the number of valid windows to the minimum of 1.

Writing a library for multithreaded execution is an example that requires explicit flushing, as is longjmp().

19.8.4. Procedure epilogue and prologue¶

The stack frame described in the previous section leads to the standard entry/exit mechanisms listed in figure 5.

function:
  save  %sp, -C, %sp

             ; perform function, leave return value,
             ; if any, in register %i0 upon exit

  ret        ; jmpl %i7+8, %g0
  restore    ; restore %g0,%g0,%g0

Figure 5 - Epilogue/prologue in procedures

The save instruction decrements the CWP, as discussed earlier, and also performs an addition. The constant C that is used in the figure to indicate the amount of space to make on the stack, and thus corresponds to the frame contents in Figure 3. The minimum is therefore the 16 words for the local and in registers, i.e. (hex) 0x40 bytes.

A confusing element of the save instruction is that the source operands (the first two parameters) are read from the old register window, and the destination operand (the rightmost parameter) is written to the new window. Thus, although %sp is indicated as both source and destination, the result is actually written into the stack pointer of the new window (the source stack pointer becomes renamed and is now the frame pointer).

The return instructions are also a bit particular. ret is a synthetic instruction, corresponding to jmpl (jump linked). This instruction jumps to the address resulting from adding 8 to the %i7 register. The source instruction address (the address of the ret instruction itself) is written to the %g0 register, i.e. it is discarded.

The restore instruction is similarly a synthetic instruction and is just a short form for a restore that chooses not to perform an addition.

The calling instruction, in turn, typically looks as follows:

call <function>    ; jmpl <address>, %o7
mov 0, %o0

Again, the call instruction is synthetic, and is actually the same instruction that performs the return. This time, however, it is interested in saving the return address, into register %o7. Note that the delay slot is often filled with an instruction related to the parameters, in this example it sets the first parameter to zero.

Note also that the return value is also generally passed in %o0.

Leaf procedures are different. A leaf procedure is an optimization that reduces unnecessary work by taking advantage of the knowledge that no call instructions exist in many procedures. Thus, the save/restore couple can be eliminated. The downside is that such a procedure may only use the out registers (since the in and local registers actually belong to the caller). See Figure 6.

function:
             ; no save instruction needed upon entry

             ; perform function, leave return value,
             ; if any, in register %o0 upon exit

  retl       ; jmpl %o7+8, %g0
  nop        ; the delay slot can be used for something else

Figure 6 - Epilogue/prologue in leaf procedures

Note in the figure that there is only one instruction overhead, namely the retl instruction. retl is also synthetic (return from leaf subroutine), is again a variant of the jmpl instruction, this time with %o7+8 as target.

Yet another variation of epilogue is caused by tail call elimination, an optimization supported by some compilers (including Sun’s C compiler but not GCC). If the compiler detects that a called function will return to the calling function, it can replace its place on the stack with the called function. Figure 7 contains an example.

    int
      foo(int n)
    {
      if (n == 0)
        return 0;
      else
        return bar(n);
    }

      cmp     %o0,0
      bne     .L1
      or      %g0,%o7,%g1
      retl
      or      %g0,0,%o0
.L1:  call    bar
      or      %g0,%g1,%o7

Figure 7 - Example of tail call elimination

Note that the call instruction overwrites register %o7 with the program counter. Therefore the above code saves the old value of %o7, and restores it in the delay slot of the call instruction. If the function call is register indirect, this twiddling with %o7 can be avoided, but of course that form of call is slower on modern processors.

The benefit of tail call elimination is to remove an indirection upon return. It is also needed to reduce register window usage, since otherwise the foo() function in Figure 7 would need to allocate a stack frame to save the program counter.

A special form of tail call elimination is tail recursion elimination, which detects functions calling themselves, and replaces it with a simple branch. Figure 8 contains an example.

      int
        foo(int n)
      {
        if (n == 0)
          return 1;
        else
          return (foo(n - 1));
      }

      cmp     %o0,0
      be      .L1
      or      %g0,%o0,%g1
      subcc   %g1,1,%g1
.L2:  bne     .L2
      subcc   %g1,1,%g1
.L1:  retl
      or      %g0,1,%o0

Figure 8 - Example of tail recursion elimination

Needless to say, these optimizations produce code that is difficult to debug.

19.8.5. Procedures, stacks, and debuggers¶

When debugging an application, your debugger will be parsing the binary and consulting the symbol table to determine procedure entry points. It will also travel the stack frames “upward” to determine the current call chain.

When compiling for debugging, compilers will generate additional code as well as avoid some optimizations in order to allow reconstructing situations during execution. For example, GCC/GDB makes sure original parameter values are kept intact somewhere for future parsing of the procedure call stack. The live in registers other than %i0 are not touched. %i0 itself is copied into a free local register, and its location is noted in the symbol file. (You can find out where variables reside by using the info address command in GDB.)

Given that much of the semantics relating to stack handling and procedure call entry/exit code is only recommended, debuggers will sometimes be fooled. For example, the decision as to whether or not the current procedure is a leaf one or not can be incorrect. In this case a spurious procedure will be inserted between the current procedure and it’s “real” parent. Another example is when the application maintains its own implicit call hierarchy, such as jumping to function pointers. In this case the debugger can easily become totally confused.

19.8.6. The window overflow and underflow traps¶

When the save instruction decrements the current window pointer (CWP) so that it coincides with the invalid window in the window invalid mask (WIM), a window overflow trap occurs. Conversely, when the restore or rett instructions increment the CWP to coincide with the invalid window, a window underflow trap occurs.

Either trap is handled by the operating system. Generally, data is written out to memory and/or read from memory, and the WIM register suitably altered.

The code in Figure 9 and Figure 10 below are bare-bones handlers for the two traps. The text is directly from the source code, and sort of works. (As far as I know, these are minimalistic handlers for SPARC V8). Note that there is no way to directly access window registers other than the current one, hence the code does additional save/restore instructions. It’s pretty tricky to understand the code, but figure 1 should be of help.

      /* a SAVE instruction caused a trap */
window_overflow:
      /* rotate WIM on bit right, we have 8 windows */
      mov %wim,%l3
      sll %l3,7,%l4
      srl %l3,1,%l3
      or  %l3,%l4,%l3
      and %l3,0xff,%l3

      /* disable WIM traps */
      mov %g0,%wim
      nop; nop; nop

      /* point to correct window */
      save

      /* dump registers to stack */
      std %l0, [%sp +  0]
      std %l2, [%sp +  8]
      std %l4, [%sp + 16]
      std %l6, [%sp + 24]
      std %i0, [%sp + 32]
      std %i2, [%sp + 40]
      std %i4, [%sp + 48]
      std %i6, [%sp + 56]

      /* back to where we should be */
      restore

      /* set new value of window */
      mov %l3,%wim
      nop; nop; nop

      /* go home */
      jmp %l1
      rett %l2

Figure 9 - window_underflow trap handler

      /* a RESTORE instruction caused a trap */
window_underflow:

      /* rotate WIM on bit LEFT, we have 8 windows */
      mov %wim,%l3
      srl %l3,7,%l4
      sll %l3,1,%l3
      or  %l3,%l4,%l3
      and %l3,0xff,%l3

      /* disable WIM traps */
      mov %g0,%wim
      nop; nop; nop

      /* point to correct window */
      restore
      restore

      /* dump registers to stack */
      ldd [%sp +  0], %l0
      ldd [%sp +  8], %l2
      ldd [%sp + 16], %l4
      ldd [%sp + 24], %l6
      ldd [%sp + 32], %i0
      ldd [%sp + 40], %i2
      ldd [%sp + 48], %i4
      ldd [%sp + 56], %i6

      /* back to where we should be */
      save
      save

      /* set new value of window */
      mov %l3,%wim
      nop; nop; nop

      /* go home */
      jmp %l1
      rett %l2

Figure 10 - window_underflow trap handler

20.1.1. CPU Model Feature Flags¶

This section presents the set of features which vary across SPARC-64 implementations and are of importance to RTEMS. The set of CPU model feature macros are defined in the file cpukit/score/cpu/sparc64/sparc64.h based upon the particular CPU model defined on the compilation command line.

20.1.2. CPU Model Implementation Notes¶

This section describes the implemenation dependencies of the CPU Models sun4u and sun4v of the SPARC V9 architecture.

20.1.3. sun4v Notes¶

XXX

20.2.1. Programming Model¶

This section discusses the programming model for the SPARC architecture.

20.2.2. Register Windows¶

The SPARC architecture includes the concept of register windows. An overly simplistic way to think of these windows is to imagine them as being an infinite supply of “fresh” register sets available for each subroutine to use. In reality, they are much more complicated.

The save instruction is used to obtain a new register window. This instruction increments the current window pointer, thus providing a new set of registers for use. This register set includes eight fresh local registers for use exclusively by this subroutine. When done with a register set, the restore instruction decrements the current window pointer and the previous register set is once again available.

The two primary issues complicating the use of register windows are that (1) the set of register windows is finite, and (2) some registers are shared between adjacent registers windows.

Because the set of register windows is finite, it is possible to execute enough save instructions without corresponding restore’s to consume all of the register windows. This is easily accomplished in a high level language because each subroutine typically performs a save instruction upon entry. Thus having a subroutine call depth greater than the number of register windows will result in a window overflow condition. The window overflow condition generates a trap which must be handled in software. The window overflow trap handler is responsible for saving the contents of the oldest register window on the program stack.

Similarly, the subroutines will eventually complete and begin to perform restore’s. If the restore results in the need for a register window which has previously been written to memory as part of an overflow, then a window underflow condition results. Just like the window overflow, the window underflow condition must be handled in software by a trap handler. The window underflow trap handler is responsible for reloading the contents of the register window requested by the restore instruction from the program stack.

The cansave, canrestore, otherwin, and cwp are used in conjunction to manage the finite set of register windows and detect the window overflow and underflow conditions. The first three of these registers must satisfy the invariant cansave + canrestore + otherwin = nwindow - 2, where nwindow is the number of register windows. The cwp contains the index of the register window currently in use. RTEMS does not use the cleanwin and otherwin registers.

The save instruction increments the cwp modulo the number of register windows, and if cansave is 0 then it also generates a window overflow. Similarly, the restore instruction decrements the cwp modulo the number of register windows, and if canrestore is 0 then it also generates a window underflow.

Unlike with the SPARC model, the SPARC-64 port does not assume that a register window is available for a trap. The window overflow and underflow conditions are not detected without hardware generating the trap. (These conditions can be detected by reading the register window registers and doing some simple arithmetic.)

The window overflow and window underflow trap handlers are a critical part of the run-time environment for a SPARC application. The SPARC architectural specification allows for the number of register windows to be any power of two less than or equal to 32. The most common choice for SPARC implementations appears to be 8 register windows. This results in the cwp ranging in value from 0 to 7 on most implementations.

The second complicating factor is the sharing of registers between adjacent register windows. While each register window has its own set of local registers, the input and output registers are shared between adjacent windows. The output registers for register window N are the same as the input registers for register window ((N + 1) modulo RW) where RW is the number of register windows. An alternative way to think of this is to remember how parameters are passed to a subroutine on the SPARC. The caller loads values into what are its output registers. Then after the callee executes a save instruction, those parameters are available in its input registers. This is a very efficient way to pass parameters as no data is actually moved by the save or restore instructions.

20.2.3. Call and Return Mechanism¶

The SPARC architecture supports a simple yet effective call and return mechanism. A subroutine is invoked via the call (call) instruction. This instruction places the return address in the caller’s output register 7 (o7). After the callee executes a save instruction, this value is available in input register 7 (i7) until the corresponding restore instruction is executed.

The callee returns to the caller via a jmp to the return address. There is a delay slot following this instruction which is commonly used to execute a restore instruction - if a register window was allocated by this subroutine.

It is important to note that the SPARC subroutine call and return mechanism does not automatically save and restore any registers. This is accomplished via the save and restore instructions which manage the set of registers windows. This allows for the compiler to generate leaf-optimized functions that utilize the caller’s output registers without using save and restore.

20.2.4. Calling Mechanism¶

All RTEMS directives are invoked using the regular SPARC calling convention via the call instruction.

20.2.5. Register Usage¶

As discussed above, the call instruction does not automatically save any registers. The save and restore instructions are used to allocate and deallocate register windows. When a register window is allocated, the new set of local registers are available for the exclusive use of the subroutine which allocated this register set.

20.2.6. Parameter Passing¶

RTEMS assumes that arguments are placed in the caller’s output registers with the first argument in output register 0 (o0), the second argument in output register 1 (o1), and so forth. Until the callee executes a save instruction, the parameters are still visible in the output registers. After the callee executes a save instruction, the parameters are visible in the corresponding input registers. The following pseudo-code illustrates the typical sequence used to call a RTEMS directive with three (3) arguments:

load third argument into o2
load second argument into o1
load first argument into o0
invoke directive

20.2.7. User-Provided Routines¶

All user-provided routines invoked by RTEMS, such as user extensions, device drivers, and MPCI routines, must also adhere to these calling conventions.

20.3.1. Flat Memory Model¶

The SPARC-64 architecture supports a flat 64-bit address space with addresses ranging from 0x0000000000000000 to 0xFFFFFFFFFFFFFFFF. Each address is represented by a 64-bit value (and an 8-bit address space identifider or ASI) and is byte addressable. The address may be used to reference a single byte, half-word (2-bytes), word (4 bytes), doubleword (8 bytes), or quad-word (16 bytes). Memory accesses within this address space are performed in big endian fashion by the SPARC. Memory accesses which are not properly aligned generate a “memory address not aligned” trap (type number 0x34). The following table lists the alignment requirements for a variety of data accesses:

RTEMS currently does not support any SPARC Memory Management Units, therefore, virtual memory or segmentation systems involving the SPARC are not supported.

20.4.1. Synchronous Versus Asynchronous Traps¶

The SPARC architecture includes two classes of traps: synchronous (precise) and asynchronous (deferred). Asynchronous traps occur when an external event interrupts the processor. These traps are not associated with any instruction executed by the processor and logically occur between instructions. The instruction currently in the execute stage of the processor is allowed to complete although subsequent instructions are annulled. The return address reported by the processor for asynchronous traps is the pair of instructions following the current instruction.

Synchronous traps are caused by the actions of an instruction. The trap stimulus in this case either occurs internally to the processor or is from an external signal that was provoked by the instruction. These traps are taken immediately and the instruction that caused the trap is aborted before any state changes occur in the processor itself. The return address reported by the processor for synchronous traps is the instruction which caused the trap and the following instruction.

20.4.2. Vectoring of Interrupt Handler¶

Upon receipt of an interrupt the SPARC automatically performs the following actions:

• The trap level is set. This provides access to a fresh set of privileged trap-state registers used to save the current state, in effect, pushing a frame on the trap stack. TL <- TL + 1

• Existing state is preserved - TSTATE[TL].CCR <- CCR - TSTATE[TL].ASI <- ASI - TSTATE[TL].PSTATE <- PSTATE - TSTATE[TL].CWP <- CWP - TPC[TL] <- PC - TNPC[TL] <- nPC

• The trap type is preserved. TT[TL] <- the trap type

• The PSTATE register is updated to a predefined state - PSTATE.MM is unchanged - PSTATE.RED <- 0 - PSTATE.PEF <- 1 if FPU is present, 0 otherwise - PSTATE.AM <- 0 (address masking is turned off) - PSTATE.PRIV <- 1 (the processor enters privileged mode) - PSTATE.IE <- 0 (interrupts are disabled) - PSTATE.AG <- 1 (global regs are replaced with alternate globals) - PSTATE.CLE <- PSTATE.TLE (set endian mode for traps)

• For a register-window trap only, CWP is set to point to the register window that must be accessed by the trap-handler software, that is:

  • If TT[TL] = 0x24 (a clean window trap), then CWP <- CWP + 1.

  • If (0x80 <= TT[TL] <= 0xBF) (window spill trap), then CWP <- CWP + CANSAVE + 2.

  • If (0xC0 <= TT[TL] <= 0xFF) (window fill trap), then CWP <- CWP1.

  • For non-register-window traps, CWP is not changed.

• Control is transferred into the trap table:

  • PC <- TBA<63:15> (TL>0) TT[TL] 0 0000

  • nPC <- TBA<63:15> (TL>0) TT[TL] 0 0100

  • where (TL>0) is 0 if TL = 0, and 1 if TL > 0.

In order to safely invoke a subroutine during trap handling, traps must be enabled to allow for the possibility of register window overflow and underflow traps.

If the interrupt handler was installed as an RTEMS interrupt handler, then upon receipt of the interrupt, the processor passes control to the RTEMS interrupt handler which performs the following actions:

• saves the state of the interrupted task on it’s stack,

• switches the processor to trap level 0,

• if this is the outermost (i.e. non-nested) interrupt, then the RTEMS interrupt handler switches from the current stack to the interrupt stack,

• enables traps,

• invokes the vectors to a user interrupt service routine (ISR).

Asynchronous interrupts are ignored while traps are disabled. Synchronous traps which occur while traps are disabled may result in the CPU being forced into an error mode.

A nested interrupt is processed similarly with the exception that the current stack need not be switched to the interrupt stack.

20.4.3. Traps and Register Windows¶

XXX

20.4.4. Interrupt Levels¶

Sixteen levels (0-15) of interrupt priorities are supported by the SPARC architecture with level fifteen (15) being the highest priority. Level zero (0) indicates that interrupts are fully enabled. Interrupt requests for interrupts with priorities less than or equal to the current interrupt mask level are ignored.

Although RTEMS supports 256 interrupt levels, the SPARC only supports sixteen. RTEMS interrupt levels 0 through 15 directly correspond to SPARC processor interrupt levels. All other RTEMS interrupt levels are undefined and their behavior is unpredictable.

20.4.5. Disabling of Interrupts by RTEMS¶

XXX

20.4.6. Interrupt Stack¶

The SPARC architecture does not provide for a dedicated interrupt stack. Thus by default, trap handlers would execute on the stack of the RTEMS task which they interrupted. This artificially inflates the stack requirements for each task since EVERY task stack would have to include enough space to account for the worst case interrupt stack requirements in addition to it’s own worst case usage. RTEMS addresses this problem on the SPARC by providing a dedicated interrupt stack managed by software.

During system initialization, RTEMS allocates the interrupt stack from the Workspace Area. The amount of memory allocated for the interrupt stack is determined by the interrupt_stack_size field in the CPU Configuration Table. As part of processing a non-nested interrupt, RTEMS will switch to the interrupt stack before invoking the installed handler.

20.7.1. HelenOS and Open Firmware¶

The provided BSPs make use of some bootstrap and low-level hardware code of the HelenOS operating system. These files can be found in the shared/helenos directory of the sparc64 bsp directory. Consult the sources for more detailed information.

The shared BSP code also uses the Open Firmware interface to re-use firmware code, primarily for console support and default trap handlers.