TweetFollow Us on Twitter

PowerPC Series
Volume Number:10
Issue Number:1
Column Tag:PowerPC Series

PowerPC Code Generation

What’s the difference between PowerPC and 68K machines?

By Peter A. Jacobson, Absoft Corp.

About the author

Peter is a principle of Absoft. Along with his partner Wood Lotz, he has been developing scientific and engineering software since 1979 for a wide variety of micro- and mini-computers.

This article will discuss some of this issues concerning code generation by high level language compilers for the IBM PowerPC RISC microprocessor. It will compare and contrast typical code generation strategies employed on CISC based architectures, such as the Motorola M68000 family of microprocessors, against the approach that might be taken with the PowerPC. The topics addressed will include addressing modes, register sets, instruction sets, instruction pipelines, and superscalar considerations. It should be understood that certain features of the PowerPC will be simplified and various aspects of code generation will be trivialized in order to facilitate this discussion.

It is difficult to arrive at a precise definition of what constitutes a RISC microprocessor. It rarely means that an individual machine actually has fewer instructions than its CISC counterpart. The PowerPC has over 230 instructions while an MC68020 has barely 100. The technology progresses so quickly that definitions are amended before they even come into common usage. In addition, features that were once ascribed only to RISC technology have found their way into CISC architectures. However, one of the most significant differences affecting code generation for RISC microprocessors is that instructions are restricted to one machine word in length and there are consequently only a limited number of instruction formats available which can access memory. Typically, load and store are the only memory operations provided and usually with extremely limited effective addressing modes. The PowerPC provides just one fundamental addressing mode: register indirect with index. The index, which may be either an immediate operand or a general purpose register, is added to a general purpose register to form the effective address. The immediate operand is encoded in the instruction and consists of a 16-bit value, sign extended to 32 bits. The register can be suppressed by specifying general purpose register R0 so that an address is formed from just the index. In this way, absolute addresses can be formed from the immediate operand, but they are limited to just the lowest and the highest 32768 bytes of memory.

Obviously, this restraint seriously affects code generation strategies. With a Motorola MC680x0, an efficient code generator could add 1 to a variable by incrementing a memory location directly with an ADDQ instruction. While for the PowerPC, it is necessary to first load the variable from memory, perform the addition, and then store the result. It might appear that this model leads to very inefficient code production, but there are many other factors that must be considered in generating code. First, most program variables are usually accessed more than once in a given procedure. Therefore, for either type of microprocessor, it is almost always more efficient to have the variable already available in a register, rather than repeatedly accessing main memory for it. Since RISC microprocessors provide such a limited number of instructions which can access memory, code generators must be capable of performing a very sophisticated analysis of program flow and allocate registers accordingly. RISC microprocessors typically provide a large set of registers that can be used to maintain copies of program variables. The Motorola MC68040, widely used in the current generation of Macintoshes, is limited to 8 data registers, 8 address registers, 8 floating point registers, and a condition code register. The PowerPC provides 32 general purpose registers, 32 floating point registers, a condition register divided into eight 4-bit fields, and six user-level special purpose registers. The general purpose registers can be used for both addresses and data. Just as with the MC68040, some of the registers are reserved for special purposes (such as stack pointer, data space pointer, etc.), but the PowerPC still provides a large number of registers for program use.

Compilers also create their own variables, many of which can have short, but very active life spans. Such compiler created variables are used for loop induction, array indexing, maintaining the intermediate results of expression evaluation, and so on. The register set represents the fastest memory available to the microprocessor and efficient register allocation is critical to program performance. Various register allocation schemes are used by code generators to insure that the most appropriate variables are allocated to registers, either temporarily within a region of code, or permanently, for the length of the procedure. Further, compilers do not necessarily immediately write the result of an assignment statement to memory. This is known as a delayed store and is employed to allow efficient scheduling of the instruction stream (discussed below). Indeed, variables which are local to a procedure may never be written to memory. However, regardless of how efficiently the compiler allocates variables to registers, it must provide a mechanism by which a programmer can indicate that a variable (and its associated memory location) is volatile. Processes in many real time systems often communicate with each other through memory locations and use memory mapped I/O to control or react to external devices. If, by setting a variable to a specific value, the programmer intends to control a valve or launch a missile, it would be inappropriate (to say the least) for the code generator not to update the associated memory location immediately.

To the programmer accustomed to the Motorola M68000 family of microprocessors and unfamiliar with RISC architectures, the instruction set of the PowerPC may seem initially puzzling. Nevertheless, the PowerPC architecture has much in common with other RISC microprocessors such as the SPARC, MC88110, R4400, and obviously POWER. The first significant difference is that most of the instructions take three operands, two sources and a destination, and several instructions take more. Also, there is no stack pointer, no instructions for calling subroutines, no obvious way to move the contents of a general purpose register to another general purpose register, and many other apparent deficiencies. (However, programmers familiar will older mainframes and mini-computers will find nothing new here.) Consider the following instruction:

 fnmsubs6,12,13,18

This is the “Floating Point Negative Multiply-Subtract (Single-Precision)” instruction. Since there is no ambiguity in the instruction set, registers are indicated by number only - register numbers cannot be confused with immediate values. This instruction says to multiply the operand in floating point register 12 by the operand in floating point register 13 and then subtract the operand in floating point register 18 from this intermediate value. The result is rounded, then negated, and finally placed in floating point register 6. The latency of this instruction is just 4 clocks - the total time it takes to execute the instruction and for the result to be available in the destination floating point register.

Since every instruction can have a destination operand different from its source(s), compilers are not forced to either copy or reload values (variables or expressions) that will be used multiple times in a block of code. This is important not only in avoiding unnecessary memory accesses, but as will be seen later, provides opportunities for exploiting the instruction pipeline and the superscalar nature of the PowerPC.

The problem of there being no stack pointer in the PowerPC architecture has been addressed by the various standards bodies concerned with the PowerPC. Through the formalization and adoption of ABIs (Application Binary Interfaces) the needs of high-level languages for a uniform stack pointer and stack frame have been addressed. General purpose register 1 is normally designated as the stack pointer and various locations in the frame have been reserved for house keeping purposes. A frame is often created by saving the current stack pointer and then subtracting the required frame amount from the stack pointer to create the new frame. In practice it is easier to accomplish this than it appears since one form of the store instruction will write the effective address of the destination into the register used to calculate the effective address:

 stwu   rS,d(rA)

This is the “Store Word with Update” instruction which says to store the contents of the source register rS at an effective address equal to the contents of general purpose register rA plus the immediate index value d and then place that effective address in rA. To create a frame, rS and rA would be 1 and d would be negative. The instruction would cause r1 to be stored at the location resulting from the calculation of the effective address r1-d and then update r1 to r1-d.

One of the most important locations in the frame is naturally where the return address for a subroutine call is stored. As stated earlier, the PowerPC does not have a subroutine call instruction - instead the branch instruction is used. A form of this instruction places the address of the instruction that follows the branch into a special purpose register called the link register. Any procedure which is not a leaf (i.e. a procedure which calls other procedures) must save the link register before calling another procedure. A subroutine return is accomplished by simply branching to the contents of the link register.

The so-called fused multiply-add instructions are another feature of the PowerPC instruction set that is important enough to be mentioned here. These instructions can perform a multiplication and an addition in the same amount of time as just a single multiplication or a single addition alone. In other words, twice as fast as the combined operations. Fortunately, this type of operation occurs often enough in mathematical software that the alert code generator will find ample opportunities to exploit them. For example, expressions of the form:

 a1 = a0 + b x c

appear in matrix operations and in polynomial expansions.

The PowerPC implements a true superscalar architecture. A superscalar machine is one which can issue multiple instructions to different execution units during each clock cycle. The PowerPC incorporates three different execution units that can operate independently and in parallel. They are the integer unit which affects the general purpose registers, the floating point unit which affects the floating point registers, and the branch unit which affects certain of the special purpose registers. Therefore, an integer shift, a floating point addition, and a branch instruction could all be issued during the same clock cycle. It is important to understand that not all of the PowerPC instructions can execute in a single clock cycle and it would be extremely difficult to schedule all three execution units for simultaneous execution on every cycle, but with careful code generation and attention paid to data dependencies, an exceptionally efficient throughput can be achieved.

It is not necessary for an instruction to completely finish in an individual execution unit before another instruction can be issued. The execution of an instruction consists of multiple stages that can be viewed (very roughly for the PowerPC is far more complicated) as fetch, decode, execute, and writeback. Each instruction is fetched from an instruction queue, decoded, executed, and the result is then written to the appropriate register file. These stages are called the pipeline and it is possible and certainly desirable for multiple instructions to be in the pipeline at once - each at a different stage. The basic limitation which would cause an instruction to stall is data dependency, which means that the execution of the instruction is dependant on the result of the preceding instruction. An instruction can also be stalled if it is waiting for an instruction with a latency greater than once clock to finish executing. That is, an instruction takes more cycles than there are stages in the pipeline for that execution unit. Instruction latency is determined by how complicated an instruction is (division takes longer than addition) and by memory access considerations. An instruction may stall while waiting for an operand to be delivered from memory. The issues of cache arbitration, both for instructions and data, are beyond the scope of this article.

A code generator which is aware of these two features, multiple execution units and their pipelines, attempts to schedule the instruction stream to make the most efficient use of the resources. Scheduling consists largely of the code generator rearranging or moving instructions to eliminate data dependencies and to keep the individual pipelines busy. This can cause expressions to executed out of order, array element address calculations to take place far from the memory references, and any number of other reorderings of the instruction stream to eliminate data dependencies. Obviously, register allocation seriously affects this scheduling process and is usually put off as long as possible to prevent any artificial or code-generator created dependencies.

“It projects a military coup!”

 

Community Search:
MacTech Search:

Software Updates via MacUpdate

Duplicate Annihilator 5.7.5 - Find and d...
Duplicate Annihilator takes on the time-consuming task of comparing the images in your iPhoto library using effective algorithms to make sure that no duplicate escapes. Duplicate Annihilator... Read more
BusyContacts 1.0.2 - Fast, efficient con...
BusyContacts is a contact manager for OS X that makes creating, finding, and managing contacts faster and more efficient. It brings to contact management the same power, flexibility, and sharing... Read more
Capture One Pro 8.2.0.82 - RAW workflow...
Capture One Pro 8 is a professional RAW converter offering you ultimate image quality with accurate colors and incredible detail from more than 300 high-end cameras -- straight out of the box. It... Read more
Backblaze 4.0.0.872 - Online backup serv...
Backblaze is an online backup service designed from the ground-up for the Mac.With unlimited storage available for $5 per month, as well as a free 15-day trial, peace of mind is within reach with... Read more
Little Snitch 3.5.2 - Alerts you about o...
Little Snitch gives you control over your private outgoing data. Track background activity As soon as your computer connects to the Internet, applications often have permission to send any... Read more
Monolingual 1.6.4 - Remove unwanted OS X...
Monolingual is a program for removing unnecesary language resources from OS X, in order to reclaim several hundred megabytes of disk space. If you use your computer in only one (human) language, you... Read more
CleanApp 5.0 - Application deinstaller a...
CleanApp is an application deinstaller and archiver.... Your hard drive gets fuller day by day, but do you know why? CleanApp 5 provides you with insights how to reclaim disk space. There are... Read more
Fantastical 2.0 - Create calendar events...
Fantastical is the Mac calendar you'll actually enjoy using. Creating an event with Fantastical is quick, easy, and fun: Open Fantastical with a single click or keystroke Type in your event details... Read more
Cocktail 8.2 - General maintenance and o...
Cocktail is a general purpose utility for OS X that lets you clean, repair and optimize your Mac. It is a powerful digital toolset that helps hundreds of thousands of Mac users around the world get... Read more
Direct Mail 4.0.4 - Create and send grea...
Direct Mail is an easy-to-use, fully-featured email marketing app purpose-built for OS X. It lets you create and send great looking email campaigns. Start your newsletter by selecting from a gallery... Read more

Fast & Furious: Legacy's Creati...
| Read more »
N-Fusion and 505's Ember is Totally...
| Read more »
These are All the Apple Watch Apps and G...
The Apple Watch is less than a month from hitting store shelves, and once you get your hands on it you're probably going to want some apps and games to install. Fear not! We've compiled a list of all the Apple Watch apps and games we've been able to... | Read more »
Appy to Have Known You - Lee Hamlet Look...
Being at 148Apps these past 2 years has been an awesome experience that has taught me a great deal, and working with such a great team has been a privilege. Thank you to Rob Rich, and to both Rob LeFebvre and Jeff Scott before him, for helping me... | Read more »
Hands-On With Allstar Heroes - A Promisi...
Let’s get this out of the way quickly. Allstar Heroes looks a lot like a certain other recent action RPG release, but it turns out that while it’s not yet available here, Allstar Heroes has been around for much longer than that other title. Now that... | Read more »
Macho Man and Steve Austin Join the Rank...
WWE Immortals, by Warner Bros. Interactive Entertainment and WWE, has gotten a superstar update. You'll now have access to Macho Man Randy Savage and Steve Austin. Both characters have two different versions: Macho Man Randy Savage Renegade or Macho... | Read more »
Fearless Fantasy is Fantastic for the iF...
I actually had my first look at Fearless Fantasy last year at E3, but it was on a PC so there wasn't much for me to talk about. But now that I've been able to play with a pre-release version of the iOS build, there's quite a bit for me to talk... | Read more »
MLB Manager 2015 (Games)
MLB Manager 2015 5.0.14 Device: iOS Universal Category: Games Price: $4.99, Version: 5.0.14 (iTunes) Description: Guide your favorite MLB franchise to glory! MLB Manager 2015, officially licensed by MLB.com and based on the award-... | Read more »
Breath of Light (Games)
Breath of Light 1.0.1421 Device: iOS Universal Category: Games Price: $2.99, Version: 1.0.1421 (iTunes) Description: Hold a quiet moment. Breath of Light is a meditative and beautiful puzzle game with a hypnotic soundtrack by... | Read more »
WWE WrestleMania Tags into the App Store
Are You ready to rumble? The official WWE WrestleMania app, by World Wrestling Entertainment, is now available. Now you can get all your WrestleMania info in one place before anyone else. The app offers details on superstar signings, interactive... | Read more »

Price Scanner via MacPrices.net

13-inch 2.5GHz MacBook Pro (refurbished) avai...
The Apple Store has Apple Certified Refurbished 13″ 2.5GHz MacBook Pros available for $829, or $270 off the cost of new models. Apple’s one-year warranty is standard, and shipping is free: - 13″ 2.... Read more
Save up to $80 on iPad Air 2s, NY tax only, f...
 B&H Photo has iPad Air 2s on sale for $80 off MSRP including free shipping plus NY sales tax only: - 16GB iPad Air 2 WiFi: $469.99 $30 off - 64GB iPad Air 2 WiFi: $549.99 $50 off - 128GB iPad... Read more
iMacs on sale for up to $205 off MSRP
B&H Photo has 21″ and 27″ iMacs on sale for up to $205 off MSRP including free shipping plus NY sales tax only: - 21″ 1.4GHz iMac: $1019 $80 off - 21″ 2.7GHz iMac: $1189 $110 off - 21″ 2.9GHz... Read more
Färbe Technik Offers iPhone Battery Charge LI...
Färbe Technik, which manufactures and markets of mobile accessories for Apple, Blackberry and Samsung mobile devices, is offering tips on how to keep your iPhone charged while in the field: •... Read more
Electronic Recyclers International CEO Urges...
Citing a recent story on CNBC about concerns some security professionals have about the forthcoming Apple Watch, John Shegerian, Chairman and CEO of Electronic Recyclers International (ERI), the... Read more
Save up to $380 with Apple refurbished iMacs
The Apple Store has Apple Certified Refurbished iMacs available for up to $380 off the cost of new models. Apple’s one-year warranty is standard, and shipping is free: - 27″ 3.5GHz 5K iMac – $2119 $... Read more
Mac minis on sale for up to $75 off, starting...
MacMall has Mac minis on sale for up to $75 off MSRP including free shipping. Their prices are the lowest available for these models from any reseller: - 1.4GHz Mac mini: $459.99 $40 off - 2.6GHz Mac... Read more
College Student Deals: Additional $50 off Mac...
Take an additional $50 off all MacBooks and iMacs at Best Buy Online with their College Students Deals Savings, valid through April 11, 2015. Anyone with a valid .EDU email address can take advantage... Read more
Mac Pros on sale for up to $260 off MSRP
B&H Photo has Mac Pros on sale for up to $260 off MSRP. Shipping is free, and B&H charges sales tax in NY only: - 3.7GHz 4-core Mac Pro: $2799, $200 off MSRP - 3.5GHz 6-core Mac Pro: $3719.99... Read more
13-inch 2.5GHz MacBook Pro on sale for $100 o...
B&H Photo has the 13″ 2.5GHz MacBook Pro on sale for $999 including free shipping plus NY sales tax only. Their price is $100 off MSRP. Read more

Jobs Board

DevOps Software Engineer - *Apple* Pay, iOS...
**Job Summary** Imagine what you could do here. At Apple , great ideas have a way of becoming great products, services, and customer experiences very quickly. Bring Read more
*Apple* Retail - Multiple Positions (US) - A...
Sales Specialist - Retail Customer Service and Sales Transform Apple Store visitors into loyal Apple customers. When customers enter the store, you're also the Read more
Sr. Technical Services Consultant, *Apple*...
**Job Summary** Apple Professional Services (APS) has an opening for a senior technical position that contributes to Apple 's efforts for strategic and transactional Read more
Lead *Apple* Solutions Consultant - Retail...
**Job Summary** Job Summary The Lead ASC is an Apple employee who serves as the Apple business manager and influencer in a hyper-business critical Reseller's store Read more
*Apple* Pay - Site Reliability Engineer - Ap...
**Job Summary** Imagine what you could do here. At Apple , great ideas have a way of becoming great products, services, and customer experiences very quickly. Bring Read more
All contents are Copyright 1984-2011 by Xplain Corporation. All rights reserved. Theme designed by Icreon.