Subversion Repositories forwardcom

% chapter included in forwardcom.tex
\documentclass[forwardcom.tex]{subfiles}
\begin{document}
\RaggedRight
 
\chapter{Instruction formats}
\section{Formats and templates}
All instructions use one of the general format templates shown below (the most significant bits are shown to the left). The basic layout of the 32-bit code word is shown in template A. Template B, C and D are derived from template A by replacing 8, 16, or 24 bits, respectively, with immediate data. Double-size and triple-size instructions can be constructed by adding one or two 32-bit words to one of these templates. For example, template A with an extra 32-bit word containing data is called A2. Template E2 is an extension to template A where the second code word contains an extra register field, extra opcode bits, mode bits, option bits, and data.
\vspace{4mm}
 
\begin{table}[h!] \label{table:templateA}
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|} \hline
 Bits & 2 & 3 & 6 & 5 & 1 & 2 & 5 & 3 & 5 \\ \hline
Field & IL & Mode & OP1 & RD & M & OT & RS & Mask & RT  \\ \hline
\multicolumn{10}{|l|}{
\textbf{Template A}. Has three operand registers and a mask register.} \\ \hline
\end{tabular}
\end{table}
\vv
 
\begin{table}[h!] \label{table:templateB}
\vv
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{24mm}|} \hline
 Bits & 2 & 3 & 6 & 5 & 1 & 2 & 5 & 8 \\ \hline
Field & IL & Mode & OP1 & RD & M & OT & RS & IM1 \\ \hline
\multicolumn{9}{|l|}{
\textbf{Template B}. Has two operand registers and an 8-bit immediate constant.} \\ \hline
\end{tabular}
\end{table}
\vv
 
\begin{table}[h!] \label{table:templateC}
\vv
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{38.5mm}|p{24mm}|} \hline
 Bits & 2 & 3 & 6 & 5 & \hspace{15mm} 8 & \hspace{8mm} 8 \\ \hline
Field & IL & Mode & OP1 & RD & \hspace{14mm} IM2 & \hspace{7mm} IM1 \\ \hline
\multicolumn{7}{|l|}{
\textbf{Template C}. Has one operand register two 8-bit immediate constants.} \\ \hline
\end{tabular}
\end{table}
\vv
 
\begin{table}[h!] \label{table:templateD}
\vv
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{81.5mm}|} \hline
 Bits & 2 & 3 & 3 & \hspace{33mm} 24 \\ \hline
Field & IL & Mode & OP1 & \hspace{32mm} IM2 \\ \hline
\multicolumn{5}{|l|}{
\textbf{Template D}. Has no register and a 24-bit immediate constant.
} \\ \hline
\end{tabular}
\end{table}
\vv
 
\begin{table}[h!] \label{table:templateA2}
\vv
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|} \hline
 Bits & 2 & 3 & 6 & 5 & 1 & 2 & 5 & 3 & 5 \\ \hline
Field & IL & Mode & OP1 & RD & M & OT & RS & Mask & RT  \\ \hline
Field & \multicolumn{9}{|c|}{ IM2 } \\ \hline
\multicolumn{10}{|l|}{
\textbf{Template A2}. 2 words. As A, with an extra 32-bit immediate constant.
} \\ \hline
\end{tabular}
\end{table}
\vv
 
\begin{table}[h!] \label{table:templateB2}
\vv
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{24mm}|} \hline
 Bits & 2 & 3 & 6 & 5 & 1 & 2 & 5 & 8 \\ \hline
Field & IL & Mode & OP1 & RD & M & OT & RS & IM1 \\ \hline
Field & \multicolumn{8}{|c|}{ IM2 } \\ \hline
\multicolumn{9}{|l|}{
\textbf{Template B2}. As B, with an extra 32-bit immediate constant.} \\ \hline
\end{tabular}
\end{table}
\vv
 
\begin{table}[h!] \label{table:templateC2}
\vv 
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{38.5mm}|p{24mm}|} \hline
 Bits & 2 & 3 & 6 & 5 & \hspace{15mm} 8 & \hspace{8mm} 8 \\ \hline
Field & IL & Mode & OP1 & RD & \hspace{14mm} IM2 & \hspace{7mm} IM1 \\ \hline
Field & \multicolumn{6}{|c|}{ IM3 } \\ \hline
\multicolumn{7}{|l|}{
\textbf{Template C2}. As C, with an extra 32-bit immediate constant.} \\ \hline
\end{tabular}
\end{table}
\vv
 
\begin{table}[h!] \label{table:templateE2}
\vv 
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|} \hline
Bits & 2 & 3 & 6 & 5 & 1 & 2 & 5 & 3 & 5 \\ \hline
Field & IL & Mode & OP1 & RD & M & OT & RS & Mask & RT  \\ \hline
Bits & 3 & 5 & 2 & 6 & \multicolumn{5}{l|}{ \hspace{29mm} 16  } \\ \hline
Field  & Mode2 & RU & OP2 & IM3 & \multicolumn{5}{l|}{ \hspace{28mm} IM2 } \\ \hline
\multicolumn{10}{|l|}{
\textbf{Template E2}. Has 4 register operands, mask, a 16-bit immediate constant, } \\
\multicolumn{10}{|l|}{
and extra bits for mode, opcode, and options. } \\ \hline
\end{tabular}
\end{table}
\vv
 
\begin{table}[h!] \label{table:templateA3}
\vv
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|} \hline
 Bits & 2 & 3 & 6 & 5 & 1 & 2 & 5 & 3 & 5 \\ \hline
Field & IL & Mode & OP1 & RD & M & OT & RS & Mask & RT  \\ \hline
Field & \multicolumn{9}{|c|}{ IM2 } \\ \hline
Field & \multicolumn{9}{|c|}{ IM3 } \\ \hline
\multicolumn{10}{|l|}{
\textbf{Template A3}. 3 words. As A, with two extra 32-bit immediate constants.
} \\ \hline
\end{tabular}
\end{table}
\vv
 
\begin{table}[h!] \label{table:templateB3}
\vv
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{24mm}|} \hline
 Bits & 2 & 3 & 6 & 5 & 1 & 2 & 5 & 8 \\ \hline
Field & IL & Mode & OP1 & RD & M & OT & RS & IM1 \\ \hline
Field & \multicolumn{8}{|c|}{ IM2 } \\ \hline
Field & \multicolumn{8}{|c|}{ IM3 } \\ \hline
\multicolumn{9}{|l|}{
\textbf{Template B3}. As B, with two extra 32-bit immediate constants.
} \\ \hline
\end{tabular}
\end{table}
\vspace{4mm}
 
\begin{table}[h!] \label{table:templateE3}
\vv
\begin{tabular}{|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|p{10mm}|} \hline
Bits & 2 & 3 & 6 & 5 & 1 & 2 & 5 & 3 & 5 \\ \hline
Field & IL & Mode & OP1 & RD & M & OT & RS & Mask & RT  \\ \hline
Bits & 3 & 5 & 2 & 6 & \multicolumn{5}{l|}{ \hspace{29mm} 16  } \\ \hline
Field  & Mode2 & RU & OP2 & IM3 & \multicolumn{5}{l|}{ \hspace{28mm} IM2 } \\ \hline
Field & \multicolumn{9}{|c|}{ IM4 } \\ \hline
\multicolumn{10}{|l|}{
\textbf{Template E3}. As E2, with an extra 32-bit immediate constant.
} \\ \hline
\end{tabular}
\end{table}
\vspace{4mm}
 
The meaning of each field is described in the following table.
 
\pagebreak
 
\begin{longtable} {|p{16mm}|p{16mm}|p{85mm}|}
\caption{Fields in instruction templates} \label{table:fieldsInTemplates} \\
\endfirsthead
\endhead
\hline
\bfseries Field name & \bfseries Meaning & \bfseries Values  \\
\hline
IL & Instruction length & 0 or 1: 1 word = 32 bits \newline
2: 2 words = 64 bits \newline
3: 3 words (possibly more in future extensions if mode > 3)  \\
\hline
Mode & Format & Determines the format template and the use of each field. 
Extended with the M bit when needed. \newline 
See details below. \\
\hline
Mode2 & Format & Extension to Mode. \\
\hline
OT & Operand type and size (OS) & 
0: 8 bit integer, OS = 1 byte  \newline
1: 16 bit integer, OS = 2 bytes \newline
2: 32 bit integer, OS = 4 bytes \newline
3: 64 bit integer, OS = 8 bytes \newline
4: 128 bit integer, OS = 16 bytes (optional) \newline
5: single precision float, OS = 4 bytes \newline
6: double precision float, OS = 8 bytes \newline
7: quadruple precision float, OS = 16 bytes (optional) \newline
The OT field is extended with the M bit when needed. \\
\hline
M & Operand type or mode & Extends the mode field when bit 1 and bit 2 of Mode are both zero (general purpose registers). Extends the OT field otherwise (vector registers).  \\
\hline
OP1 & Opcode & Decides the operation, for example add or move.  \\
\hline
OP2 & Opcode & Opcode extension for single-format instructions. \newline
               May also be used as an extension to IM3. \\
\hline
RD & Destination register & r0 – r31 or v0 – v31. Also used for first source operand and fallback if the instruction format does not specify enough operands. \\
\hline
RS & Source register & r0 – r31 or v0 – v31. Source register, pointer, or fallback. \\
\hline
RT & Source register & r0 – r31 or v0 – v31. Source register, index, or vector length.  \\
\hline
RU & Source register & r0 – r31 or v0 – v31. Source register or fallback. \\
\hline
Mask & mask register & 0-6 means that a general purpose register or vector register is used for mask and option bits. 7 means no mask.  \\
\hline
IM1 IM2 IM3 IM4 & Immediate data & 8, 16, 24, or 32 bits immediate operand or address offset or option bits. Adjacent IM fields can be merged to make a larger constant. \\
\hline
\end{longtable}
\vv
 
Instructions have several different formats, defined by the IL and mode bits, according to  table \ref{table:instructionFormats} below. The different formats specify different sizes of immediate data or memory operands with different addressing modes. \\
\vv
 
Instructions can have up to three source operands (input), one destination operand (output), and a mask. The destination operand always uses the RD field, except where the destination is a memory operand. The source operands are using the available operand fields according to the following algorithm: The required source operands are assigned to the available
operand fields defined by table \ref{table:instructionFormats} in the following order of priority: immediate data field, memory operand, RT, RS, RU, RD. 
The operands are assigned in reverse order so that the last operand gets the field that comes first in this order of priority. For example, the instruction r1 = r2 - r3 using template A will be RD = RS - RT. RD is used for both destination and the first source operand only if there are no other vacant register fields.
\vv
 
The coding of instructions with two or three source operands is indicated in the table in the following way: \\
RD = f2(RS,RT)  means that instructions with two input operands (f2) use the register specified in RD as destination operand and RS and RT as source operands.\\
RD=f3(RD, RU, [RS+RT*OS+IM2])  means that instructions with three input operands (f3) use the register specified in RD as both destination and the first source operand. The second source operand is RU. The third source operand is a memory operand with RS as base pointer, RT as index scaled by the operand size, and the constant IM2 as offset.\\
Instructions with only one input operand are coded as f2 with the first source operand omitted.
 
\begin{longtable} {|p{10mm}|p{6mm}|p{9mm}|p{7mm}|p{80mm}|}
\caption{List of instruction formats} \label{table:instructionFormats} \\
\endfirsthead
\endhead
\hline
\bfseries Format name & \bfseries IL & \bfseries Mode. \small Mode2 & \bfseries Tem-plate & \bfseries Use \\
\hline
0.0 & 0 & 0 & A & Three general purpose register operands.\newline 
RD = f2(RS, RT). RD = f3(RD, RS, RT).\\
 
\hline
0.1 & 0 & 1 & B & Two general purpose registers and 8-bit immediate operand. \newline
RD = f2(RS, IM1). RD = f3(RD, RS, IM1).\\
 
\hline
0.2 & 0 & 2 & A & Three vector register operands.\newline 
RD = f2(RS, RT). RD = f3(RD, RS, RT).\\
 
\hline
0.3 & 0 & 3 & B & Two vector registers and a broadcast 8-bit immediate operand. \newline
RD = f2(RS, IM1). RD = f3(RD, RS, IM1).\\
 
\hline
0.4 & 0 & 4 & A & One vector register and memory operand. Vector length specified by general purpose register. \newline
RD = f2(RD, [RS]). length=RT.\\
 
\hline
0.5 & 0 & 5 & A & One vector register and a memory operand with base pointer and negative index.  This is used for vector loops as explained on page \pageref{vectorLoops}. \newline
RD = f2(RD, [RS-RT]). length=RT.\\
 
\hline
0.6 & 0 & 6 & A & One vector register and a scalar memory operand with base pointer and scaled index. \newline
RD = f2(RD, [RS+RT*OS]).\\
 
\hline
0.7 & 0 & 7 & B & One vector register and a scalar memory operand with base pointer and 8-bit offset. \newline
RD = f2(RD, [RS+IM1*OS]).\\
 
\hline
0.8 & 0 & 0 M=1 & A & One general purpose register and a memory operand with base pointer and scaled index. \newline
RD = f2(RD, [RS+RT*OS]).\\
 
\hline
0.9 & 0 & 1 M=1 & B & One general purpose register and a memory operand with base pointer and 8-bit offset. \newline
RD = f2(RD, [RS+IM1*OS]).\\
 
\hline
1.0 & 1 & 0 & A & Single-format instructions. Three general purpose register operands. \newline 
RD = f2(RS, RT). RD = f3(RD, RS, RT).\\
 
\hline
1.1 & 1 & 1 & C & Single-format instructions. One general purpose register and a 16-bit immediate operand. \newline 
RD = f2(RD, IM1-2).\\
 
\hline
1.2 & 1 & 2 & A & Single-format instructions. Three vector register operands. \newline 
RD = f2(RS, RT). RD = f3(RD, RS, RT).\\
 
\hline
1.3 & 1 & 3 & B & Single-format instructions. Two vector registers and a broadcast 8-bit immediate operand. \newline 
RD = f2(RS, IM1). RD = f3(RD, RS, IM1). \\
 
\hline
1.4 & 1 & 4 & C & Single-format instructions. One vector register and a broadcast 16-bit immediate operand. \newline 
RD = f2(RD, IM1-2). \\
 
\hline
1.5 & 1 & 5 &  & Vacant. May be used for application-specific vector instructions. \\
 
\hline
1.6 A & 1 & 6 & A & Multiway jump instructions and system calls with three register operands.\\
 
\hline
1.6 B & 1 & 6 & B & Jump instructions with two register operands and 8 bit offset.\\
 
\hline
1.7C & 1 & 7 & C & Jump instructions with one register operand, 8 bit constant (IM2) and 8 bit offset (IM1).\\
 
\hline
1.7D & 1 & 7 & D & Jump instructions with no register and 24 bit offset.  \\
 
\hline
1.8 & 1 & 0 M=1 & B & Single-format instructions. Two general purpose registers and an 8-bit immediate operand.\newline 
RD = f2(RS, IM1). RD = f3(RD, RS, IM1).\\
 
\hline
1.9 &  &  &  & There is no format 1.9 because 1.1 has no M bit.\\
 
\hline
2.0.0 & 2 & 0.0  & E2 & Three general purpose registers and a memory operand with base and 16 bit offset.\newline 
RD = f2(RT, [RS+IM2]). \newline 
RD = f3(RU, RT, [RS+IM2]).\\
 
\hline
2.0.1 & 2 & 0.1  & E2 & Two general purpose registers and a memory operand with base, index and optional 16 bit offset, no scale.\newline  
RD = f2(RU, [RS+RT+IM2]).\newline  
RD = f3(RD, RU, [RS+RT+IM2]). \\
 
\hline
2.0.2 & 2 & 0.2  & E2 & Two general purpose registers and a memory operand with base,  scaled index, and optional 16 bit offset.\newline   
RD = f2(RU, [RS+RT*OS+IM2]). \newline 
RD = f3(RD, RU, [RS+RT*OS+IM2]). \\
 
\hline
2.0.3 & 2 & 0.3  & E2 & Two general purpose registers and a memory operand with base, scaled index, and 16-bit limit. Optional. \newline 
RD = f2(RU, [RS+RT*OS]). \newline 
RD = f3(RD, RU, [RS+RT*OS]). \newline
Limit RT $\leq$ IM2 (unsigned).\newline 
Support for this format is optional. \\
 
\hline
2.0.5 & 2 & 0.5  & E2 & One general purpose register and a memory operand with base, scaled index, 16-bit offset, and an 8-bit immediate operand using IM3 extended with OP2.  Optional. \newline 
RD = f2([RS+RT*OS+IM2], IM3). \newline 
RD = f3(RU, [RS+RT*OS+IM2], IM3). \\ 
 
\hline
2.0.6 & 2 & 0.6  & E2 & Four general purpose registers.\newline 
RD = f2(RS, RT). \newline 
RD = f3(RU, RS, RT).\\
 
\hline
2.0.7 & 2 & 0.7  & E2 & Three general purpose registers and a 16-bit integer with left shift.\newline 
RD = f2(RT, IM2). \newline 
RD = f3(RS, RT, IM2).\newline 
IM2 (signed) is shifted left by the 6-bit unsigned value of IM3, or whithout shift if IM3 is used for other purposes. \\
 
\hline
2.1 & 2 & 1 & A2 & Two general purpose registers and a memory operand with base and 32 bit offset (IM2). \newline 
RD = f2(RT, [RS+IM2]). \newline 
RD = f3(RD, RT, [RS+IM2]).\\
 
\hline
2.2.0 & 2 & 2.0 & E2 & Two vector registers and a broadcast scalar memory operand with base  and 16 bit offset.\newline 
RD = f2(RU, [RS+IM2]). \newline 
RD = f3(RD, RU, [RS+IM2]). \newline
Broadcast to length RT.\\
 
\hline
2.2.1 & 2 & 2.1 & E2 & Two vector registers and a memory operand with base and 16 bit offset.\newline 
RD = f2(RU, [RS+IM2]). \newline 
RD = f3(RD, RU, [RS+IM2]).\newline
Length=RT.\\
 
\hline
2.2.2 & 2 & 2.2 & E2 & Two vector registers and a scalar memory operand with base and scaled index. \newline 
RD = f2(RU, [RS+RT*OS+IM2]). \newline 
RD = f3(RD, RU, [RS+RT*OS+IM2]). \\
 
\hline
2.2.3 & 2 & 2.3 & E2 & Two vector registers and a scalar memory operand with base, scaled index, and 16-bit limit. Optional. \newline 
RD = f2(RU, [RS+RT*OS]). \newline 
RD = f3(RD, RU, [RS+RT*OS]).\newline
Limit RT $\leq$ IM2 (unsigned).\\
 
\hline
2.2.4 & 2 & 2.4 & E2 & Two vector registers and a memory operand with base and negative index. \newline 
RD = f2(RU, [RS-RT+IM2]). \newline 
RD = f3(RD, RU, [RS-RT+IM2]). \newline 
Length=RT. \\
 
\hline
2.2.5 & 2 & 2.5  & E2 & One vector register and a memory operand with base, 16-bit offset, and an 8-bit immediate operand using IM3 extended with OP2. Optional. \newline 
RD = f2([RS+IM2], IM3). \newline 
RD = f3(RU, [RS+IM2], IM3). \newline 
Length=RT.\\
 
\hline
2.2.6 & 2 & 2.6 & E2 & Four vector registers.\newline 
RD = f2(RS, RT). \newline 
RD = f3(RU, RS, RT).\\
 
\hline
2.2.7 & 2 & 2.7 & E2 & Three vector registers and a broadcast immediate half-precision float or 16-bit integer with left shift.\newline 
RD = f2(RT, IM2). \newline 
RD = f3(RS, RT, IM2).\newline
Floating point operands: IM2 is half precision.
Integer operands: IM2 (signed) is shifted left by the 6-bit unsigned value of IM3, or whithout shift if IM3 is used for other purposes. \\
 
\hline
2.3 & 2 & 3 & A2 & Three vector registers and a broadcast 32-bit immediate operand.\newline 
RD = f2(RT, IM2). \newline 
RD = f3(RS, RT, IM2).\\
 
\hline
2.4 & 2 & 4 & A2 & One vector register and a memory operand with base and 32 bit offset.\newline
RD = f2(RD, [RS+IM2]). length=RT.\\
 
\hline
2.5 & 2 & 5 & A2, B2, C2 & Jump instructions for OP1 $<$ 8. Single format instructions with memory operands or mixed register types for OP1 $\geq$ 8.\\
 
\hline
2.6 & 2 & 6 & A2 & Single-format instructions. Three vector registers and a 32-bit immediate operand.\newline 
RD = f2(RT, IM2). \newline 
RD = f3(RS, RT, IM2).\\
 
\hline
2.7 & 2 & 7 &  & Currently unused.\\
 
\hline
2.8 & 2 & 0 M=1 & A2 & Three general purpose registers and a 32-bit immediate operand.\newline 
RD = f2(RT, IM2). \newline 
RD = f3(RS, RT, IM2).\\
 
\hline
2.9 & 2 & 1 M=1 & A2 & Single-format instructions. Three general purpose registers and a 32-bit immediate operand.\newline 
RD = f2(RT, IM2). \newline 
RD = f3(RS, RT, IM2).\\
 
 
\hline
3.0.0 & 3 & 0.0  & E3 & Three general purpose registers and a memory operand with base and 32 bit offset.\newline 
RD = f2(RT, [RS+IM4]). \newline 
RD = f3(RU, RT, [RS+IM4]).\\
 
\hline
3.0.2 & 3 & 0.2  & E3 & Two general purpose registers and a memory operand w. base, scaled index, and 32 bit offset.\newline 
RD = f2(RU, [RS+RT*OS+IM4]). \newline 
RD = f3(RD, RU, [RS+RT*OS+IM4]). \\
 
\hline
3.0.3 & 3 & 0.3  & E3 & Two general purpose registers and a memory operand with base, scaled index, and 32-bit limit. Optional. \newline 
RD = f2(RU, [RS+RT*OS]). \newline 
RD = f3(RD, RU, [RS+RT*OS]). \newline 
Limit RT $\leq$ IM4 (unsigned). \\
 
\hline
3.0.5 & 3 & 0.5  & E3 & One general purpose register and a memory operand with base, scaled index, 16-bit offset, and a 32-bit immediate operand. Optional. \newline 
RD = f2([RS+RT*OS+IM2], IM4). \newline 
RD = f3(RU, [RS+RT*OS+IM2], IM4). \\ 
 
\hline
3.0.7 & 3 & 0.7  & E3 & Three general purpose registers and a 32-bit integer with left shift.\newline 
RD = f2(RS, IM4 $<<$ IM2). \newline 
RD = f3(RS, RT, IM4 $<<$ IM2). \newline 
IM4 (signed) is shifted left by the unsigned value of IM2. \\
 
\hline
3.1 & 3 & 1 & A3, B3 & Jump instructions for OP1 $<$ 8. Single format instructions with memory operands or mixed register types for OP1 $\geq$ 8.\\
 
\hline
3.2.0 & 3 & 2.0 & E3 & Two vector registers and a broadcast scalar memory operand with base and 32 bit offset.\newline 
RD = f2(RU, [RS+IM4]). \newline 
RD = f3(RD, RU, [RS+IM4]). \newline 
Broadcast to length RT.\\
 
\hline
3.2.1 & 3 & 2.1 & E3 & Two vector registers and a memory operand with base and 32 bit offset.\newline 
RD = f2(RU, [RS+IM4]). \newline 
RD = f3(RD, RU, [RS+IM4]). \newline 
Length=RT.\\
 
\hline
3.2.2 & 3 & 2.2 & E3 & Two vector registers and a scalar memory operand w. base, scaled index, and 32-bit offset. Optional. \newline 
RD = f2(RU, [RS+RT*OS+IM4]). \newline 
RD = f3(RD, RU, [RS+RT*OS+IM4]).\\
 
\hline
3.2.3 & 3 & 2.3 & E3 & Two vector registers and a scalar memory operand with base, scaled index, and 32-bit limit. Optional. \newline 
RD = f2(RU, [RS+RT*OS]). \newline 
RD = f3(RD, RU, [RS+RT*OS]). \newline 
Limit RT $\leq$ IM4 (unsigned).\\
 
\hline
3.2.5 & 3 & 2.5  & E3 & One vector register and a memory operand with base, 16-bit offset, and a 32-bit immediate operand. Optional. \newline 
RD = f2([RS+IM2], IM4). \newline 
RD = f3(RU, [RS+IM2], IM4). \newline 
Length=RT.\\
 
\hline
3.2.7 & 3 & 2.7 & E3 & Three vector registers and a broadcast single precision float or 32-bit integer with left shift.\newline 
RD = f2(RT, IM4). \newline 
RD = f3(RS, RT, IM4).\newline
Floating point operands: IM4 is single precision.
Integer operands: IM4 (signed) is shifted left by the unsigned value of IM2. \\
 
\hline
3.3 & 3 & 3 & A3 & Three vector registers and a broadcast 64-bit immediate operand.\newline 
RD = f2(RT, IM3:IM2). \newline 
RD = f3(RS, RT, IM3:IM2).\\
 
\hline
3.8 & 3 & 0 M=1 & A3 & Three general purpose registers and a 64-bit immediate operand. \newline 
RD = f2(RT, IM3:IM2). \newline 
RD = f3(RS, RT, IM3:IM2).\\
 
\hline
3.9 &  &  &  & There is no format 3.9 because 3.1 uses the M bit.\\
 
\hline
4.x & 3 & 4-7 &  & Reserved for future 4-word instructions and longer. \\
\hline
\end{longtable}
 
 
%\vspace{2mm}
%\subsection{Maximum number of input operands}
%The hardware supports a maximum of four or five input dependencies. Three-input instructions cannot have both a mask and a memory operand with base and index or vector length if the hardware has a limit of four input dependencies. For example the mul\_add instruction cannot have a mask in format 2.2.0 if the hardware does not support five inputs.
 
\vv
\section{Coding of operands}
\subsection{Operand type}
The type and size of operands is determined by the OT field as indicated above. The operand type is 32 bit integer if there is no OT field unless otherwise specified. The operand size (OS) is the size in bytes of a scalar operand or a vector element. This is equal to the number of bits divided by 8.
 
\subsection{Register type}
The instructions can use either general purpose registers or vector registers. General purpose registers are used for source and destination operands and for masks if the Mode field is 0 or 1 (with M = 0 or 1). Vector registers are used for source and destination operands and for masks if Mode is 2-7. Jump instructions use vector registers if M = 1. A few single-format instructions deviate from this rule and use mixed register types.
 
\subsection{Pointer register}
Instructions with a memory operand always use an address relative to a base pointer. The base pointer can be a general purpose register, the data section pointer, the thread data pointer, the instruction pointer, or the stack pointer. The pointer is determined by the RS field. This field is interpreted as follows.
\vv
 
Single-size instructions with a memory operand (formats 0.4 - 0.9) can use any of the registers r0-r31 as base pointer. r31 is the stack pointer.
\vv
 
Larger instructions with a memory operand and an offset field of at least 16 bits (formats 2.0.x, 2.1, 2.2.x, 2.4, 2.9, 3.0.x, 3.2.x) can use the same registers, except r28 - r30,  which are replaced by the thread pointer (THREADP), data section pointer (DATAP), and instruction pointer (IP), respectively.
\vv
 
The instruction pointer may be used for addressing data in a read-only data section. This works in the following way. The address of the end of the current instruction is used as a reference point. This is the same as the address of the next instruction. The reason for using the end of the instruction as reference point is that it makes relocation in the linker independent of the instruction length in most cases. This address is multiplied by 4 when used as a data address because the instruction pointer is addressing 32 bit word units while data pointers are addressing byte units.
\vv
 
 
\subsection{Index register}
Instruction formats with an index can use r0 - r30 as index in the RT field. 
A value of 31 in the index field means no index. The signed index is multiplied by the operand size (OS) for formats 0.6, 0.8, 2.0.2, 2.0.3, 2.0.5, 2.2.3, 3.0.3, 3.2.3; by 1 for format 2.0.1; or by -1 for format 0.5 and 2.2.2. The result is added to the address given by the base pointer.
 
\subsection{Offsets}
Offsets can be 8, 16, or 32 bits. The value is sign-extended to 64 bits. An 8-bit offset is multiplied by the operand size OS, as given by the OT field. An offset of 16 or 32 bits is not scaled. The result is added to the address given by the base pointer and the index.
\vv
 
Support for addressing modes with both index and offset is optional 
(format 2.0.1, 2.0.2, 2.0.5, 2.2.2, 2.2.4, 3.0.2, 3.0.5, 3.2.2). 
Hardware implementations where the use of two additions in the address calculation would cause timing problems may allow having an index with a offset of zero or an offset with no index (RT = 31).
\vv
 
\subsection{Limit on index}
Formats 2.0.3, 2.2.3, 3.0.3, and 3.2.3 have a 16-bit or 32-bit limit on the index register. This is useful for checking array limits. A trap is generated if the value of the index register, interpreted as unsigned, is bigger than the unsigned limit. This feature is optional.
 
\subsection{Vector length}
The vector length of memory operands is specified by r0-r30 in the RT field for formats with a vector memory operand. A value of 31 in the RT field indicates a scalar with the same length as the operand size (OS).
\vv
 
The value of the vector length register indicates the vector length in bytes (not the number of elements). If the value is bigger than the maximum vector length then the maximum vector length is used. 
If the indicated vector length is zero or negative then the resulting vector will be empty and nothing will be read or written.
\vv
 
The vector length must be a multiple of the operand size OS, as indicated by the OT field. If the vector length is not a multiple of the operand size then the partial vector element will be zero.
\vv
 
The vector length for source operands in vector registers is stored in the register itself.
 
\subsection{Combining vectors with different lengths}
The length of the destination register of a vector instruction will be the same as the vector length of the first source operand.
\vv
 
A consequence of this is that the length of the result is determined by the order of the operands when vectors of different lengths are combined.
\vv
 
If the source operands have different lengths then the lengths will be adjusted as follows. If a vector source operand is too long then the extra elements will be ignored. If a vector source operand is too short then the missing elements will be zero.
\vv
 
A scalar memory operand is not broadcast but treated as a short vector. It is padded with zeroes to the vector length of the destination.
\vv
 
A broadcast memory operand will use the vector length given by the vector length register. If this is less than the length of the destination then it is padded with zeroes.
\vv
 
An immediate operand will be broadcast to the vector length of the destination.
\vv
 
\subsection{Immediate constants}
Immediate constants can be 8, 16, 32, and 64 bits. Immediate fields are aligned to natural addresses. They are interpreted as follows.
\vv
 
If OT specifies an integer type then the field is interpreted as an integer. If the field is smaller than the operand size then it is sign-extended to the appropriate size. If the field is larger than the operand size then the superfluous upper bits are ignored. The truncation of a too large immediate operand will not trigger any overflow condition.
\vv
 
If OT specifies a floating point type then the field is interpreted as follows. Immediate fields of 8 bits are interpreted as signed integers and converted to floating point numbers of the desired precision. A 16-bit field is interpreted as a half precision floating point number (subnormal numbers are supported for float16). 
A 32-bit field is interpreted as a single precision floating point number. It is converted to the desired precision if necessary. A 64-bit field is interpreted as a double precision floating point number. A 64-bit field is not allowed with a single precision operand type.
\vv
 
Some instruction formats allow immediate integer constants with a left shift. Large integer constants with a limited number of significant bits can be represented with fewer bits in this way.
 
Format 2.0.7 and 2.2.7 allow a 16-bit immediate constant in IM2 to be shifted left by the unsigned value of IM3 to give a 64-bit signed value, except for instructions that use IM3 for other purposes.
 
Format 3.0.7 and 3.2.7 allow a 32-bit immediate constant in IM4 to be shifted left by the unsigned value of IM2.
Any overflow beyond 64 bits is ignored.
 
Some single-format instructions also use shifted constants.
\vv
 
An instruction can be made compact by using the smallest size that fits the actual value of the constant.
 
 
\subsection{Mask register and fallback register} \label{MaskAndFallback}
The 3-bit mask field in formats with templates of type A or E indicates a mask register. Register r0-r6 can be used as masks if the destination is a general purpose register. Vector register v0-v6 can be used as masks if the destination is a vector register. A value of 7 in the mask field means no mask and unconditional execution using the options specified in the numeric control register.
\vv
 
If the mask is a vector register then it is interpreted as a vector with the same element size as indicated by the OT field. Each element of the mask register is applied to the corresponding element of the result.
\vv
 
The mask has multiple purposes. The primary purpose is for conditional execution. An instruction is not executed if bit 0 of the mask is zero. In this case, the destination will get a fallback value instead of the result of the calculation, and any numerical error condition will be suppressed. Vector instructions are executed conditionally for each vector element separately, so that each vector element is enabled if bit 0 of the corresponding vector element of the mask register is 1.
\vv
 
The fallback value is taken from an extra register if the instruction has less than three source operands and the format has a vacant register field, or from the first source register operand otherwise. The fallback cannot be different from the first source register if the instruction has three source operands, even if there is a vacant register field. 
If the instruction format has more than one vacant register field, then the field that would be used for the first source operand if the instruction had three source operands is used for the fallback register.
\vv
 
Register r31 (stack pointer) and v31 cannot be used as fallback register. Instead, the fallback value will be zero if a register number of 31 is indicated. 
Register r31 and v31 should not be used as first source register if it is also used as feedback because this would cause ambiguity about the fallback value. (The fallback value will not be zero in this case).
\vv
 
A memory write has no fallback register. Instead, the value of the memory operand will be unchanged if the mask has a zero in bit 0.
\vv
 
The remaining bits of the mask are used for specifying various options.
The meanings of these mask bits are described in the next section.
 
\section{Coding of masks}
A mask register can be a general purpose register r0-r6 or a vector register v0-v6. A value of 7 in the mask field means no mask.
\vv
 
The bits in the mask register are coded as follows.
 
\begin{longtable}
{|p{15mm}|p{90mm}|}
\caption{Bits in mask register and numeric control register}
\label{table:maskBits}
\endfirsthead
\endhead
\hline
\bfseries Bit number & \bfseries Meaning \\
 \hline
0 & Predicate or mask. The operation is executed only if this bit is one.\\
1 & Guaranteed to be ignored. \\
\hline
2-7 & Numerical exception control. See page \pageref{table:FPExceptionResults}. \\
2 & Floating point division by zero generates NAN \\
3 & Floating point overflow generates NAN \\
4 & Floating point underflow generates NAN \\
5 & Floating point inexact generates NAN \\
  & Bits 2-7 may also be used for controlling integer division by zero and integer overflow \\ \hline
10-12 & Floating point rounding mode: \newline
000 = nearest or even \newline
001 = down \newline
010 = up \newline
011 = towards zero \newline
This feature is optional.\\ \hline
13 & Support subnormal numbers in single and higher precision. 
(Subnormal numbers are always supported for half precision). This feature is optional.\\ \hline  
 
18-23 & Instruction-specific option bits.\\
\hline
26 - 30 & Possible use for enabling numerical traps. Not used in the standard version. \\
\hline
31 & Constant execution time. This bit makes instructions take the same number of clock cycles regardless of the values of mask and operands. The guarantee provided by this bit is useful for cryptographic applications. This feature is optional. \\
\hline
\end{longtable}
 
Bits 8, 16, 24, etc. in a vector mask register can be used like bit 0 for 8-bit and 16-bit operand sizes. All other bits are reserved for future use.
\vv
 
Vector instructions treat the mask register as a vector with the same element size (OS) as the operands. Each element of the mask vector has the bit codes as listed above. The different vector elements can have different mask bits.
\vv
 
The numeric control register (NUMCONTR) is used as mask when the mask field is 7 or absent. The NUMCONTR register is broadcast to all elements of a vector, using as many bits of NUMCONTR as indicated by the operand size, when an instruction has no mask register. The number of bits in NUMCONTR is implementation dependent (usually 16 or more). Any missing bits will be zero. 
The same NUMCONTR value is applied to all vector elements. 
Bit 0 of NUMCONTR is always 1.
\vv
 
The instruction-specific option bits (bit 18-23) may be used for various options in specific instructions. The option bits in the mask are considered zero in vector operands with an 8-bit or 16-bit operand type because each mask element has too few bits in this case. 
\vv
 
\section{Format for jump, call and branch instructions}
Most branches in software are based on the result of an arithmetic or logic instruction (ALU). The ForwardCom design combines the ALU instruction and the conditional jump into a single instruction. For example, a loop control can be implemented with a single instruction that counts down and jumps until it reaches zero or counts up until it reaches a certain limit.
\vv
 
The jumps, calls, branches, and multiway branches will use the following formats.
 
\begin{longtable}
{|p{10mm}|p{8mm}|p{8mm}|p{8mm}|p{8mm}|p{70mm}|}
\caption{List of formats for control transfer instructions}
\label{table:jumpInstructionFormats}
\endfirsthead
\endhead
\hline
\bfseries Format & \bfseries IL & \bfseries Mode & \bfseries OP1 & \bfseries Tem-plate & \bfseries Description \\
\hline
1.6 A & 1 & 6 & OPJ & B & Multiway jump and calls with three register operands.  \\
\hline
1.6 B & 1 & 6 & OPJ & B & Short jump with two register operands (RD, RS) and 8 bit offset  (IM1).  \\
\hline
1.7 C & 1 & 7 & OPJ & C & Short jump with one register operand (RD), an 8-bit immediate constant (IM2) and 8 bit offset (IM1). \\
\hline
1.7 D & 1 & 7 & 0-15 & D & Jump or call with 24-bit offset. \\
\hline
 
2.5.0 & 2 & 5 & 3 & A2 & Double size jump with three register operands (RD, RS, RT), 
and a 24-bit address offset (IM2). OPJ in upper 8 bits of IM2. \\
\hline
 
2.5.1 & 2 & 5 & 1 & B2 & Double size jump with a register destination operand, a register source operand, a 16-bit immediate operand (IM2 lower half), and 
a 16-bit jump offset (IM2 upper half). OPJ in IM1. \\
\hline
 
2.5.2 & 2 & 5 & 2 & B2 & Double size jump with one register operand (RD), 
a memory operand with base RS and 16-bit address offset (IM2 lower half), 
and a 16-bit jump offset (IM2 upper half). OPJ in IM1. Optional. \\
\hline
 
2.5.4 & 2 & 5 & 4 & C2 & Double size jump with one register operand (RD), one 8-bit immediate constant (IM2) and 32 bit offset (IM3). OPJ in IM1. \\
\hline
 
2.5.5 & 2 & 5 & 5 & C2 & Double size jump with one register operand (RD), an 8-bit offset (IM2) and a 32-bit immediate constant (IM3).  OPJ in IM1. \\
\hline
 
2.5.7 & 2 & 5 & 7 & C2 & Double size system call, 16 bit constant (IM1,IM2) and 32-bit constant (IM3). No OPJ. \\
\hline
 
3.1.0 & 3 & 1 & 0 & A3 & Triple size jump with two register operands (RD, RT), 
a 24-bit jump offset (IM2), and a memory operand with base RS and 32-bit address offset (IM3). OPJ in last byte of IM2. Optional. \\
\hline
 
3.1.1 & 3 & 1 & 1 & B3 & Triple size jump with a register destination operand, a register source operand (RS), a 32-bit jump offset (IM2), and a 32-bit immediate operand (IM3). OPJ in IM1. Optional. \\
\hline
 
 
\end{longtable}
 
The jump, call, and branch instructions have signed offsets of 8, 16, 24, or 32 bits relative to the instruction pointer. Or, more precisely, relative to the end of the instruction. This offset is multiplied by the instruction word size (= 4 bytes) to cover an address range of $\pm$ 512 bytes for short conditional jumps with 8 bits offset, $\pm$ 128 kilobytes for jumps and calls with 16 bits offset, $\pm$ 32 megabytes for 24 bits offset, and $\pm$ 8 gigabytes for 32 bits offsets.
\vv
 
The OPJ field defines the operation and jump condition. This field has 6 bits in the single size version and 8 bits in the longer format versions. The two extra bits in the longer versions are reserved for future use.
\vv
 
The versions with template C and C2 have no OT field. The operand type is 32-bit integer when there is no OT field, unless otherwise noted. It is not possible to use formats with template C or C2 with other operand types. 
\vv
 
The instructions will use vector registers when there is an OT field and M = 1. In other words, the combined ALU-and-branch instructions will use vector registers only when a floating point type is specified (or 128-bit integer type, if supported). General purpose registers are used in all other cases. Only the first element of a vector register is used. 
Logical instructions will interpret the value in a vector register as an integer, when a floating point type is specified. Only the compare instructions interpret the operands as floating point when a floating point type is indicated. Branch instructions with addition and subtraction cannot use floating point operands. The codes that these instructions would use are used for floating point compare instructions instead.
\vv
 
The combined ALU and conditional jump instructions can be coded in the formats listed above. Subtraction with a constant cannot be coded in format 1.7 C. The assembler will replace subtraction with a small immediate constant by addition with the negative constant. The code space that would have been used by subtraction in format 1.7 C is instead used for coding direct jump and call instructions with a 24-bit offset using format 1.7 D, where the lower three bits of OP1 are used as part of the 24-bit offset.
\vv
 
Unconditional and indirect jumps and calls use the formats indicated above, where unused fields must be zero. Bit 0 of the OPJ field is zero for unconditional jump instructions and one for call instructions.
\vv
 
See page \pageref{table:controlTransferInstructions} for a list of OPJ condition codes.
\vv
 
 
\section{Assignment of opcodes}
The opcodes and formats for new instructions can be assigned according to the following rules.
 
\begin{itemize}
\item Multi-format instructions. Often-used instructions that need to support many different operand types, addressing modes, and formats use all or most of the following formats: 0.0 - 0.9, 2.0.x, 2.1, 2.2.x, 2.3, 2.4, 2.8, 3.0.x, 3.2.x, 3.3, and 3.8. The same value of OP1 is used in all these formats. OP2 must be 0, except in formats 2.0.5 and 2.2.5 that use OP2 for other purposes. Instructions with few source operands should have the lowest values of OP1. Available OP1 values is a limited resource that should be economized. Instructions for integers only and instructions for floating point only may share the same OP1 value.
 
\item Control transfer instructions, i. e. jumps, branches, calls and returns, can be coded as short instructions with IL = 1, mode = 6 - 7, and OP1 = 0 - 63 or as double-size instructions with IL = 2, mode = 5, OP1 = 0 - 7, and optionally as triple-size instructions with IL = 3, mode = 1, OP1 = 0-7. See page \pageref{table:jumpInstructionFormats}.
 
\item Short single-format instructions with general purpose registers. Use mode 1.0, 1.1, and 1.8, with any value of OP1. Mode 1.0 is currently unused and may be reserved for future purposes.
 
\item Short single-format instructions with vector registers. Use mode 1.2, 1.3, and 1.4 with any value of OP1.
 
\item Double-size single-format instructions with general purpose registers can use mode 2.9 with any value of OP1, and mode 2.0.x (except 2.0.5) with any value of OP1 and OP2 $\neq$ 0 (give similar instructions the same value of OP2). If more combinations are needed then use IM3 for further subdivision of the code space.
 
\item Double-size single-format instructions with vector registers can use mode 2.6 with with any value of OP1, and mode 2.2.x (except 2.2.5) with any value of OP1 and OP2 $\neq$ 0 (give similar instructions the same value of OP2). If more combinations are needed then use IM3 for further subdivision of the code space.
 
\item Double-size single-format instructions with mixed vector and general purpose registers or with memory operands can use mode 2.5 with OP1 in the range 8-63.
 
\item Triple-size single-format instructions with general purpose registers can use mode 3.0.x with with any value of OP1 and OP2 $\neq$ 0.
 
\item Triple-size single-format instructions with vector registers can use mode 3.2.x with with any value of OP1 and OP2 $\neq$ 0.
 
\item Triple-size single-format instructions with mixed register types can use mode 3.1 with with OP1 in the range 8-63.
 
\item Possible future instructions longer than three 32-bit words should be coded with IL = 3, mode = 4-7.
 
\item New options or other modifications to existing instructions can use IM3 bits in template E or mask register bits.
 
\item New addressing modes and formats may be implemented as single-format read and write instructions. Template E formats use Mode2 for distinguishing between different formats. 
Other single-format templates may be divided into groups of eight consecutive OP1 values with the same format.
New addressing modes or other formats that apply to all multi-format instructions can use vacant values of Mode2 with E templates.
 
\item Format 1.0 is intended for single-format instructions with three general purpose registers. There are currently no such instructions. Therefore, format 1.0 A or B may be used for application-specific single-size instructions or for other purposes. Note that the M bit is not available in format 1.0 because this bit is used for distinguishing format 1.8 from 1.0. This means that format 1.0 cannot be used for vector instructions without violating the general coding scheme.
 
\item Format 1.5 is vacant to use for single-format instructions with vector registers.
 
\end{itemize}
 
Application-specific instructions may preferably use E template formats with OP2 $\neq$ 0. There are many vacant opcodes in these formats. General multi-purpose instructions may use some of the more crowded formats.
\vv
 
Unused register fields may have the same value as the first source register operand in order to avoid false dependences. Unused mask fields have the value 7 in instructions that can have a mask.
All other unused fields must be zero. The instructions with the fewest input operands should preferably have the lowest OP1 codes. 
\vv
 
The file forwardcom\_sourcecode\_documentation has a checklist of what to do when making or modifying instructions.
\vv
 
 
\end{document}