Flush-to-zero mode | ARM VERSION 1.2 instruction

Vector Floating-point Programming

6.6Flush-to-zero mode

Some implementations of VFP use support code to handle denormalized numbers. The performance of such systems, in calculations involving denormalized numbers, is much less than it is in normal calculations.

Flush-to-zero mode replaces denormalized numbers with +0. This does not comply with IEEE 754 arithmetic, but in some circumstances can improve performance considerably.

6.6.1When to use flush-to-zero mode

You should select flush-to-zero mode if all the following are true:

•IEEE 754 compliance is not a requirement for your system

•the algorithms you are using are such that they sometimes generate denormalized numbers

•your system uses support code to handle denormalized numbers

•the algorithms you are using do not depend for their accuracy on the preservation of denormalized numbers

•the algorithms you are using do not generate frequent exceptions as a result of replacing denormalized numbers with +0.

You can change between flush-to-zero and normal mode at any time, if different parts of your code have different requirements. Numbers already in registers are not affected by changing mode.

6.6.2The effects of using flush-to-zero mode

With certain exceptions (see Operations not affected by flush-to-zero mode on

page 6-14), flush-to-zero mode has the following effects on floating-point operations:

•A denormalized number is treated as +0 when used as an input to a floating point operation. The source register is not altered.

•If the result of a single-precision floating-point operation, before rounding, is in the range –2–126to +2–126, it is replaced by +0.

•If the result of a double-precision floating-point operation, before rounding, is in the range –2–1022to +2–1022, it is replaced by +0.

An inexact exception occurs whenever a denormalized number is used as an operand, or a result is flushed to zero. Underflow exceptions do not occur in flush-to-zero mode.

ARM DUI 0068B

Copyright © 2000, 2001 ARM Limited. All rights reserved.

6-13

Image 251

ARM VERSION 1.2 manual Flush-to-zero mode, When to use flush-to-zero mode, Effects of using flush-to-zero mode

Contents

ARM Developer Suite Copyright 2000, 2001 ARM Limited. All rights reserved ARM Developer SuiteAssembler Guide ARM Developer Suite Assembler Guide Glossary Feedback on About this book on This book is organized into the following chapters Using this bookAbout this book Intended audience ARM publications Typographical conventionsItalic Further reading Other publications ARM Reference Peripheral Specification ARM DDI Feedback on this book FeedbackFeedback on the ARM Developer Suite ARM DUI 0068B About the ARM Developer Suite assemblers on Introduction ARM Developer Suite ADS has About the ARM Developer Suite assemblers Writing ARM and Thumb Assembly Language Or run the module in AXD with interleaving on See IntroductionCode examples ARM processors always start executing code in ARM state Overview of the ARM architectureArchitecture versions ARM and Thumb state General-purpose, 32-bit registers Processor modeRegisters Program counter pc Data processing instructions ARM instruction set overviewBranch instructions Semaphore instructions Status register access instructionsSingle register load and store instructions Multiple register load and store instructions Following general points apply to ARM instructions Register accessAccess to the inline barrel shifter ARM instruction capabilities Following general points apply to Thumb instructions Thumb instruction set overviewThumb instruction capabilities Differences between Thumb and ARM instruction sets Access to the barrel shifter Single register load and store instructions Case rules Structure of assembly language modulesLayout of assembly language source files General form of source lines in assembly language is Comments LabelsLocal labels Is a base between 2 Xxx is a number in that base Constants ELF sections and the Area directive An example ARM assembly language module END directive Entry directiveApplication execution Application termination Calling subroutines BX instruction An example Thumb assembly language moduleCODE32 and CODE16 directives Example 2-4 Preprocessing an assembly language source file Using the C preprocessor ALU status flags Conditional execution Examples Execution conditions Example of the use of conditional execution Using conditional execution in ARM state Conditional branches only Branch prediction and caches Converting to Thumb Loading constants into registers Right, 2 bits Direct loading with MOV and MVN Loading with LDR Rd, =const Direct loading with MOV in Thumb state Placing literal pools Loading floating-point constants Direct loading with ADR and Adrl Loading addresses into registers ADR Implementing a jump table with ADR Example 2-7 ARM code jump table Example 2-8 Thumb code jump table Loading addresses with LDR Rd, = label R1, =Darea + = LDR R1,PC, #offset into Literal Pool Increments r1 by An LDR Rd, =label example string copyingAddress registers. For example, the instruction Load register Load and store multiple register instructions Where ARM LDM and STM instructionsSyntax Syntax of the LDM instructions is LDM and STM addressing modes Usage Descending or ascending Implementing stacks with LDM and STM Stacking registers for nested subroutines Example 2-11 Block copy Block copy with LDM and STM Movs Thumb-state block copy example Thumb LDM and STM instructionsLDM and STM Push and POP LSR After substitution this becomes Using macrosTest-and-branch macro example This macro can be invoked as follows If only the remainder is required Unsigned integer division macro exampleRegister that holds the divisor After the instructions are executed, it holds the remainder Ratio DivMod r0,r5,r4,r2 Describing data structures with MAP and Field directives Relative maps Register-based maps Program-relative maps Finding the end of the allocated data Forcing correct alignment EndOfChars Defining register-based symbols Using register-based MAP and Field directives Is equivalent to the C code Setting up a C-type structure 2-27 for an explanation of these Making faster access possibleIf you want the equivalent of the C code Not This example, the MAP directive is Field MAP Using two register-based structures Avoiding problems with MAP and Field directives ArrayBase RN r9 Using frame directives Expressions, literals, and operators on Symbols on Command syntax Position-independent. The default is /norwpi Specifies that the content of inputfile is read-onlyPosition-independent. The default is /noropi Specifies that the content of inputfile is read-write Selects no floating-point option. This makes your assembled Assembled for the wrong target FPUSelection of libraries, accordingly Valid options are Allow unaligned LDRs Command-line optionsSets the maximum source cache size to n. The default is 8MB Object file, for use by the debugger see Keep on Turns off warning messages First pass and reads them from memory on the second passAs \n and \t Register names Assembler Reference Format of source lines Predeclared floating-point register names Predefined register and coprocessor namesPredeclared register names Predeclared program status register names Expressions or conditions, for example Built-in variablesLists the built-in variables defined by the ARM assembler Determining the armasm version at assembly time Labels on SymbolsSymbol naming rules Numeric constants on Numeric constants Variables Gbls Assembly time substitution of variables Register-relative labels Program-relative labelsDCD and Dcdu on Dcfd and Dcfdu on Dcfs and Dcfsu on DCQ and Dcqu on DCW and Dcwu on Syntax of a reference to a local label is Syntax of a local label is Assembler Reference This section contains the following subsections Expressions, literals, and operators String literals String expressionsString literals Example Numeric expressions Numeric code of the character Is a sequence of characters using only the digits 0 to nNumeric literals Numeric literals can take any of the following forms Floating-point literals can take any of the following forms Floating-point literals There are only two logical literals Register-relative and program-relative expressionsLogical expressions Logical literals String manipulation Operator precedence Operator precedence in C Precedence Operator Usage Description Unary operators Example of use of SBOFFSET1912 and SBOFFSET11 String manipulation operators Binary operatorsMultiplicative operators Addition, subtraction, and logical operators Shift operatorsSHR is a logical shift and does not propagate the sign bit Relational operators 10 shows the Boolean operators Boolean operators ARM DUI 0068B ARM Instruction Reference Add with carry, Add All Logical Branch Move not All ARM condition codes Q flag LDR and STR, doublewords on ARM memory access instructionsLDR and STR, words and unsigned bytes on LDR and STR, halfwords and signed bytes on Otherwise, a 32-bit word is transferred Where Is either LDR Load Register or STR Store RegisterLDR and STR, words and unsigned bytes Post-indexed offset Zero offsetPre-indexed offset Program-relative Flexible offset syntax Loading to r15 Address alignment for word transfers Architectures Saving from r15 LDR and STR, halfwords and signed bytes Is often a numeric constant see examples below Must be within ±255 bytes of the current instructionIs an offset applied to the value in Rn see Offset syntax Offset syntax Incorrect example Offset syntax is the same for LDR and STR, doublewords onAddress alignment for halfword transfers You cannot load halfwords or bytes to r15 Pre-indexed without writeback LDR and STR, doublewordsIs an optional condition code see Conditional execution on Must be an even numbered register, and not r14 Address alignment Not be the same as Rd or Rd+1 Incorrect examples Increment address before each transfer Is either LDM or STMIs any one of the following Increment address after each transfer Loading or storing the base register, with writeback Non word-aligned addresses Is the register on which the memory address is based 5 PLD Alignment Both Rd and Rm Is swapped with the contents of the memory location6 SWP ARM general data processing instructions Flexible second operand Bits of the register are set to LSR and LSL Instruction substitution Carry flag Is the ARM register holding the first operand 2 ADD, SUB, RSB, ADC, SBC, and RSCIs one of ADD, SUB, RSB, ADC, SBC, or RSC Is the ARM register for the result Multiword arithmetic examples Condition flagsUse of r15 These instructions subtract one 96-bit integer from another Is one of AND, ORR, EOR, or BIC 3 AND, ORR, EOR, and BICLogical AND, OR, Exclusive or and Bit Clear Orreq Is the ARM register for the result MOV and MVNMove and Move Not Mvnne Compare and Compare Negative CMP and CMN CMN Test and Test Equivalence TST and TEQ TEQ Is the operand register 7 CLZCount Leading Zeroes Is the ARM register for the result. Rd must not be r15 UMULL, UMLAL, Smull and Smlal on ARM multiply instructionsMUL and MLA on MUL and MLA MUL Is one of UMULL, UMLAL, SMULL, or Smlal UMULL, UMLAL, Smull and Smlal Umull Are the ARM registers holding the values to be multiplied SMULxyUse the top end bits 3116 of Rm Use the top end bits 3116 of Rs Smulbt Is the ARM register holding the value to be added SMLAxy Smlatb Are the ARM registers holding the operands SMULWyUse the top end bits 3116 of Rs Is the ARM register holding the value to be added SMLAWyAre the ARM registers holding the values to be multiplied Smlawt SMLALxy Smlaltt R15 cannot be used for either Rm or Rs 8 MIA, MIAPH, and MIAxyCurrent processors Use the top end bits 3116 of Rm These instructions are only available in XScale Are the ARM registers holding the operands ARM saturating arithmetic instructionsQADD, QSUB, QDADD, and Qdsub Is one of QADD, QSUB, QDADD, or Qdsub Qadd ARM branch instructions Branch, and Branch with Link 2 BX 3 BLX Blxmi MCR, MCR2, Mcrr on ARM coprocessor instructions 1 CDP, CDP2 Are ARM source registers. They must not be r15 Is p n, where n is an integer in the range2 MCR, MCR2, Mcrr Affected 3 MRC, MRC2 Mrrc Is the coprocessor register to load or save 5 LDC, STCIs either LDC or STC Is an optional suffix specifying a long transfer Architectures Is either LDC2 or STC2 6 LDC2, STC2 Architectures Miscellaneous ARM instructions Software interrupt 1 SWI Is either Cpsr or Spsr 2 MRSWhere Is the destination register. Rd must not be r15 See MRS on 3 MSRIs either Cpsr or Spsr MSR CPSRf, r5 Breakpoint Bkpt Are general-purpose registers 5 MAR, MRAFor current processors LDR ARM pseudo-instruction on ARM pseudo-instructionsADR ARM pseudo-instructionon Adrl ARM pseudo-instructionon Non word-aligned address within ±255 bytes ADR ARM pseudo-instructionIs an optional condition code Is the register to load Word-aligned address within 256KB Adrl ARM pseudo-instructionNon word-aligned address within 64KB R4,start + = ADD R4,pc,#0xe800 LDR ARM pseudo-instruction R3,=0xff0 Loads Into = MOV r3,#0xff0 NOP ARM pseudo-instruction Thumb Instruction Reference Add with carry Rotate right LDR and STR, pc or sp relative on Thumb memory access instructionsLDR and STR, immediate offset on LDR and STR, register offset on Store register Where Is eitherLDR and STR, immediate offset Load register Address alignment for word and halfword transfers LDR and STR, register offset Strsh 4 in the range 0 to LDR and STR, pc or sp relative R2,pc,#1016 POP reglist These instructions do not affect the flags POP reglist, pc Store multiple, increment after Ldmia and StmiaLoad and store multiple registers Load multiple, increment after R3!, r0,r4 ADC, SBC, and MUL on Thumb arithmetic instructionsADD and SUB, sp on ADD, pc or sp relative on 255 to +255 ADD and SUB, low registersTo +7 Restrictions These instructions update the N, Z, C, and V flags Is a register containing the second operand 2 ADD, high or low registers Range -508 to +508 ADD and SUB, sp Range 4 ADD, pc or sp relativeIs the destination register. Rd must be in the range r0- r7 Is either sp or pc 5 ADC, SBC, and MUL Where Is one of ADC, SBC, or MUL TST on page 5-30 Test bits Thumb general data processing instructionsCMP and CMN on page 5-26 Compare and Compare Negative MOV, MVN, and NEG on page 5-28 Move, Move NOT, and Negate Range r0- r7 Where Is one of AND, ORR, EOR, or BIC1 AND, ORR, EOR, and BIC Bitwise logical operations 2 ASR, LSL, LSR, and ROR Immediate shift Register-controlled shift Where Is the register containing the first operand Examples Is the source register Where Is the destination register4 MOV, MVN, and NEG Move, Move NOT, and Negate Condition flags Rn and Rm must be in the range r0-r7 Where Is the register containing the first operand5 TST Test bits Thumb branch instructions Label must be within 1 BIs an optional condition code see -2 on 3-23 for more information Condition codes for Thumb B instruction Long branch with Link 2 BL 3 BX BLX label always causes a change to ARM state 4 BLXBranch with Link, and optionally exchange instruction set Instruction clears the T flag in the CPSR. Code at Bkpt on Thumb software interrupt and breakpoint instructions Bkpt immed8 Thumb pseudo-instructions 1KB. expr must be defined locally, it cannot be imported ADR Thumb pseudo-instruction Assembler generates the instruction LDR Thumb pseudo-instructionIf the value of expr is within range of a MOV instruction, =labelname Syntax for NOP is NOP Thumb pseudo-instruction ARM DUI 0068B Vector Floating-point Programming Absolute value Vector All Negate Vector All VFP architectures Vector floating-point coprocessorReference Manual Register banks Floating-point registers Restriction on vector length VectorsVector wrap-around Vector stride Vector operations Vector and scalar operationsControl of scalar, vector and mixed operations Scalar operations VFP and condition codes Vector Floating-point Programming FPSCR, the floating-point status and control register VFP system registers 0b000 See FMRX, FMXR, and Fmstat on FPEXC, the floating-point exception registerFPSID, the floating-point system ID register Modifying individual bits of a VFP system register Effects of using flush-to-zero mode Flush-to-zero modeWhen to use flush-to-zero mode Operations not affected by flush-to-zero mode Ftosi and Ftoui on VFP instructionsFmrrs and Fmsrr on FMRX, FMXR, and Fmstat onPage Fabsd d3, d5 Fnegsmi s15, s15 Is the VFP register holding the second operand Fadd and FsubIs the VFP register for the result Is the VFP register holding the first operand Fcmp instructions can produce Invalid Operation exceptions FcmpFloating-point compare Fcmp is always scalar With zero instruction Is a single-precision VFP register holding the operand FcvtdsIs a double-precision VFP register for the result Is a double-precision VFP register holding the operand FcvtsdIs a single-precision VFP register for the result Fdiv Address used for the transfer FLD and FSTFloating-point load and store Precision specified in precision Fldsne Fldm and Fstm Following instructions are equivalent Unspecified precision Must be one of FMAC, FNMAC, FMSC, or Fnmsc FMAC, FNMAC, FMSC, and Fnmsc Fnmscsle These instructions do not produce any exceptions Fmdrr and FmrrdIs the VFP double-precision register Are ARM registers. Do not use r15 These instructions are used together as matched pairs FMDHR, FMDLR, FMRDH, and FmrdlIs the ARM register. Rd must not be r15 Is the VFP single-precision register Fmrs and Fmsr Are the ARM registers. Do not use r15 Fmrrs and FmsrrAre two consecutive VFP single-precision registers Is the ARM register FMRX, FMXR, and Fmstat Fmul and Fnmul Fsito and Fuito Is the VFP register holding the operand Fsqrt Is a single-precision VFP register for the integer result Ftosi and Ftoui Can be S for single-precision, or D for double-precision VFP pseudo-instructionThere is one VFP pseudo-instruction FLD pseudo-instruction D1,=3.12E106 Loads 3.12E106 into d1 Vfpassert Scalar on Vfpassert Vector on VFP directives and vector notation VFP directives and vector notation on Vfpassert Vector on Vfpassert Scalar VFP directives and vector notation on Vfpassert Scalar on Where Is the vector length Is the vector stride R10,FPSCR ARM DUI 0068B Conditional assembly, looping, inclusions, and macros Assembly control directives on Location of descriptions of directives Alphabetical list of directives Declare a global arithmetic, logical, or string variable Symbol definition directivesThis section describes the following directives GBLA, GBLL, and Gbls on GBLA, GBLL, and Gbls Armasm -pd objectsize Seta 0xFF -o objectfile sourcefile LCLA, LCLL, and Lcls SETA, SETL, and Sets Names on Rlist Evaluates to a coprocessor register number from 0 to 5 CNCN directive defines a name for a coprocessor register Coprocessor names on Evaluates to a coprocessor number from 0 to 6 CP DN and SN Evaluates to a floating-point register number from 0 to 8 FN DCQ and Dcqu on Data definition directivesDCD and Dcdu on Ltorg During the first pass of the assembler 2 MAPIs a numeric or program-relative expression Set to this address Storage counter FieldBy the value of expr Expressions on Space Consecutive bytes of store 5 DCBTo 255 see Numeric expressions on Quoted string. The characters of the string are loaded into DCW and Dcwu on DCQ and Dcqu on DCD and DcduNumeric expression see Numeric expressions on Program-relative expression Dcdo Dcfd and Dcfdu Dcfs and Dcfsu Is a numeric expression see Numeric expressions on 10 DCI DCD and Dcdu on DCW and Dcwu on DCQ and Dcqu Data DCW and DcwuTo 65535 see Numeric expressions on DCD and Dcdu on DCQ and Dcqu on IF, ELSE, and Endif on While and Wend on Assembly control directivesMacro and Mend on Mexit on Macro and Mend BGE Mexit Using a macro to produce assembly-time diagnostics See Relational operators on 4 IF, ELSE, and Endif Example 7-3 Assembly conditional on a variable being defined While and Wend Frame POP on Frame Push on Frame Register on Frame Restore onFrame description directives Frame Address on Can omit it Is sp unless the function uses a separate frame pointer Is the number of bytes that the stack pointer moves There are two alternative syntaxes for Frame POP There are two alternative syntaxes for Frame Push Frame Push Is the register in which the value is preserved Frame Register Frame Restore Frame Save Frame State Remember Frame State Restore on Function or Proc on Frame State Remember on Function or Proc on Frame State Restore Function or Proc Endfunc or Endp Is an assertion that can evaluate to either True or False Reporting directivesAssert Is an expression that evaluates to a string Info Specify the -listassembler option to turn on listing Where Is the OPT directive setting. -2 lists valid settings3 OPT OPT Is the subtitle TTL and SubtIs the title GET or Include on Miscellaneous directivesCODE16 and CODE32 on Export or Global on Align Cacheable, CODE, ALIGN=3 Area Example,CODE,READONLY An example code section CODE16 and CODE32 4 END Entry directive declares an entry point to a program Entry Is optional. type can be any one Is the symbolic name to assign to the value6 EQU Address, or a 32-bit integer constant Export or Global Exportas Extern Nesting directives on GET or Include Import File, or library. The symbol name is case-sensitiveGlobal See Export or Global on See GET or Include on IncbinInclude Keep Symbols are kept except register-relative symbols Require directive specifies a dependency between sections Assembly failsNofp Require REQUIRE8 and PRESERVE8 Evaluates to a register number from 0 to 19 RN Rout Platforms See ARM Developer SuiteAmong other things, computer software Thumb state Otherwise stated Used to find errors in the application program flowDefault thread runs See also Saved Processor Status Register See Read Write Position Independent See also RopiSee also Rwpi See Read Only Position Independent See Saved Processor Status Register Normally set to zero on resetBlock of software code or data for an Image By ARM to handle semihosting Index Align Gbll Adrl Symbols Index-6