2 X86 CPU FAMILY MPN SUBROUTINES
5 This file has some notes on things common to all the x86 family code.
11 The x86 .asm files are BSD style x86 assembler code, first put through m4
12 for macro processing. The generic mpn/asm-defs.m4 is used, together with
13 mpn/x86/x86-defs.m4. Detailed notes are in those files.
15 The code is meant for use with GNU "gas" or a system "as". There's no
16 support for assemblers that demand Intel style, and with gas freely
17 available and easy to use that shouldn't be a problem.
23 m4 macros are used to define the parameters passed on the stack, and these
24 act like comments on what the stack frame looks like too. For example,
25 mpn_mul_1() has the following.
27 defframe(PARAM_MULTIPLIER, 16)
28 defframe(PARAM_SIZE, 12)
29 defframe(PARAM_SRC, 8)
30 defframe(PARAM_DST, 4)
32 Here PARAM_MULTIPLIER gets defined as `FRAME+16(%esp)', and the others
33 similarly. The return address is at offset 0, but there's not normally any
36 FRAME is redefined as necessary through the code so it's the number of bytes
37 pushed on the stack, and hence the offsets in the parameter macros stay
38 correct. At the start of a routine FRAME should be zero.
47 Helper macros FRAME_pushl(), FRAME_popl(), FRAME_addl_esp() and
48 FRAME_subl_esp() exist to adjust FRAME for the effect of those instructions,
49 and can be used instead of explicit definitions if preferred.
50 defframe_pushl() is a combination FRAME_pushl() and defframe().
52 There's generally some slackness in redefining FRAME. If new values aren't
53 going to get used, then the redefinitions are omitted to keep from
54 cluttering up the code. This happens for instance at the end of a routine,
55 where there might be just four register pops and then a ret, so FRAME isn't
58 Local variables and saved registers can be similarly defined, with negative
59 offsets representing stack space below the initial stack pointer. For
62 defframe(SAVE_ESI, -4)
63 defframe(SAVE_EDI, -8)
64 defframe(VAR_COUNTER,-12)
66 deflit(STACK_SPACE, 12)
68 Here STACK_SPACE gets used in a "subl $STACK_SPACE, %esp" to allocate the
69 space, and that instruction must be followed by a redefinition of FRAME
70 (setting it equal to STACK_SPACE) to reflect the change in %esp.
72 Definitions for pushed registers are only put in when they're going to be
73 used. If registers are just saved and restored with pushes and pops then
74 definitions aren't made.
80 Only addition and subtraction seem to be universally available, certainly
81 that's all the Solaris 8 "as" seems to accept. If expressions are wanted
82 then m4 eval() should be used.
84 In particular note that a "/" anywhere in a line starts a comment in Solaris
85 "as", and in some configurations of gas too.
87 addl $32/2, %eax <-- wrong
89 addl $eval(32/2), %eax <-- right
91 Binutils gas/config/tc-i386.c has a choice between "/" being a comment
92 anywhere in a line, or only at the start. FreeBSD patches 2.9.1 to select
93 the latter, and as of 2.9.5 it's the default for GNU/Linux too.
99 Solaris "as" doesn't support "#" commenting, using /* */ instead,
100 unfortunately. For that reason "C" commenting is used (see asm-defs.m4) and
101 the intermediate ".s" files have no comments.
107 In a couple of places addressing modes like 0(%ebx) with a byte-sized zero
108 displacement are wanted, rather than (%ebx) with no displacement. These are
109 either for computed jumps or to get desirable code alignment. Explicit
110 .byte sequences are used to ensure the assembler doesn't turn 0(%ebx) into
111 (%ebx). The Zdisp() macro in x86-defs.m4 is used for this.
113 Current gas 2.9.5 or recent 2.9.1 leave 0(%ebx) as written, but old gas
114 1.92.3 changes it. In general changing would be the sort of "optimization"
115 an assembler might perform, hence explicit ".byte"s are used where
120 SHLD/SHRD INSTRUCTIONS
122 The %cl count forms of double shift instructions like "shldl %cl,%eax,%ebx"
123 must be written "shldl %eax,%ebx" for some assemblers. gas takes either,
124 Solaris "as" doesn't allow %cl, gcc generates %cl for gas and NeXT (which is
125 gas), and omits %cl elsewhere.
127 For GMP an autoconf test is used to determine whether %cl should be used and
128 the macros shldl, shrdl, shldw and shrdw in mpn/x86/x86-defs.m4 then pass
129 through or omit %cl as necessary. See comments with those macros for usage.
135 The x86 calling conventions say that the direction flag should be clear at
136 function entry and exit. (See iBCS2 and SVR4 ABI books, references below.)
138 Although this has been so since the year dot, it's not absolutely clear
139 whether it's universally respected. Since it's better to be safe than
140 sorry, gmp follows glibc and does a "cld" if it depends on the direction
141 flag being clear. This happens only in a few places.
145 POSITION INDEPENDENT CODE
147 Defining the symbol PIC in m4 processing selects position independent code.
148 This mainly affects computed jumps, and these are implemented in a
149 self-contained fashion (without using the global offset table). The few
150 calls from assembly code to global functions use the normal procedure
153 PIC is necessary for ELF shared libraries because they can be mapped into
154 different processes at different virtual addresses. Text relocations in
155 shared libraries are allowed, but that presumably means a page with such a
156 relocation isn't shared. The use of the PLT for PIC adds a fixed cost to
157 every function call, which is small but might be noticeable when working with
160 Calls from one library function to another don't need to go through the PLT,
161 since of course the call instruction uses a displacement, not an absolute
162 address, and the relative locations of object files are known when libgmp.so
163 is created. "ld -Bsymbolic" (or "gcc -Wl,-Bsymbolic") will resolve calls
164 this way, so that there's no jump through the PLT, but of course leaving
165 setups of the GOT address in %ebx that may be unnecessary.
167 The %ebx setup could be avoided in assembly if a separate option controlled
168 PIC for calls as opposed to computed jumps etc. But there's only ever
169 likely to be a handful of calls out of assembler, and getting the same
170 optimization for C intra-library calls would be more important. There seems
171 no easy way to tell gcc that certain functions can be called non-PIC, and
172 unfortunately many gmp functions use the global memory allocation variables,
173 so they need the GOT anyway. Object files with no global data references
174 and only intra-library calls could go into the library as non-PIC under
175 -Bsymbolic. Integrating this into libtool and automake is left as an
176 exercise for the reader.
182 The overheads in setting up for an unrolled loop can mean that at small
183 sizes a simple loop is faster. Making small sizes go fast is important,
184 even if it adds a cycle or two to bigger sizes. To this end various
185 routines choose between a simple loop and an unrolled loop according to
186 operand size. The path to the simple loop, or to special case code for
187 small sizes, is always as fast as possible.
189 Adding a simple loop requires a conditional jump to choose between the
190 simple and unrolled code. The size of a branch misprediction penalty
191 affects whether a simple loop is worthwhile.
193 The convention is for an m4 definition UNROLL_THRESHOLD to set the crossover
194 point, with sizes < UNROLL_THRESHOLD using the simple loop, sizes >=
195 UNROLL_THRESHOLD using the unrolled loop. If position independent code adds
196 a couple of cycles to an unrolled loop setup, the threshold will vary with
197 PIC or non-PIC. Something like the following is typical.
200 deflit(UNROLL_THRESHOLD, 10)
202 deflit(UNROLL_THRESHOLD, 8)
205 There's no automated way to determine the threshold. Setting it to a small
206 value and then to a big value makes it possible to measure the simple and
207 unrolled loops each over a range of sizes, from which the crossover point
208 can be determined. Alternately, just adjust the threshold up or down until
209 there's no more speedups.
215 The x86 addressing modes allow a byte displacement of -128 to +127, making
216 it possible to access 256 bytes, which is 64 limbs, without adjusting
217 pointer registers within the loop. Dword sized displacements can be used
218 too, but they increase code size, and unrolling to 64 ought to be enough.
220 When unrolling to the full 64 limbs/loop, the limb at the top of the loop
221 will have a displacement of -128, so pointers have to have a corresponding
222 +128 added before entering the loop. When unrolling to 32 limbs/loop
223 displacements 0 to 127 can be used with 0 at the top of the loop and no
224 adjustment needed to the pointers.
226 Where 64 limbs/loop is supported, the +128 adjustment is done only when 64
227 limbs/loop is selected. Usually the gain in speed using 64 instead of 32 or
228 16 is small, so support for 64 limbs/loop is generally only for comparison.
234 When working from least significant limb to most significant limb (most
235 routines) the computed jump and pointer calculations in preparation for an
236 unrolled loop are as follows.
238 S = operand size in limbs
239 N = number of limbs per loop (UNROLL_COUNT)
240 L = log2 of unrolling (UNROLL_LOG2)
241 M = mask for unrolling (UNROLL_MASK)
242 C = code bytes per limb in the loop
243 B = bytes per limb (4 for x86)
245 computed jump (-S & M) * C + entrypoint
246 subtract from pointers (-S & M) * B
247 initial loop counter (S-1) >> L
248 displacements 0 to B*(N-1)
250 The loop counter is decremented at the end of each loop, and the looping
251 stops when the decrement takes the counter to -1. The displacements are for
252 the addressing accessing each limb, eg. a load with "movl disp(%ebx), %eax".
254 Usually the multiply by "C" can be handled without an imul, using instead an
255 leal, or a shift and subtract.
257 When working from most significant to least significant limb (eg. mpn_lshift
258 and mpn_copyd), the calculations change as follows.
260 add to pointers (-S & M) * B
261 displacements 0 to -B*(N-1)
267 This version comes with FreeBSD 2.2.8 and has a couple of gremlins that
270 Firstly, an expression involving two forward references to labels comes out
271 as zero. For example,
278 This should lead to "addl $1, %eax", but it comes out as "addl $0, %eax".
279 When only one forward reference is involved, it works correctly, as for
287 Secondly, an expression involving two labels can't be used as the
288 displacement for an leal. For example,
293 leal bar-foo(%eax,%ebx,8), %ecx
295 A slightly cryptic error is given, "Unimplemented segment type 0 in
296 parse_operand". When only one label is used it's ok, and the label can be a
297 forward reference too, as for example,
299 leal foo(%eax,%ebx,8), %ecx
303 These problems only affect PIC computed jump calculations. The workarounds
304 are just to do an leal without a displacement and then an addl, and to make
305 sure the code is placed so that there's at most one forward reference in the
312 "Intel Architecture Software Developer's Manual", volumes 1 to 3, 1999,
313 order numbers 243190, 243191 and 243192. Available on-line,
315 ftp://download.intel.com/design/PentiumII/manuals/243190.htm
316 ftp://download.intel.com/design/PentiumII/manuals/243191.htm
317 ftp://download.intel.com/design/PentiumII/manuals/243192.htm
319 "Intel386 Family Binary Compatibility Specification 2", Intel Corporation,
320 published by McGraw-Hill, 1991, ISBN 0-07-031219-2.
322 "System V Application Binary Interface", Unix System Laboratories Inc, 1992,
323 published by Prentice Hall, ISBN 0-13-880410-9. And the "Intel386 Processor
324 Supplement", AT&T, 1991, ISBN 0-13-877689-X. (These have details of ELF
325 shared library PIC coding.)