2 INTEL P6 MPN SUBROUTINES
6 This directory contains code optimized for Intel P6 class CPUs, meaning
7 PentiumPro, Pentium II and Pentium III. The mmx and p3mmx subdirectories
8 have routines using MMX instructions.
14 Times for the loops, with all code and data in L1 cache, are as follows.
15 Some of these might be able to be improved.
29 mpn_addmul/submul_1 6.35
33 mpn_mul_basecase 8.2 cycles/crossproduct (approx)
34 mpn_sqr_basecase 4.0 cycles/crossproduct (approx)
35 or 7.75 cycles/triangleproduct (approx)
37 Pentium II and III have MMX and get the following improvements.
39 mpn_divrem_1 25.0 integer part, 17.5 fractional part
49 Write-allocate L1 data cache means prefetching of destinations is unnecessary.
51 Mispredicted branches have a penalty of between 9 and 15 cycles, and even up
52 to 26 cycles depending how far speculative execution has gone. The 9 cycle
53 minimum penalty comes from the issue pipeline being 9 stages.
55 A copy with rep movs seems to copy 16 bytes at a time, since speeds for 4,
56 5, 6 or 7 limb operations are all the same. The 0.75 cycles/limb would be 3
57 cycles per 16 byte block.
64 Instructions in general code have been shown grouped if they can execute
65 together, which means up to three instructions with no successive
66 dependencies, and with only the first being a multiple micro-op.
68 P6 has out-of-order execution, so the groupings are really only showing
69 dependent paths where some shuffling might allow some latencies to be
77 "Intel Architecture Optimization Reference Manual", 1999, revision 001 dated
78 02/99, order number 245127 (order number 730795-001 is in the document too).
81 http://download.intel.com/design/PentiumII/manuals/245127.htm
83 "Intel Architecture Optimization Manual", 1997, order number 242816. This
84 is an older document mostly about P5 and not as good as the above.
87 http://download.intel.com/design/PentiumII/manuals/242816.htm