Geode instruction set: Difference between revisions

From OLPC
Jump to navigation Jump to search
mNo edit summary
Line 2: Line 2:


==Overview==
==Overview==
Generally speaking, the '''Geode''' processor supports the '''i686''' (Pentium Pro) instruction set plus '''MMX''', the '''parts of SSE''' that do not involve SSE registers or prefetch, and '''3DNow! Enhanced'''. It also supports a couple Geode-specific instructions and a few SSE2 instructions.
Generally speaking, the '''Geode''' processor supports the '''i686''' (Pentium Pro) instruction set plus '''MMX''', the '''parts of SSE''' that do not involve SSE registers, and '''3DNow! Enhanced'''. It also supports a couple Geode-specific instructions and a few SSE2 instructions.


There are also numerous kernel-only instructions and MSRs for dealing with an extra debug mode, for directly loading and saving the normally-hidden segment state (base, limit, etc.) and so on.
There are also numerous kernel-only instructions and MSRs for dealing with an extra debug mode, for directly loading and saving the normally-hidden segment state (base, limit, etc.) and so on.

Revision as of 23:41, 2 September 2007

See Hardware specification, Geode, and AMD's Geode LX data book.

Overview

Generally speaking, the Geode processor supports the i686 (Pentium Pro) instruction set plus MMX, the parts of SSE that do not involve SSE registers, and 3DNow! Enhanced. It also supports a couple Geode-specific instructions and a few SSE2 instructions.

There are also numerous kernel-only instructions and MSRs for dealing with an extra debug mode, for directly loading and saving the normally-hidden segment state (base, limit, etc.) and so on.

Geode-specific 3DNow! instructions

PFRCPV
Opcode "0F 0F / 86" is "PFRCPV xr,xr/m64" or "PFRCPV Pq,Qq". It performs floating-point reciprocal approximation for a pair of 32-bit floats in an MMX register. It is like the PFRCP instruction, except that it operates on a pair of values.
PFRSQRTV
Opcode "0F 0F / 86" is "PFRSQRTV xr,xr/m64" or "PFRSQRTV Pq,Qq". It performs floating-point reciprocal square root approximation for a pair of 32-bit floats in an MMX register. It is like the PFRSQRT instruction, except that it operates on a pair of values.

useful "undocumented" instructions

INT1
Opcode "F1" is "INT1", a 1-byte version of the "INT 1" ("CD 01") instruction. This instruction is also known as ICEBP, and is in fact available (but mostly undocumented) on all x86 processors since the 80386.
SALC
Opcode "D6" is "SALC", documented as "SETALC" with an incorrect description. This instruction sets all bits of AL to the carry flag. As with INT1, this instruction is available on all processors.

supported SSE1

PAVGB
packed average of unsigned byte, a=(a+b+1)/2 for 8 bytes
PAVGW
packed average of unsigned word, a=(a+b+1)/2 for 4 words
MASKMOVQ
cache-bypassing MMX store with byte mask (uses edi as destination)
MOVNTQ
cache-bypassing MMX store
PEXTRW
extract word into integer register, copies any word from an MMX register into a regular register
PINSRW
insert word, copies any word from memory or an integer register into an MMX register
PMAXSW
signed max() on MMX words
PMAXUB
unsigned max() on MMX bytes
PMINSW
signed min() on MMX words
PMINUB
unsigned min() on MMX bytes
PMOVMSKB
move byte mask (MSB of MMX bytes) to integer register
PMULHUW
packed multiply high unsigned word (useful for division)
PREFETCHNTA
prefetch using NTA register
PREFETCH0
prefetch using T0 register
PREFETCH1
prefetch using T1 register
PREFETCH2
prefetch using T2 register
PSADBW
packed sum of absolute byte differences
PSHUFW
packed shuffle word
SFENCE
store fence (see mfence)

unsupported SSE1

ADDPS ADDSS ANDNPS ANDPS CMPPS CMPSS COMISS CVTPI2PS CVTPS2PI CVTSI2SS CVTSS2SI CVTTPS2PI CVTTSS2SI DIVPS DIVSS LDMXCSR MAXPS MAXSS MINPS MINSS MOVAPS MOVHLPS MOVHPS MOVLHPS MOVLPS MOVNTPS MOVSS MOVUPS MULPS MULSS ORPS RCPPS RCPSS RSQRTPS RSQRTSS SHUFPS SQRTPS SQRTSS STMXCSR SUBPS SUBSS UCOMISS UNPCKHPS UNPCKLPS XORPS

supported SSE2

clflush
unprivileged cache line write back and invalidate (use "+m" for inline asm)
lfence
load fence (see mfence)
mfence
any load or store before the mfence is globally ordered with respect to any load or store after the mfence. This is needed when non-temporal operations are in use.
pause
really just a "rep nop", to give other CPUs a chance

Other