[PATCH v5 3/8] x86: re-work insn/suffix recognition

[H.J. Lu]

On Fri, Oct 28, 2022 at 2:00 AM Jan Beulich <jbeulich@suse.com> wrote:
>
> On 27.10.2022 19:21, H.J. Lu wrote:
> > On Tue, Oct 25, 2022 at 12:26 AM Jan Beulich <jbeulich@suse.com> wrote:
> >> @@ -989,13 +976,13 @@ emms, 0xf77, None, CpuMMX, No_bSuf|No_wS
> >>  // copying between Reg64/Mem64 and RegXMM/RegMMX, as is mandated by Intel's
> >>  // spec). AMD's spec, having been in existence for much longer, failed to
> >>  // recognize that and specified movd for 32- and 64-bit operations.
> >> -movd, 0x666e, None, CpuAVX, D|Modrm|Vex128|Space0F|VexW0|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|SSE2AVX, { Reg32|Unspecified|BaseIndex, RegXMM }
> >> +movd, 0x666e, None, CpuAVX, D|Modrm|Vex128|Space0F|VexW0|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|SSE2AVX|Pass2, { Reg32|Unspecified|BaseIndex, RegXMM }
> >>  movd, 0x666e, None, CpuAVX|Cpu64, D|Modrm|Vex=1|Space0F|VexW1|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|Size64|SSE2AVX, { Reg64|BaseIndex, RegXMM }
> >>  movd, 0x660f6e, None, CpuSSE2, D|Modrm|IgnoreSize|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { Reg32|Unspecified|BaseIndex, RegXMM }
> >>  movd, 0x660f6e, None, CpuSSE2|Cpu64, D|Modrm|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|Size64, { Reg64|BaseIndex, RegXMM }
> >>  movd, 0xf6e, None, CpuMMX, D|Modrm|IgnoreSize|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { Reg32|Unspecified|BaseIndex, RegMMX }
> >>  movd, 0xf6e, None, CpuMMX|Cpu64, D|Modrm|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|Size64, { Reg64|BaseIndex, RegMMX }
> >> -movq, 0xf37e, None, CpuAVX, Load|Modrm|Vex=1|Space0F|VexWIG|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|SSE2AVX, { Qword|Unspecified|BaseIndex|RegXMM, RegXMM }
> >> +movq, 0xf37e, None, CpuAVX, Load|Modrm|Vex=1|Space0F|VexWIG|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|SSE2AVX|Pass2, { Qword|Unspecified|BaseIndex|RegXMM, RegXMM }
> >>  movq, 0x66d6, None, CpuAVX, Modrm|Vex=1|Space0F|VexWIG|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|SSE2AVX, { RegXMM, Qword|Unspecified|BaseIndex|RegXMM }
> >>  movq, 0x666e, None, CpuAVX|Cpu64, D|Modrm|Vex=1|Space0F|VexW1|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|Size64|SSE2AVX, { Reg64|Unspecified|BaseIndex, RegXMM }
> >>  movq, 0xf30f7e, None, CpuSSE2, Load|Modrm|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { Unspecified|Qword|BaseIndex|RegXMM, RegXMM }
> >> @@ -1159,7 +1146,7 @@ andpd<sse2>, 0x660f54, None, <sse2:cpu>,
> >>  cmp<frel>pd<sse2>, 0x660fc2, <frel:imm>, <sse2:cpu>, Modrm|<sse2:attr>|<sse2:vvvv>|<frel:comm>|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|ImmExt, { RegXMM|Unspecified|BaseIndex, RegXMM }
> >>  cmp<frel>sd<sse2>, 0xf20fc2, <frel:imm>, <sse2:cpu>, Modrm|<sse2:scal>|<sse2:vvvv>|<frel:comm>|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|ImmExt, { RegXMM|Qword|Unspecified|BaseIndex, RegXMM }
> >>  cmppd<sse2>, 0x660fc2, None, <sse2:cpu>, Modrm|<sse2:attr>|<sse2:vvvv>|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { Imm8, RegXMM|Unspecified|BaseIndex, RegXMM }
> >> -cmpsd<sse2>, 0xf20fc2, None, <sse2:cpu>, Modrm|<sse2:scal>|<sse2:vvvv>|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { Imm8, Qword|Unspecified|BaseIndex|RegXMM, RegXMM }
> >> +cmpsd<sse2>, 0xf20fc2, None, <sse2:cpu>, Modrm|<sse2:scal>|<sse2:vvvv>|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|Pass2, { Imm8, Qword|Unspecified|BaseIndex|RegXMM, RegXMM }
> >>  comisd<sse2>, 0x660f2f, None, <sse2:cpu>, Modrm|<sse2:scal>|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { Qword|Unspecified|BaseIndex|RegXMM, RegXMM }
> >>  cvtpi2pd, 0x660f2a, None, CpuSSE2, Modrm|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { RegMMX, RegXMM }
> >>  cvtpi2pd, 0xf3e6, None, CpuAVX, Modrm|Vex|Space0F|VexW0|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|SSE2AVX, { Qword|Unspecified|BaseIndex, RegXMM }
> >> @@ -1184,7 +1171,7 @@ movlpd, 0x6613, None, CpuAVX, Modrm|Vex|
> >>  movlpd, 0x660f12, None, CpuSSE2, D|Modrm|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { Qword|Unspecified|BaseIndex, RegXMM }
> >>  movmskpd<sse2>, 0x660f50, None, <sse2:cpu>, Modrm|<sse2:attr>|IgnoreSize|No_bSuf|No_wSuf|No_sSuf|No_ldSuf|NoRex64, { RegXMM, Reg32|Reg64 }
> >>  movntpd<sse2>, 0x660f2b, None, <sse2:cpu>, Modrm|<sse2:attr>|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { RegXMM, Xmmword|Unspecified|BaseIndex }
> >> -movsd, 0xf210, None, CpuAVX, D|Modrm|VexLIG|Space0F|VexW0|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|SSE2AVX, { Qword|Unspecified|BaseIndex, RegXMM }
> >> +movsd, 0xf210, None, CpuAVX, D|Modrm|VexLIG|Space0F|VexW0|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|SSE2AVX|Pass2, { Qword|Unspecified|BaseIndex, RegXMM }
> >>  movsd, 0xf210, None, CpuAVX, D|Modrm|Vex=3|Space0F|VexVVVV=1|VexW=1|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf|SSE2AVX, { RegXMM, RegXMM }
> >>  movsd, 0xf20f10, None, CpuSSE2, D|Modrm|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { Qword|Unspecified|BaseIndex|RegXMM, RegXMM }
> >>  movupd<sse2>, 0x660f10, None, <sse2:cpu>, D|Modrm|<sse2:attr>|No_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf, { RegXMM|Unspecified|BaseIndex, RegXMM }
> >>
> >
> > All these instructions with Pass2 have a vector register operand.  Can you
> > use it instead?
>
> I probably could, but that would go against your request of limiting as
> much as possible the allocation of a copy of the input line: Far more
> insns would then also undergo this setup for a possible 2nd parsing pass.
> It may be possible to limit the set some by taking into account further
> attributes (like, leaving aside move-with-sign-extend which you object
> to being corrected, SSE2AVX), but the collection of those checks would
> likely be clumsy and harder to maintain than the new attribute. For
> example the consistency check in i386-gen would then not be possible
> anymore, and hence the whole thing would become _silently_ dependent on
> the ordering of templates in the file.

Order dependency is the part of the implementation.

> Furthermore keying this condition to insn attributes would be less
> future proof. While it may not be very likely for new colliding
> mnemonics to appear, if one would which then doesn't fit the chosen
> pattern, the tweaking of the conditional would be more fragile than
> the mere adding of Pass2 in the appropriate template. (Even if you
> continue to object to it, simply consider the last patch of this series
> an example.)
>

I like to avoid the second pass for many common instructions.
The current assembler may not be ideal.  But it works.

-- 
H.J.