Re: [Qemu-devel] [PATCH 0/7] Define "deposit" tcg operation, v2
On Mon, Jan 10, 2011 at 07:23:41PM -0800, Richard Henderson wrote: > Changes since v1: > * No attempt to pack pos+len into one operand. Updated backends > to match this change. > > * Example in the README is a bit more complex. > > * Define an official tcg_scratch_alloc routine, used by the i386 > target for the case in which we need a scratch register. I had > said that triggering this wasn't possible with mainline, but I > was wrong. The rlwimi example I posted in the other thread hits > this case. > > Aurelien suggested using a shorter name like "dep", but I think that > would be more confusing than not. This opcode is not really used > enough to warrent cropping 4 characters, in my opinion. Thanks Richard. In agreement with Aurelien I've applied patch 1 and 6 as a first step. I've also applied a follow-up patch to fix the copy-pastos in the README (loc -> len). Hopefully as more acks come in, we'll apply the rest. Cheers
Re: [Qemu-devel] [PATCH 0/7] Define "deposit" tcg operation, v2
On Mon, Jan 10, 2011 at 07:23:41PM -0800, Richard Henderson wrote: > Changes since v1: > * No attempt to pack pos+len into one operand. Updated backends > to match this change. > > * Example in the README is a bit more complex. > > * Define an official tcg_scratch_alloc routine, used by the i386 > target for the case in which we need a scratch register. I had > said that triggering this wasn't possible with mainline, but I > was wrong. The rlwimi example I posted in the other thread hits > this case. > > Aurelien suggested using a shorter name like "dep", but I think that > would be more confusing than not. This opcode is not really used > enough to warrent cropping 4 characters, in my opinion. > It was just a suggestion, I am also fine with "deposit". The implementation of the deposit op (that is the first patch) looks fine to me, and I am convince this new op is useful. It's probably going to be useful for the s390 target, that tried to solve the same problem with a sync op. To go further, I think we should split the patch series into two (no need to resend it), the op itself, and the implementation/usage in various host/targets. The latter can go part by part, so the various persons can review them separately. The whole series is still useful anyway to see the potential of such a new op. Even if the code for the new op is quite simple, adding a new op has some consequences (the main one being is that it's difficult to remove it latter), so I propose that it is reviewed/acked by a few persons before being committed. People, please do that or send your comments. -- Aurelien Jarno GPG: 1024D/F1BCDB73 [email protected] http://www.aurel32.net
Re: [Qemu-devel] [PATCH 0/7] Define "deposit" tcg operation
On 01/07/2011 03:10 PM, Peter Maydell wrote: > Unless I've missed something, deposit_i32 is basically the same > as the ARM BFI instruction, so for ARM we could use deposit_i32 in > the implementation of BFI (and conversely implement deposit_i32 > using BFI when we're generating for an ARMv6T2 or better host.) That's correct. I'll admit to not knowing the arm/thumb variants very well. While I can see that gcc sometimes implements its similar "insv" operation with BFC and BFI, knowing under what conditions those are available is ... slightly complex to say the least. r~
Re: [Qemu-devel] [PATCH 0/7] Define "deposit" tcg operation
On 7 January 2011 22:42, Richard Henderson wrote: > Emulating i386 -- particularly in 16-bit mode -- requires quite a few > bitfield insert operations, to handle byte and word stores into the > dword registers. On several hosts, this can be done natively, without > resorting to a sequence of and and or instructions. +* deposit_i32/i64 dest, t1, t2, loc + +Deposit T2 as a bitfield into T1, placing the result in DEST. +The bitfield is described by LOC, an immediate value: + + bits 0:7 - the length of the bitfield + bits 8:15 - the position of the first bit + +For example, 0x101 indicates a 1-bit field at bit 1. +This operation would be equivalent to + + dest = (t1 & ~2) | ((t2 << 1) & 2) Unless I've missed something, deposit_i32 is basically the same as the ARM BFI instruction, so for ARM we could use deposit_i32 in the implementation of BFI (and conversely implement deposit_i32 using BFI when we're generating for an ARMv6T2 or better host.) -- PMM
