481761 - TM: ARM backend cleanup

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

This is the start of a merge with upstream tamarin tracemonkey ARM stuff, though with some changes.  There's a number of patches here in series.

Comment 1

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: 1 - simple cleanup

Some simple cleanup -- rename ccName to condName, rename Scratch register to IP.

Comment 2

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: 2 - ALU macros, step 1

Step 1 of ALU conversion macros; also gets rid of some x86isms (MOV instead of MR, etc).

Comment 3

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: 3 - ALU macros, step 2

Converts ADD, SUB, some shifts, TEST, and CMP to ALU macros.

Comment 4

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: 4 - minor asm_cmov cleanup

Some minor asm_cmov cleanup.

Comment 5

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: 5 - rename condition codes for MOV to match ARM names

Rename condition codes for MOV to match ARM instead of x86 names

Comment 6

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: 6 - make load/store functions

Make load and store functions instead of convoluted macros.  Also, apparently, changes the MOV_cond macro so the condition comes first in the args instead of last.

Comment 7

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: 7 - add code for using MOVT/MOVW for 32-bit loads

On ARMv6t2 or newer, use MOVW/MOVT to load 32-bit constants instead of reading from memory.  Note that this patch has this protected by NJ_ARM_HAVE_MOVW, but the patch in bug 480796 builds on top of this series to remove the ifdefs and do the the detection at runtime.

Comment 8

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: 8 - silly debug-only assert bug

_Dm -> _Dn in a NanoAssert; might have problems building with DEBUG without this on Windows CE (since those macros end up being used on CE only).

Comment 9

[Graydon Hoare :graydon]

Comment on attachment 365767 [details] [diff] [review]
2 - ALU macros, step 1

Hmm, actually, can you modify the comment on the ALU macros to explain the cond argument as well? And/or possibly wire it to AL in the macro, excluding the cond argument altogether, if we're actually not using the condition-bit system in instruction selection? It looks to me like we're not (it's sort of an exotic thing that requires semantics beyond our assembler interface, for most uses...)

Comment 10

[Graydon Hoare :graydon]

Comment on attachment 365772 [details] [diff] [review]
3 - ALU macros, step 2

I'm liking that the macros are getting smaller and/or being factored out into functions. I wonder if the ALU ones can be turned into calls to real functions -- after CPP stringifying and token-pasting the 'op' argument -- the same way you converted the add and sub immediates. Thoughts?

Also (though it can be a followup patch) all the final ALU macros should probably use _arg forms not arg forms, as the other macros do, to reduce possible name collisions.

Comment 11

[Graydon Hoare :graydon]

Comment on attachment 365783 [details] [diff] [review]
7 - add code for using MOVT/MOVW for 32-bit loads

Fine but same comment as before: the macros are doing more than CPP-only stuff, should move majority to helper function. Fine as follow-up patch.

Comment 12

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: 9 - use LD32_nochk instead of hardcoded for quad load

Two more quick ones..

Get rid of the hardcoded loads in asm_quad and just use LD32, so that intelligent things happen for 0, 16-bit values, etc.

Also change the jmp macros to use the existing B_cond macro.

Comment 13

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: 10 - underrunProtect some more

(Final one, I'll land this whole set today, I promise!)

We were missing a few underrunProtects, usually with LD32_nochk.. fix that, and get rid of LD32_nochks as we can.

Comment 14

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

All checked in, finally.