Skip to content

Commit 1d0e035

Browse files
andrewcarlottiandrewcarlotti
committed
Remove useless function multiversioning features
Supporting a feature in function multiversioning requires that the rest of the toolchain can support enabling the feature. In GCC/Binutils, this requires either that there is an equivalent feature extension available in the `-march` command line option, or that the feature introduces no new instructions (and therefore does not need to be enabled when passing to the assembler). However, many of the features listed in in the original function multiversioning specification do not meet this criteria. These fall into four categories: 1. Features that were originally linked to a specific architecture version (fcma, jscvt, frintts, flagm2, wfxt, rcpc2): these should get their own flags across the toolchain (and I've already added this support in Binutils). 2. Features that are combined with other features in existing command line options: these should also be merged in the function multiversioning specification. 3. Features that indicate support for hint instructions: these can be dropped, since the instructions can be used unconditionally. 4. Features that enable existing instruction behaviour to be changed when a system register flag is set: the function multiversioning resolvers don't check for runtime enablement of the control flags, so this isn't a suitable way of exploiting the behaviour enabled by those flags. 5. Features that can also be expressed as a combination of two other features. We therefore remove support for the following features from the specification: - sha1 (2): included within +sha2 - pmull (2): included within +aes - dit (4) - dgh (3) - ebf16 (4) - sve-bf16 (5) - sve-ebf16 (4) - sve-i8mm (5) - sve2-pmull128 (2): included within +sve2-aes - memtag2 (2): included within +memtag - memtag3 (4) - ssbs2 (2): included within +ssbs - bti (3) - ls64_v, ls64_accdata (2): included within +ls64
1 parent a8e69f2 commit 1d0e035

File tree

1 file changed

+14
-22
lines changed

1 file changed

+14
-22
lines changed

‎main/acle.md

+14-22
Original file line numberDiff line numberDiff line change
@@ -375,7 +375,14 @@ Armv8.4-A [[ARMARMv84]](#ARMARMv84). Support is added for the Dot Product intrin
375375
* Change name mangling of the default version.
376376
* Align priorities to account for feature dependencies.
377377
* Introduce alternative names (aliases) `rdma` for `rdm`.
378-
* Correct FEAT_BTI feature register value.
378+
* Combine some features, to align more closely with compiler command line
379+
options (eliminating sha1, pmull, sve2-pmull128, memtag2, ssbs2, ls64_v
380+
and ls64_accdata).
381+
* Remove features that are combinations of other features (sve-bf16 and
382+
sve-i8mm).
383+
* Remove features that only enable hint instructions (dgh and bti).
384+
* Remove features that indicate only that a specific behaviour can be
385+
separately switched on (dit, ebf16, sve-ebf16 and memtag3).
379386
* Introduced the `__ARM_FEATURE_PAUTH_LR` feature macro in section
380387
[Pointer Authentication](#pointer-authentication) to indicate target support
381388
for the Armv9.5-A's PAC Enhancements.
@@ -2590,14 +2597,11 @@ The following table lists the architectures feature mapping for AArch64
25902597
| 106 | `FEAT_SM3`, `FEAT_SM4` | sm4 | ```ID_AA64ISAR0_EL1.SM4 == 0b0001 AND ``` <br> ```ID_AA64ISAR0_EL1.SM3 == 0b0001``` |
25912598
| 108 | `FEAT_RDM` | rdm, rdma | ```ID_AA64ISAR0_EL1.RDM == 0b0001``` |
25922599
| 110 | `FEAT_CRC32` | crc | ```ID_AA64ISAR0_EL1.CRC32 == 0b0001``` |
2593-
| 120 | `FEAT_SHA1` | sha1 | ```ID_AA64ISAR0_EL1.SHA1 == 0b0001``` |
2594-
| 130 | `FEAT_SHA256` | sha2 | ```ID_AA64ISAR0_EL1.SHA2 == 0b0001``` |
2600+
| 130 | `FEAT_SHA1`,`FEAT_SHA256`| sha2 | ```ID_AA64ISAR0_EL1.SHA2 == 0b0001``` |
25952601
| 140 | `FEAT_SHA512`,`FEAT_SHA3`| sha3 | ```ID_AA64ISAR0_EL1.SHA3 != 0b0000``` |
2596-
| 150 | `FEAT_AES` | aes | ```ID_AA64ISAR0_EL1.AES >= 0b0001``` |
2597-
| 160 | `FEAT_PMULL` | pmull | ```ID_AA64ISAR0_EL1.AES == 0b0010``` |
2602+
| 150 | `FEAT_AES`,`FEAT_PMULL` | aes | ```ID_AA64ISAR0_EL1.AES >= 0b0010``` |
25982603
| 170 | `FEAT_FP16` | fp16 | ```ID_AA64PFR0_EL1.FP == 0b0001``` |
25992604
| 175 | `FEAT_FHM` | fp16fml | ```ID_AA64ISAR0_EL1.FHM == 0b0001``` |
2600-
| 180 | `FEAT_DIT` | dit | ```ID_AA64PFR0_EL1.DIT == 0b0001``` |
26012605
| 190 | `FEAT_DPB` | dpb | ```ID_AA64ISAR1_EL1.DPB >= 0b0001``` |
26022606
| 200 | `FEAT_DPB2` | dpb2 | ```ID_AA64ISAR1_EL1.DPB == 0b0010``` |
26032607
| 210 | `FEAT_JSCVT` | jscvt | ```ID_AA64ISAR1_EL1.JSCVT == 0b0001``` |
@@ -2606,35 +2610,23 @@ The following table lists the architectures feature mapping for AArch64
26062610
| 240 | `FEAT_LRCPC2` | rcpc2 | ```ID_AA64ISAR1_EL1.LRCPC == 0b0010``` |
26072611
| 241 | `FEAT_LRCPC3` | rcpc3 | ```ID_AA64ISAR1_EL1.LRCPC == 0b0011``` |
26082612
| 250 | `FEAT_FRINTTS` | frintts | ```ID_AA64ISAR1_EL1.FRINTTS == 0b0001``` |
2609-
| 260 | `FEAT_DGH` | dgh | ```ID_AA64ISAR1_EL1.DGH == 0b0001``` |
26102613
| 270 | `FEAT_I8MM` | i8mm | ```ID_AA64ISAR1_EL1.I8MM == 0b0001``` |
26112614
| 280 | `FEAT_BF16` | bf16 | ```ID_AA64ISAR1_EL1.BF16 != 0b0000``` |
2612-
| 290 | `FEAT_EBF16` | ebf16 | ```ID_AA64ISAR1_EL1.BF16 == 0b0010``` |
26132615
| 300 | `FEAT_RPRES` | rpres | ```ID_AA64ISAR2_EL1.RPRES == 0b0001``` |
26142616
| 310 | `FEAT_SVE` | sve | ```ID_AA64PFR0_EL1.SVE != 0b0000 AND ``` <br> ```ID_AA64ZFR0_EL1.SVEver == 0b0000``` |
2615-
| 320 | `FEAT_BF16` | sve-bf16 | ```ID_AA64ZFR0_EL1.BF16 != 0b0000``` |
2616-
| 330 | `FEAT_EBF16` | sve-ebf16 | ```ID_AA64ZFR0_EL1.BF16 == 0b0010``` |
2617-
| 340 | `FEAT_I8MM` | sve-i8mm | ```ID_AA64ZFR0_EL1.I8MM == 0b00001``` |
26182617
| 350 | `FEAT_F32MM` | f32mm | ```ID_AA64ZFR0_EL1.F32MM == 0b00001``` |
26192618
| 360 | `FEAT_F64MM` | f64mm | ```ID_AA64ZFR0_EL1.F64MM == 0b00001``` |
26202619
| 370 | `FEAT_SVE2` | sve2 | ```ID_AA64PFR0_EL1.SVE != 0b0000 AND ``` <br> ```ID_AA64ZFR0_EL1.SVEver == 0b0001``` |
2621-
| 380 | `FEAT_SVE_AES` | sve2-aes | ```ID_AA64ZFR0_EL1.AES == 0b0001 OR ``` <br> ```ID_AA64ZFR0_EL1.AES == 0b0010``` |
2622-
| 390 | `FEAT_SVE_PMULL128` | sve2-pmull128 | ```ID_AA64ZFR0_EL1.AES == 0b0010``` |
2620+
| 380 | `FEAT_SVE_AES`,<br>`FEAT_SVE_PMULL128`| sve2-aes | ```ID_AA64ZFR0_EL1.AES == 0b0010``` |
26232621
| 400 | `FEAT_SVE_BitPerm` | sve2-bitperm | ```ID_AA64ZFR0_EL1.BitPerm == 0b0001``` |
26242622
| 410 | `FEAT_SVE_SHA3` | sve2-sha3 | ```ID_AA64ZFR0_EL1.SHA3 == 0b0001``` |
26252623
| 420 | `FEAT_SM3`,`FEAT_SVE_SM4`| sve2-sm4 | ```ID_AA64ZFR0_EL1.SM4 == 0b0001``` |
26262624
| 430 | `FEAT_SME` | sme | ```ID_AA64PFR1_EL1.SME == 0b0001``` |
2627-
| 440 | `FEAT_MTE` | memtag | ```ID_AA64PFR1_EL1.MTE >= 0b0001``` |
2628-
| 450 | `FEAT_MTE2` | memtag2 | ```ID_AA64PFR1_EL1.MTE >= 0b0010``` |
2629-
| 460 | `FEAT_MTE3` | memtag3 | ```ID_AA64PFR1_EL1.MTE >= 0b0011``` |
2625+
| 440 | `FEAT_MTE`,`FEAT_MTE2` | memtag | ```ID_AA64PFR1_EL1.MTE >= 0b0010``` |
26302626
| 470 | `FEAT_SB` | sb | ```ID_AA64ISAR1_EL1.SB == 0b0001``` |
26312627
| 480 | `FEAT_SPECRES` | predres | ```ID_AA64ISAR1_EL1.SPECRES == 0b0001``` |
2632-
| 490 | `FEAT_SSBS` | ssbs | ```ID_AA64PFR1_EL1.SSBS == 0b0001``` |
2633-
| 500 | `FEAT_SSBS2` | ssbs2 | ```ID_AA64PFR1_EL1.SSBS == 0b0010``` |
2634-
| 510 | `FEAT_BTI` | bti | ```ID_AA64PFR1_EL1.BT == 0b0001``` |
2635-
| 520 | `FEAT_LS64` | ls64 | ```ID_AA64ISAR1_EL1.LS64 >= 0b0001``` |
2636-
| 530 | `FEAT_LS64_V` | ls64_v | ```ID_AA64ISAR1_EL1.LS64 >= 0b0010``` |
2637-
| 540 | `FEAT_LS64_ACCDATA` | ls64_accdata | ```ID_AA64ISAR1_EL1.LS64 >= 0b0011``` |
2628+
| 490 | `FEAT_SSBS`,`FEAT_SSBS2` | ssbs | ```ID_AA64PFR1_EL1.SSBS >= 0b0010``` |
2629+
| 520 | `FEAT_LS64`,`FEAT_LS64_V`,<br>`FEAT_LS64_ACCDATA`| ls64 | ```ID_AA64ISAR1_EL1.LS64 >= 0b0011``` |
26382630
| 550 | `FEAT_WFxT` | wfxt | ```ID_AA64ISAR2_EL1.WFxT == 0b0001``` |
26392631
| 560 | `FEAT_SME_F64F64` | sme-f64f64 | ```ID_AA64SMFR0_EL1.F64F64 == 0b0001``` |
26402632
| 570 | `FEAT_SME_I16I64` | sme-i16i64 | ```ID_AA64SMFR0_EL1.I16I64 == 0b1111``` |

0 commit comments

Comments
 (0)