[2/3,aarch64] Add function multiversioning support

  This adds initial support for function multiversion on aarch64 using the
target_version and target_clones attributes. This mostly follows the
Beta specification in the ACLE [1], with a few diffences that remain to
be fixed:

- Symbol mangling for target_clones differs from that for target_version
  and does not match the mangling specified in the ACLE. This
  inconsistency is also present in i386 and rs6000 mangling.
- The target_clones attribute does not currently support an implicit
  "default" version.
- Unrecognised target names in a target_clones attribute should be
  ignored (with an optional warning), but currently cause an error to be
  raised instead.
- There is no option to disable function multiversioning at compile
  time.
- There is no support for function multiversioning in C, since this is
  not yet enabled in the frontend. On the other hand, this patch
  happens to enable multiversioning in Ada and D as well, using their
  existing frontend support.

This patch relies on adding functionality to libgcc, to support:
- struct { unsigned long long features; } __aarch64_cpu_features;
- void __init_cpu_features (void);
- void __init_cpu_features_resolver (unsigned long hwcap,
				     const __ifunc_arg_t *arg);
This support matches the interface currently used in LLVM's compiler-rt,
and will be implemented in a future patch (which will be merged before
merging this patch).

This version of the patch incorrectly uses __init_cpu_features in the
ifunc resolvers, which could lead to invalid library calls at load time.
I will fix this to use __init_cpu_features_resolver in a future version
of the patch.

[1] https://github.com/ARM-software/acle/blob/main/main/acle.md#function-multi-versioning

gcc/ChangeLog:

	* attribs.cc (decl_attributes): Pass attribute name to target
	hook.
	* config/aarch64/aarch64.cc
	(aarch64_process_target_version_attr): New.
	(aarch64_option_valid_attribute_p): Add check and support for
	target_version attribute.
	(enum CPUFeatures): New list of for bitmask positions.
	(aarch64_fmv_feature_data): New.
	(get_feature_bit): New.
	(get_feature_mask_for_version): New.
	(compare_feature_masks): New.
	(aarch64_compare_version_priority): New.
	(make_resolver_func): New.
	(add_condition_to_bb): New.
	(compare_feature_version_info): New.
	(dispatch_function_versions): New.
	(aarch64_generate_version_dispatcher_body): New.
	(aarch64_get_function_versions_dispatcher): New.
	(aarch64_common_function_versions): New.
	(aarch64_mangle_decl_assembler_name): New.
	(TARGET_OPTION_VALID_VERSION_ATTRIBUTE_P): New implementation.
	(TARGET_OPTION_EXPANDED_CLONES_ATTRIBUTE): New implementation.
	(TARGET_OPTION_FUNCTION_VERSIONS): New implementation.
	(TARGET_COMPARE_VERSION_PRIORITY): New implementation.
	(TARGET_GENERATE_VERSION_DISPATCHER_BODY): New implementation.
	(TARGET_GET_FUNCTION_VERSIONS_DISPATCHER): New implementation.
	(TARGET_MANGLE_DECL_ASSEMBLER_NAME): New implementation.
  

Andrew Carlotti <andrew.carlotti@arm.com> writes:
> This adds initial support for function multiversion on aarch64 using the
> target_version and target_clones attributes. This mostly follows the
> Beta specification in the ACLE [1], with a few diffences that remain to
> be fixed:
>
> - Symbol mangling for target_clones differs from that for target_version
>   and does not match the mangling specified in the ACLE. This
>   inconsistency is also present in i386 and rs6000 mangling.
> - The target_clones attribute does not currently support an implicit
>   "default" version.
> - Unrecognised target names in a target_clones attribute should be
>   ignored (with an optional warning), but currently cause an error to be
>   raised instead.
> - There is no option to disable function multiversioning at compile
>   time.
> - There is no support for function multiversioning in C, since this is
>   not yet enabled in the frontend. On the other hand, this patch
>   happens to enable multiversioning in Ada and D as well, using their
>   existing frontend support.
>
> This patch relies on adding functionality to libgcc, to support:
> - struct { unsigned long long features; } __aarch64_cpu_features;
> - void __init_cpu_features (void);
> - void __init_cpu_features_resolver (unsigned long hwcap,
> 				     const __ifunc_arg_t *arg);
> This support matches the interface currently used in LLVM's compiler-rt,
> and will be implemented in a future patch (which will be merged before
> merging this patch).
>
> This version of the patch incorrectly uses __init_cpu_features in the
> ifunc resolvers, which could lead to invalid library calls at load time.
> I will fix this to use __init_cpu_features_resolver in a future version
> of the patch.
>
> [1] https://github.com/ARM-software/acle/blob/main/main/acle.md#function-multi-versioning
>
> gcc/ChangeLog:
>
> 	* attribs.cc (decl_attributes): Pass attribute name to target
> 	hook.
> 	* config/aarch64/aarch64.cc
> 	(aarch64_process_target_version_attr): New.
> 	(aarch64_option_valid_attribute_p): Add check and support for
> 	target_version attribute.
> 	(enum CPUFeatures): New list of for bitmask positions.
> 	(aarch64_fmv_feature_data): New.
> 	(get_feature_bit): New.
> 	(get_feature_mask_for_version): New.
> 	(compare_feature_masks): New.
> 	(aarch64_compare_version_priority): New.
> 	(make_resolver_func): New.
> 	(add_condition_to_bb): New.
> 	(compare_feature_version_info): New.
> 	(dispatch_function_versions): New.
> 	(aarch64_generate_version_dispatcher_body): New.
> 	(aarch64_get_function_versions_dispatcher): New.
> 	(aarch64_common_function_versions): New.
> 	(aarch64_mangle_decl_assembler_name): New.
> 	(TARGET_OPTION_VALID_VERSION_ATTRIBUTE_P): New implementation.
> 	(TARGET_OPTION_EXPANDED_CLONES_ATTRIBUTE): New implementation.
> 	(TARGET_OPTION_FUNCTION_VERSIONS): New implementation.
> 	(TARGET_COMPARE_VERSION_PRIORITY): New implementation.
> 	(TARGET_GENERATE_VERSION_DISPATCHER_BODY): New implementation.
> 	(TARGET_GET_FUNCTION_VERSIONS_DISPATCHER): New implementation.
> 	(TARGET_MANGLE_DECL_ASSEMBLER_NAME): New implementation.

Nice.  Mostly LGTM, but some comments below.

> diff --git a/gcc/attribs.cc b/gcc/attribs.cc
> index a3c4a81e8582ea4fd06b9518bf51fad7c998ddd6..cc935b502028392ebdc105f940900f01f79196a7 100644
> --- a/gcc/attribs.cc
> +++ b/gcc/attribs.cc
> @@ -657,7 +657,8 @@ decl_attributes (tree *node, tree attributes, int flags,
>       options to the attribute((target(...))) list.  */
>    if (TREE_CODE (*node) == FUNCTION_DECL
>        && current_target_pragma
> -      && targetm.target_option.valid_attribute_p (*node, NULL_TREE,
> +      && targetm.target_option.valid_attribute_p (*node,
> +						  get_identifier("target"),
>  						  current_target_pragma, 0))
>      {
>        tree cur_attr = lookup_attribute ("target", attributes);
> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> index 9c3c0e705e2e6ea3b55b4a5f1e7d3360f91eb51d..ca0e2a2507ffdbf99e17b77240504bf2d175b9c0 100644
> --- a/gcc/config/aarch64/aarch64.cc
> +++ b/gcc/config/aarch64/aarch64.cc
> @@ -19088,11 +19088,70 @@ aarch64_process_target_attr (tree args)
>    return true;
>  }
>  
> +/* Parse the tree in ARGS that contains the targeti_version attribute
> +   information and update the global target options space.  */
> +
> +bool
> +aarch64_process_target_version_attr (tree args)
> +{
> +  if (TREE_CODE (args) == TREE_LIST)
> +    {
> +      if (TREE_CHAIN (args))
> +	{
> +	  error ("attribute %<target_version%> has multiple values");
> +	  return false;
> +	}
> +      args = TREE_VALUE (args);
> +    }
> +
> +  if (!args || TREE_CODE (args) != STRING_CST)
> +    {
> +      error ("attribute %<target_version%> argument not a string");
> +      return false;
> +    }
> +
> +  const char *str = TREE_STRING_POINTER (args);
> +  if (strcmp (str, "default") == 0)
> +    return true;
> +
> +  auto with_plus = std::string ("+") + str;
> +  enum aarch_parse_opt_result parse_res;
> +  auto isa_flags = aarch64_asm_isa_flags;
> +
> +  std::string invalid_extension;
> +  parse_res = aarch64_parse_extension (with_plus.c_str(), &isa_flags,
> +				       &invalid_extension);
> +
> +  if (parse_res == AARCH_PARSE_OK)
> +    {
> +      aarch64_set_asm_isa_flags (isa_flags);
> +      return true;
> +    }
> +
> +  switch (parse_res)
> +    {
> +      case AARCH_PARSE_MISSING_ARG:
> +	error ("missing value in %<target_version%> attribute");
> +	break;
> +
> +      case AARCH_PARSE_INVALID_FEATURE:
> +	error ("invalid feature modifier %qs of value %qs in "
> +	       "%<target_version%> attribute", invalid_extension.c_str (),
> +	       with_plus.c_str());
> +	break;
> +
> +      default:
> +	gcc_unreachable ();
> +    }
> +
> +  return false;
> +}
> +
>  /* Implement TARGET_OPTION_VALID_ATTRIBUTE_P.  This is used to
>     process attribute ((target ("..."))).  */
>  
>  static bool
> -aarch64_option_valid_attribute_p (tree fndecl, tree, tree args, int)
> +aarch64_option_valid_attribute_p (tree fndecl, tree name, tree args, int)
>  {
>    struct cl_target_option cur_target;
>    bool ret;
> @@ -19100,13 +19159,22 @@ aarch64_option_valid_attribute_p (tree fndecl, tree, tree args, int)
>    tree new_target, new_optimize;
>    tree existing_target = DECL_FUNCTION_SPECIFIC_TARGET (fndecl);
>  
> +  bool target_version_p;
> +  const char *attr_name = IDENTIFIER_POINTER (name);
> +  if (strcmp (attr_name, "target") == 0)
> +    target_version_p = false;
> +  else if (strcmp (attr_name, "target_version") == 0)
> +    target_version_p = true;
> +  else
> +    gcc_assert (false);
> +
>    /* If what we're processing is the current pragma string then the
>       target option node is already stored in target_option_current_node
>       by aarch64_pragma_target_parse in aarch64-c.cc.  Use that to avoid
>       having to re-parse the string.  This is especially useful to keep
>       arm_neon.h compile times down since that header contains a lot
>       of intrinsics enclosed in pragmas.  */
> -  if (!existing_target && args == current_target_pragma)
> +  if (!target_version_p && !existing_target && args == current_target_pragma)
>      {
>        DECL_FUNCTION_SPECIFIC_TARGET (fndecl) = target_option_current_node;
>        return true;
> @@ -19142,7 +19210,25 @@ aarch64_option_valid_attribute_p (tree fndecl, tree, tree args, int)
>      cl_target_option_restore (&global_options, &global_options_set,
>  			      TREE_TARGET_OPTION (target_option_current_node));
>  
> -  ret = aarch64_process_target_attr (args);
> +  if (!target_version_p)
> +    {
> +      ret = aarch64_process_target_attr (args);
> +      if (ret)
> +	{
> +	  tree version_attr = lookup_attribute ("target_version",
> +						DECL_ATTRIBUTES (fndecl));
> +	  if (version_attr != NULL_TREE)
> +	    {
> +	      /* Reapply any target_version attribute after target attribute.
> +		 This should be equivalent to applying the target_version once
> +		 after processing all target attributes.  */
> +	      tree version_args = TREE_VALUE (version_attr);
> +	      ret = aarch64_process_target_version_attr (version_args);
> +	    }
> +	}
> +    }
> +  else
> +    ret = aarch64_process_target_version_attr (args);
>  
>    /* Set up any additional state.  */
>    if (ret)
> @@ -19173,6 +19259,730 @@ aarch64_option_valid_attribute_p (tree fndecl, tree, tree args, int)
>    return ret;
>  }
>  
> +/* This enum needs to match the enum used in libgcc cpuinfo.c.  */
> +//TODO: Does this clash with or overlap an existing list of target features?

Yeah, good question :)  Might be worth sticking it in an anonymous namespace
just in case.  We can do something more drastic if that turns out not
to be enough.

libgcc should have access to gcc/ headers, so I think we could probably
share this enum between them.

> +enum CPUFeatures {
> +  FEAT_RNG,
> +  FEAT_FLAGM,
> +  FEAT_FLAGM2,
> +  FEAT_FP16FML,
> +  FEAT_DOTPROD,
> +  FEAT_SM4,
> +  FEAT_RDM,
> +  FEAT_LSE,
> +  FEAT_FP,
> +  FEAT_SIMD,
> +  FEAT_CRC,
> +  FEAT_SHA1,
> +  FEAT_SHA2,
> +  FEAT_SHA3,
> +  FEAT_AES,
> +  FEAT_PMULL,
> +  FEAT_FP16,
> +  FEAT_DIT,
> +  FEAT_DPB,
> +  FEAT_DPB2,
> +  FEAT_JSCVT,
> +  FEAT_FCMA,
> +  FEAT_RCPC,
> +  FEAT_RCPC2,
> +  FEAT_FRINTTS,
> +  FEAT_DGH,
> +  FEAT_I8MM,
> +  FEAT_BF16,
> +  FEAT_EBF16,
> +  FEAT_RPRES,
> +  FEAT_SVE,
> +  FEAT_SVE_BF16,
> +  FEAT_SVE_EBF16,
> +  FEAT_SVE_I8MM,
> +  FEAT_SVE_F32MM,
> +  FEAT_SVE_F64MM,
> +  FEAT_SVE2,
> +  FEAT_SVE_AES,
> +  FEAT_SVE_PMULL128,
> +  FEAT_SVE_BITPERM,
> +  FEAT_SVE_SHA3,
> +  FEAT_SVE_SM4,
> +  FEAT_SME,
> +  FEAT_MEMTAG,
> +  FEAT_MEMTAG2,
> +  FEAT_MEMTAG3,
> +  FEAT_SB,
> +  FEAT_PREDRES,
> +  FEAT_SSBS,
> +  FEAT_SSBS2,
> +  FEAT_BTI,
> +  FEAT_LS64,
> +  FEAT_LS64_V,
> +  FEAT_LS64_ACCDATA,
> +  FEAT_WFXT,
> +  FEAT_SME_F64,
> +  FEAT_SME_I64,
> +  FEAT_SME2,
> +  FEAT_RCPC3, //TODO: Check this index - needs to agree with LLVM.
> +  FEAT_MAX
> +};
> +
> +typedef struct
> +{
> +  const char *name;
> +  int priority;
> +  unsigned long long feature_mask;
> +} aarch64_fmv_feature_datum;
> +
> +/* List these in priority order, to make it easier to sort target strings.  */
> +static aarch64_fmv_feature_datum aarch64_fmv_feature_data[] = {
> +  {"default",        0, 0ULL},
> +  {"rng",            10, 1ULL << FEAT_RNG},
> +  {"flagm",          20, 1ULL << FEAT_FLAGM},
> +  {"flagm2",         30, 1ULL << FEAT_FLAGM2},
> +  {"fp16fml",        40, 1ULL << FEAT_FP16FML},
> +  {"dotprod",        50, 1ULL << FEAT_DOTPROD},
> +  {"sm4",            60, 1ULL << FEAT_SM4},
> +  {"rdm",            70, 1ULL << FEAT_RDM},
> +  {"lse",            80, 1ULL << FEAT_LSE},
> +  {"fp",             90, 1ULL << FEAT_FP},
> +  {"simd",          100, 1ULL << FEAT_SIMD},
> +  {"crc",           110, 1ULL << FEAT_CRC},
> +  {"sha1",          120, 1ULL << FEAT_SHA1},
> +  {"sha2",          130, 1ULL << FEAT_SHA2},
> +  {"sha3",          140, 1ULL << FEAT_SHA3},
> +  {"aes",           150, 1ULL << FEAT_AES},
> +  {"pmull",         160, 1ULL << FEAT_PMULL},
> +  {"fp16",          170, 1ULL << FEAT_FP16},
> +  {"dit",           180, 1ULL << FEAT_DIT},
> +  {"dpb",           190, 1ULL << FEAT_DPB},
> +  {"dpb2",          200, 1ULL << FEAT_DPB2},
> +  {"jscvt",         210, 1ULL << FEAT_JSCVT},
> +  {"fcma",          220, 1ULL << FEAT_FCMA},
> +  {"rcpc",          230, 1ULL << FEAT_RCPC},
> +  {"rcpc2",         240, 1ULL << FEAT_RCPC2},
> +  {"rcpc3",         241, 1ULL << FEAT_RCPC3},
> +  {"frintts",       250, 1ULL << FEAT_FRINTTS},
> +  {"dgh",           260, 1ULL << FEAT_DGH},
> +  {"i8mm",          270, 1ULL << FEAT_I8MM},
> +  {"bf16",          280, 1ULL << FEAT_BF16},
> +  {"ebf16",         290, 1ULL << FEAT_EBF16},
> +  {"rpres",         300, 1ULL << FEAT_RPRES},
> +  {"sve",           310, 1ULL << FEAT_SVE},
> +  {"sve-bf16",      320, 1ULL << FEAT_SVE_BF16},

I suppose we need a way of handling things like this.  We deliberately
don't have a +sve-bf16 feature flag because it's implied by the
combination +sve+bf16.

We probably need to do something similar to LLVM and have a mapping
from these strings to a set of supported feature flags, either as
a string or (preferably) as an aarch64_feature_flags.

> +  {"sve-ebf16",     330, 1ULL << FEAT_SVE_EBF16},
> +  {"sve-i8mm",      340, 1ULL << FEAT_SVE_I8MM},
> +  {"f32mm",         350, 1ULL << FEAT_SVE_F32MM},
> +  {"f64mm",         360, 1ULL << FEAT_SVE_F64MM},
> +  {"sve2",          370, 1ULL << FEAT_SVE2},
> +  {"sve2-aes",      380, 1ULL << FEAT_SVE_AES},
> +  {"sve2-pmull128", 390, 1ULL << FEAT_SVE_PMULL128},
> +  {"sve2-bitperm",  400, 1ULL << FEAT_SVE_BITPERM},
> +  {"sve2-sha3",     410, 1ULL << FEAT_SVE_SHA3},
> +  {"sve2-sm4",      420, 1ULL << FEAT_SVE_SM4},
> +  {"sme",           430, 1ULL << FEAT_SME},
> +  {"memtag",        440, 1ULL << FEAT_MEMTAG},
> +  {"memtag2",       450, 1ULL << FEAT_MEMTAG2},
> +  {"memtag3",       460, 1ULL << FEAT_MEMTAG3},
> +  {"sb",            470, 1ULL << FEAT_SB},
> +  {"predres",       480, 1ULL << FEAT_PREDRES},
> +  {"ssbs",          490, 1ULL << FEAT_SSBS},
> +  {"ssbs2",         500, 1ULL << FEAT_SSBS2},
> +  {"bti",           510, 1ULL << FEAT_BTI},
> +  {"ls64",          520, 1ULL << FEAT_LS64},
> +  {"ls64_v",        530, 1ULL << FEAT_LS64_V},
> +  {"ls64_accdata",  540, 1ULL << FEAT_LS64_ACCDATA},
> +  {"wfxt",          550, 1ULL << FEAT_WFXT},
> +  {"sme-f64f64",    560, 1ULL << FEAT_SME_F64},
> +  {"sme-i16i64",    570, 1ULL << FEAT_SME_I64},
> +  {"sme2",          580, 1ULL << FEAT_SME2}
> +};

Do we anticipate any mask having more than one bit set?  I wonder if
aarch64_fmv_feature_datum should have a CPUFeatures instead, since
that would be more future-proof when we run out of bits.

Also, if we did that, it might make sense to put the list of features
in a .def file.  It would then be possible to use that to define the
enum and aarch64_fmv_feature_data, which would help to make sure that
they stay in sync.

> +
> +/* Look up a single feature name, and return the bitmask.  */
> +unsigned long long
> +get_feature_bit (char *name)
> +{
> +  /* Skip default entry here.  */
> +  for (int i = 1; i < FEAT_MAX; i++)
> +      if (strcmp(aarch64_fmv_feature_data[i].name, name) == 0)
> +	return aarch64_fmv_feature_data[i].feature_mask;

Formatting nit (sorry): the if and return are indented too far.
Missing space between "strcmp" and "(".

> +  return 0;
> +}
> +
> +/* This parses the attribute arguments to target_version in DECL and the
> +   feature mask required to select those targets.  No adjustments are made to
> +   add or remove redundant feature requirements.  */
> +
> +unsigned long long

This is probably personal preference, but how about hiding this behind
a typedef now?  I assume even unsigned long long won't be enough for long.
And it's harder to retroactively find which unsigned long longs are
for this and which are for something else.

> +get_feature_mask_for_version (tree decl)
> +{
> +  tree version_attr = lookup_attribute ("target_version", DECL_ATTRIBUTES (decl));

Formatting nit: long line.

> +  if (version_attr == NULL)
> +    return 0;
> +
> +  const char *version_string = TREE_STRING_POINTER (TREE_VALUE (TREE_VALUE
> +						    (version_attr)));
> +  if (strcmp (version_string, "default") == 0
> +      || strcmp (version_string, "") == 0)
> +    return 0;
> +
> +  int attr_len = strlen (version_string);
> +
> +  char *feature_string = XNEWVEC (char, attr_len+ 1);
> +  strcpy (feature_string, version_string);

There's a checked xstrdup we can use for this.

> +
> +  int count = 1;
> +  for (int i = 0; i < attr_len; i++)
> +    {
> +      if (feature_string[i] == '+')
> +	{
> +	  feature_string[i] = '\0';
> +	  count++;
> +	}
> +    }
> +
> +  unsigned long long feature_mask = 0ULL;
> +  char *cur_feature = feature_string;
> +  for (int i = 0; i < count; i++)
> +    {
> +      unsigned long long feature_bit = get_feature_bit (cur_feature);
> +      if (feature_bit == 0)
> +	{
> +	  /* TODO: For target_clones, we should just ignore this version
> +	     instead.  */
> +	  error_at (DECL_SOURCE_LOCATION (decl), 0,
> +		      "Unrecognised feature %s in function version string",
> +		      cur_feature);

GCC's convention is that error messages start with lower case.

> +	  feature_mask = -1ULL;
> +	}
> +      feature_mask |= feature_bit;
> +      cur_feature += strlen(cur_feature) + 1;

Formatting nit: missing space between "strlen" and "(".

> +    }
> +  XDELETEVEC (feature_string);
> +  return feature_mask;
> +}
> +
> +/* Compare priorities of two feature masks. Return:
> +     1: mask1 is higher priority
> +    -1: mask2 is higher priority
> +     0: masks are equal.  */
> +
> +int
> +compare_feature_masks (unsigned long long mask1, unsigned long long mask2)
> +{
> +  int pop1 = __builtin_popcountll(mask1);
> +  int pop2 = __builtin_popcountll(mask2);

Not sure we can use __builtin_popcountll unconditionally.  Probably best
to use popcount_hwi instead.

> +  if (pop1 > pop2)
> +    return 1;
> +  if (pop2 > pop1)
> +    return -1;

I suppose this is implementing the ACLE rule:

4. Select the most specific version else

But I wonder if the rule is the right approach.  E.g. simd+fp16 would
be considered more specific than sve, even though sve implies fp16.

The behaviour without this seems reasonable on the face of it
(rank by the highest priority point of difference), but I've not
thought about it much.

> +  unsigned long long diff_mask = mask1 ^ mask2;
> +  if (diff_mask == 0ULL)
> +    return 0;
> +  for (int i = FEAT_MAX - 1; i > 0; i--)
> +    {
> +      unsigned long long bit_mask = aarch64_fmv_feature_data[i].feature_mask;
> +      if (diff_mask & bit_mask)
> +	return (mask1 & bit_mask) ? 1 : -1;
> +    }
> +  gcc_unreachable();
> +}
> +
> +int
> +aarch64_compare_version_priority (tree decl1, tree decl2)
> +{
> +  unsigned long long mask1 = get_feature_mask_for_version (decl1);
> +  unsigned long long mask2 = get_feature_mask_for_version (decl2);
> +
> +  return compare_feature_masks (mask1, mask2);
> +}
> +
> +/* Make the resolver function decl to dispatch the versions of
> +   a multi-versioned function,  DEFAULT_DECL.  IFUNC_ALIAS_DECL is
> +   ifunc alias that will point to the created resolver.  Create an
> +   empty basic block in the resolver and store the pointer in
> +   EMPTY_BB.  Return the decl of the resolver function.  */
> +
> +static tree
> +make_resolver_func (const tree default_decl,
> +		    const tree ifunc_alias_decl,
> +		    basic_block *empty_bb)
> +{
> +  tree decl, type, t;
> +
> +  /* Create resolver function name based on default_decl.  */
> +  tree decl_name = clone_function_name (default_decl, "resolver");
> +  const char *resolver_name = IDENTIFIER_POINTER (decl_name);
> +
> +  /* The resolver function should return a (void *). */
> +  type = build_function_type_list (ptr_type_node, NULL_TREE);
> +
> +  decl = build_fn_decl (resolver_name, type);
> +  SET_DECL_ASSEMBLER_NAME (decl, decl_name);
> +
> +  DECL_NAME (decl) = decl_name;
> +  TREE_USED (decl) = 1;
> +  DECL_ARTIFICIAL (decl) = 1;
> +  DECL_IGNORED_P (decl) = 1;
> +  TREE_PUBLIC (decl) = 0;
> +  DECL_UNINLINABLE (decl) = 1;
> +
> +  /* Resolver is not external, body is generated.  */
> +  DECL_EXTERNAL (decl) = 0;
> +  DECL_EXTERNAL (ifunc_alias_decl) = 0;
> +
> +  DECL_CONTEXT (decl) = NULL_TREE;
> +  DECL_INITIAL (decl) = make_node (BLOCK);
> +  DECL_STATIC_CONSTRUCTOR (decl) = 0;
> +
> +  if (DECL_COMDAT_GROUP (default_decl)
> +      || TREE_PUBLIC (default_decl))
> +    {
> +      /* In this case, each translation unit with a call to this
> +	 versioned function will put out a resolver.  Ensure it
> +	 is comdat to keep just one copy.  */
> +      DECL_COMDAT (decl) = 1;
> +      make_decl_one_only (decl, DECL_ASSEMBLER_NAME (decl));
> +    }
> +  else
> +    TREE_PUBLIC (ifunc_alias_decl) = 0;
> +
> +  /* Build result decl and add to function_decl. */
> +  t = build_decl (UNKNOWN_LOCATION, RESULT_DECL, NULL_TREE, ptr_type_node);
> +  DECL_CONTEXT (t) = decl;
> +  DECL_ARTIFICIAL (t) = 1;
> +  DECL_IGNORED_P (t) = 1;
> +  DECL_RESULT (decl) = t;
> +
> +  gimplify_function_tree (decl);
> +  push_cfun (DECL_STRUCT_FUNCTION (decl));
> +  *empty_bb = init_lowered_empty_function (decl, false,
> +					   profile_count::uninitialized ());
> +
> +  cgraph_node::add_new_function (decl, true);
> +  symtab->call_cgraph_insertion_hooks (cgraph_node::get_create (decl));
> +
> +  pop_cfun ();
> +
> +  gcc_assert (ifunc_alias_decl != NULL);
> +  /* Mark ifunc_alias_decl as "ifunc" with resolver as resolver_name.  */
> +  DECL_ATTRIBUTES (ifunc_alias_decl)
> +    = make_attribute ("ifunc", resolver_name,
> +		      DECL_ATTRIBUTES (ifunc_alias_decl));
> +
> +  /* Create the alias for dispatch to resolver here.  */
> +  cgraph_node::create_same_body_alias (ifunc_alias_decl, decl);
> +  return decl;
> +}

It looks like this is identical to the x86 version, is that right?
If so, I think it'd be worth moving it to target-independent code.

> +
> +/* This adds a condition to the basic_block NEW_BB in function FUNCTION_DECL
> +   to return a pointer to VERSION_DECL if all feature bits specified in
> +   FEATURE_MASK are not set in MASK_VAR.  This function will be called during
> +   version dispatch to decide which function version to execute.  It returns
> +   the basic block at the end, to which more conditions can be added.  */
> +static basic_block
> +add_condition_to_bb (tree function_decl, tree version_decl,
> +		     unsigned long long feature_mask,
> +		     tree mask_var, basic_block new_bb)
> +{
> +  gimple *return_stmt;
> +  tree convert_expr, result_var;
> +  gimple *convert_stmt;
> +  gimple *if_else_stmt;
> +
> +  basic_block bb1, bb2, bb3;
> +  edge e12, e23;
> +
> +  gimple_seq gseq;
> +
> +  push_cfun (DECL_STRUCT_FUNCTION (function_decl));
> +
> +  gcc_assert (new_bb != NULL);
> +  gseq = bb_seq (new_bb);
> +
> +
> +  convert_expr = build1 (CONVERT_EXPR, ptr_type_node,
> +			 build_fold_addr_expr (version_decl));
> +  result_var = create_tmp_var (ptr_type_node);
> +  convert_stmt = gimple_build_assign (result_var, convert_expr);
> +  return_stmt = gimple_build_return (result_var);
> +
> +
> +  if (feature_mask == 0)
> +    {
> +      /* Default version.  */
> +      gimple_seq_add_stmt (&gseq, convert_stmt);
> +      gimple_seq_add_stmt (&gseq, return_stmt);
> +      set_bb_seq (new_bb, gseq);
> +      gimple_set_bb (convert_stmt, new_bb);
> +      gimple_set_bb (return_stmt, new_bb);
> +      pop_cfun ();
> +      return new_bb;
> +    }
> +
> +  tree and_expr_var = create_tmp_var (long_long_unsigned_type_node);
> +  tree and_expr = build2 (BIT_AND_EXPR,
> +			  long_long_unsigned_type_node,
> +			  mask_var,
> +			  build_int_cst (long_long_unsigned_type_node,
> +					 feature_mask));
> +  gimple *and_stmt = gimple_build_assign (and_expr_var, and_expr);
> +  gimple_set_block (and_stmt, DECL_INITIAL (function_decl));
> +  gimple_set_bb (and_stmt, new_bb);
> +  gimple_seq_add_stmt (&gseq, and_stmt);
> +
> +  tree zero_llu = build_int_cst (long_long_unsigned_type_node, 0);
> +  if_else_stmt = gimple_build_cond (EQ_EXPR, and_expr_var, zero_llu,
> +				    NULL_TREE, NULL_TREE);
> +  gimple_set_block (if_else_stmt, DECL_INITIAL (function_decl));
> +  gimple_set_bb (if_else_stmt, new_bb);
> +  gimple_seq_add_stmt (&gseq, if_else_stmt);
> +
> +  gimple_seq_add_stmt (&gseq, convert_stmt);
> +  gimple_seq_add_stmt (&gseq, return_stmt);
> +  set_bb_seq (new_bb, gseq);
> +
> +  bb1 = new_bb;
> +  e12 = split_block (bb1, if_else_stmt);
> +  bb2 = e12->dest;
> +  e12->flags &= ~EDGE_FALLTHRU;
> +  e12->flags |= EDGE_TRUE_VALUE;
> +
> +  e23 = split_block (bb2, return_stmt);
> +
> +  gimple_set_bb (convert_stmt, bb2);
> +  gimple_set_bb (return_stmt, bb2);
> +
> +  bb3 = e23->dest;
> +  make_edge (bb1, bb3, EDGE_FALSE_VALUE);
> +
> +  remove_edge (e23);
> +  make_edge (bb2, EXIT_BLOCK_PTR_FOR_FN (cfun), 0);
> +
> +  pop_cfun ();
> +
> +  return bb3;
> +}

It looks like there's quite a bit of common code here too,
but also significant differences.

Perhaps the target-indendent code could contain a base class for generating
resolvers that targets can customise via derivation?  Just an idea though.
Maybe the amount of common code beyond make_resolver_func doesn't make
it worth it.

> +
> +/* Used when sorting the decls into dispatch order.  */
> +static int compare_feature_version_info (const void *p1, const void *p2)
> +{
> +  typedef struct _function_version_info
> +    {
> +      tree version_decl;
> +      unsigned long long feature_mask;
> +    } function_version_info;

There's no need for the typedef struct construct now that the
codebase is C++.

> +  const function_version_info v1 = *(const function_version_info *)p1;
> +  const function_version_info v2 = *(const function_version_info *)p2;
> +  return - compare_feature_masks (v1.feature_mask, v2.feature_mask);
> +}
> +
> +static int
> +dispatch_function_versions (tree dispatch_decl,
> +			    void *fndecls_p,
> +			    basic_block *empty_bb)
> +{
> +  gimple *ifunc_cpu_init_stmt;
> +  gimple_seq gseq;
> +  int ix;
> +  tree ele;
> +  vec<tree> *fndecls;
> +  unsigned int num_versions = 0;
> +  unsigned int actual_versions = 0;
> +  unsigned int i;
> +
> +  struct _function_version_info
> +    {
> +      tree version_decl;
> +      unsigned long long feature_mask;
> +    }*function_version_info;
> +
> +  gcc_assert (dispatch_decl != NULL
> +	      && fndecls_p != NULL
> +	      && empty_bb != NULL);
> +
> +  /*fndecls_p is actually a vector.  */
> +  fndecls = static_cast<vec<tree> *> (fndecls_p);
> +
> +  /* At least one more version other than the default.  */
> +  num_versions = fndecls->length ();
> +  gcc_assert (num_versions >= 2);
> +
> +  function_version_info = (struct _function_version_info *)
> +    XNEWVEC (struct _function_version_info, (num_versions));
> +
> +  push_cfun (DECL_STRUCT_FUNCTION (dispatch_decl));
> +
> +  gseq = bb_seq (*empty_bb);
> +  /* Function version dispatch is via IFUNC.  IFUNC resolvers fire before
> +     constructors, so explicity call __builtin_cpu_init here.  */

Comment needs updating for AArch64 (but I realise you said in the
covering note that this wasn't final).

> +  tree init_fn_type = build_function_type_list (void_type_node, NULL);
> +  tree init_fn_decl = build_decl (UNKNOWN_LOCATION, FUNCTION_DECL,
> +				  get_identifier ("init_cpu_features"),
> +				  init_fn_type);
> +  ifunc_cpu_init_stmt = gimple_build_call (init_fn_decl, 0);
> +  gimple_seq_add_stmt (&gseq, ifunc_cpu_init_stmt);
> +  gimple_set_bb (ifunc_cpu_init_stmt, *empty_bb);
> +
> +  /* Build the struct type for __aarch64_cpu_features.  */
> +  tree global_type = lang_hooks.types.make_type (RECORD_TYPE);
> +  tree field1 = build_decl (UNKNOWN_LOCATION, FIELD_DECL,
> +			    get_identifier ("features"),
> +			    long_long_unsigned_type_node);
> +  DECL_FIELD_CONTEXT (field1) = global_type;
> +  TYPE_FIELDS (global_type) = field1;
> +  layout_type (global_type);
> +
> +  tree global_var = build_decl (UNKNOWN_LOCATION, VAR_DECL, get_identifier
> +				("__aarch64_cpu_features"), global_type);
> +  DECL_EXTERNAL (global_var) = 1;
> +  tree mask_var = create_tmp_var (long_long_unsigned_type_node);
> +
> +  tree component_expr = build3 (COMPONENT_REF, long_long_unsigned_type_node,
> +				global_var, field1, NULL_TREE);
> +  gimple *component_stmt = gimple_build_assign (mask_var, component_expr);
> +  gimple_set_block (component_stmt, DECL_INITIAL (dispatch_decl));
> +  gimple_set_bb (component_stmt, *empty_bb);
> +  gimple_seq_add_stmt (&gseq, component_stmt);
> +
> +  tree not_expr = build1 (BIT_NOT_EXPR, long_long_unsigned_type_node, mask_var);
> +  gimple *not_stmt = gimple_build_assign (mask_var, not_expr);
> +  gimple_set_block (not_stmt, DECL_INITIAL (dispatch_decl));
> +  gimple_set_bb (not_stmt, *empty_bb);
> +  gimple_seq_add_stmt (&gseq, not_stmt);
> +
> +  set_bb_seq (*empty_bb, gseq);
> +
> +  pop_cfun ();
> +
> +  for (ix = 0; fndecls->iterate (ix, &ele); ++ix)

Very minor, but we can use:

  for (tree version_decl : *fndecls)

now.

Thanks,
Richard

> +    {
> +      tree version_decl = ele;
> +      unsigned long long feature_mask;
> +      /* Get attribute string, parse it and find the right features.  */
> +      feature_mask = get_feature_mask_for_version (version_decl);
> +      function_version_info [actual_versions].version_decl = version_decl;
> +      function_version_info [actual_versions].feature_mask = feature_mask;
> +      actual_versions++;
> +    }
> +
> +  /* Sort the versions according to descending order of dispatch priority.  */
> +  qsort (function_version_info, actual_versions,
> +	 sizeof (struct _function_version_info), compare_feature_version_info);
> +
> +  for (i = 0; i < actual_versions; ++i)
> +    *empty_bb = add_condition_to_bb (dispatch_decl,
> +				     function_version_info[i].version_decl,
> +				     function_version_info[i].feature_mask,
> +				     mask_var,
> +				     *empty_bb);
> +
> +  free (function_version_info);
> +  return 0;
> +}
> +
> +
> +tree
> +aarch64_generate_version_dispatcher_body (void *node_p)
> +{
> +  tree resolver_decl;
> +  basic_block empty_bb;
> +  tree default_ver_decl;
> +  struct cgraph_node *versn;
> +  struct cgraph_node *node;
> +
> +  struct cgraph_function_version_info *node_version_info = NULL;
> +  struct cgraph_function_version_info *versn_info = NULL;
> +
> +  node = (cgraph_node *)node_p;
> +
> +  node_version_info = node->function_version ();
> +  gcc_assert (node->dispatcher_function
> +	      && node_version_info != NULL);
> +
> +  if (node_version_info->dispatcher_resolver)
> +    return node_version_info->dispatcher_resolver;
> +
> +  /* The first version in the chain corresponds to the default version.  */
> +  default_ver_decl = node_version_info->next->this_node->decl;
> +
> +  /* node is going to be an alias, so remove the finalized bit.  */
> +  node->definition = false;
> +
> +  resolver_decl = make_resolver_func (default_ver_decl,
> +				      node->decl, &empty_bb);
> +
> +  node_version_info->dispatcher_resolver = resolver_decl;
> +
> +  push_cfun (DECL_STRUCT_FUNCTION (resolver_decl));
> +
> +  auto_vec<tree, 2> fn_ver_vec;
> +
> +  for (versn_info = node_version_info->next; versn_info;
> +       versn_info = versn_info->next)
> +    {
> +      versn = versn_info->this_node;
> +      /* Check for virtual functions here again, as by this time it should
> +	 have been determined if this function needs a vtable index or
> +	 not.  This happens for methods in derived classes that override
> +	 virtual methods in base classes but are not explicitly marked as
> +	 virtual.  */
> +      if (DECL_VINDEX (versn->decl))
> +	sorry ("virtual function multiversioning not supported");
> +
> +      fn_ver_vec.safe_push (versn->decl);
> +    }
> +
> +  dispatch_function_versions (resolver_decl, &fn_ver_vec, &empty_bb);
> +  cgraph_edge::rebuild_edges ();
> +  pop_cfun ();
> +  return resolver_decl;
> +}
> +
> +/* Make a dispatcher declaration for the multi-versioned function DECL.
> +   Calls to DECL function will be replaced with calls to the dispatcher
> +   by the front-end.  Returns the decl of the dispatcher function.  */
> +
> +tree
> +aarch64_get_function_versions_dispatcher (void *decl)
> +{
> +  tree fn = (tree) decl;
> +  struct cgraph_node *node = NULL;
> +  struct cgraph_node *default_node = NULL;
> +  struct cgraph_function_version_info *node_v = NULL;
> +  struct cgraph_function_version_info *first_v = NULL;
> +
> +  tree dispatch_decl = NULL;
> +
> +  struct cgraph_function_version_info *default_version_info = NULL;
> +
> +  gcc_assert (fn != NULL && DECL_FUNCTION_VERSIONED (fn));
> +
> +  node = cgraph_node::get (fn);
> +  gcc_assert (node != NULL);
> +
> +  node_v = node->function_version ();
> +  gcc_assert (node_v != NULL);
> +
> +  if (node_v->dispatcher_resolver != NULL)
> +    return node_v->dispatcher_resolver;
> +
> +  /* Find the default version and make it the first node.  */
> +  first_v = node_v;
> +  /* Go to the beginning of the chain.  */
> +  while (first_v->prev != NULL)
> +    first_v = first_v->prev;
> +  default_version_info = first_v;
> +  while (default_version_info != NULL)
> +    {
> +      if (get_feature_mask_for_version
> +	    (default_version_info->this_node->decl) == 0ULL)
> +	break;
> +      default_version_info = default_version_info->next;
> +    }
> +
> +  /* If there is no default node, just return NULL.  */
> +  if (default_version_info == NULL)
> +    return NULL;
> +
> +  /* Make default info the first node.  */
> +  if (first_v != default_version_info)
> +    {
> +      default_version_info->prev->next = default_version_info->next;
> +      if (default_version_info->next)
> +	default_version_info->next->prev = default_version_info->prev;
> +      first_v->prev = default_version_info;
> +      default_version_info->next = first_v;
> +      default_version_info->prev = NULL;
> +    }
> +
> +  default_node = default_version_info->this_node;
> +
> +  if (targetm.has_ifunc_p ())
> +    {
> +      struct cgraph_function_version_info *it_v = NULL;
> +      struct cgraph_node *dispatcher_node = NULL;
> +      struct cgraph_function_version_info *dispatcher_version_info = NULL;
> +
> +      /* Right now, the dispatching is done via ifunc.  */
> +      dispatch_decl = make_dispatcher_decl (default_node->decl);
> +      TREE_NOTHROW (dispatch_decl) = TREE_NOTHROW (fn);
> +
> +      dispatcher_node = cgraph_node::get_create (dispatch_decl);
> +      gcc_assert (dispatcher_node != NULL);
> +      dispatcher_node->dispatcher_function = 1;
> +      dispatcher_version_info
> +	= dispatcher_node->insert_new_function_version ();
> +      dispatcher_version_info->next = default_version_info;
> +      dispatcher_node->definition = 1;
> +
> +      /* Set the dispatcher for all the versions.  */
> +      it_v = default_version_info;
> +      while (it_v != NULL)
> +	{
> +	  it_v->dispatcher_resolver = dispatch_decl;
> +	  it_v = it_v->next;
> +	}
> +    }
> +  else
> +    {
> +      error_at (DECL_SOURCE_LOCATION (default_node->decl),
> +		"multiversioning needs %<ifunc%> which is not supported "
> +		"on this target");
> +    }
> +
> +  return dispatch_decl;
> +}
> +
> +bool
> +aarch64_common_function_versions (tree fn1, tree fn2)
> +{
> +  if (TREE_CODE (fn1) != FUNCTION_DECL
> +      || TREE_CODE (fn2) != FUNCTION_DECL)
> +    return false;
> +
> +  return (aarch64_compare_version_priority (fn1, fn2) != 0);
> +}
> +
> +
> +tree
> +aarch64_mangle_decl_assembler_name (tree decl, tree id)
> +{
> +  /* For function version, add the target suffix to the assembler name.  */
> +  if (TREE_CODE (decl) == FUNCTION_DECL
> +      && DECL_FUNCTION_VERSIONED (decl))
> +    {
> +      unsigned long long feature_mask = get_feature_mask_for_version (decl);
> +
> +      /* No suffix for the default version.  */
> +      if (feature_mask == 0ULL)
> +	return id;
> +
> +      char suffix[2048];
> +      int pos = 0;
> +      const char *base = IDENTIFIER_POINTER (id);
> +
> +      for (int i = 1; i < FEAT_MAX; i++)
> +	{
> +	  if (feature_mask & aarch64_fmv_feature_data[i].feature_mask)
> +	    {
> +	      suffix[pos] = 'M';
> +	      strcpy (&suffix[pos+1], aarch64_fmv_feature_data[i].name);
> +	      pos += strlen(aarch64_fmv_feature_data[i].name) + 1;
> +	    }
> +	}
> +      suffix[pos] = '\0';
> +
> +      char *ret = XNEWVEC (char, strlen (base) + strlen (suffix) + 3);
> +      sprintf (ret, "%s._%s", base, suffix);
> +
> +      if (DECL_ASSEMBLER_NAME_SET_P (decl))
> +	SET_DECL_RTL (decl, NULL);
> +
> +      id = get_identifier (ret);
> +    }
> +  return id;
> +}
> +
> +
>  /* Helper for aarch64_can_inline_p.  In the case where CALLER and CALLEE are
>     tri-bool options (yes, no, don't care) and the default value is
>     DEF, determine whether to reject inlining.  */
> @@ -27804,6 +28614,12 @@ aarch64_libgcc_floating_mode_supported_p
>  #undef TARGET_OPTION_VALID_ATTRIBUTE_P
>  #define TARGET_OPTION_VALID_ATTRIBUTE_P aarch64_option_valid_attribute_p
>  
> +#undef TARGET_OPTION_VALID_VERSION_ATTRIBUTE_P
> +#define TARGET_OPTION_VALID_VERSION_ATTRIBUTE_P aarch64_option_valid_attribute_p
> +
> +#undef TARGET_OPTION_EXPANDED_CLONES_ATTRIBUTE
> +#define TARGET_OPTION_EXPANDED_CLONES_ATTRIBUTE "target_version"
> +
>  #undef TARGET_SET_CURRENT_FUNCTION
>  #define TARGET_SET_CURRENT_FUNCTION aarch64_set_current_function
>  
> @@ -28128,6 +28944,24 @@ aarch64_libgcc_floating_mode_supported_p
>  #undef TARGET_CONST_ANCHOR
>  #define TARGET_CONST_ANCHOR 0x1000000
>  
> +#undef TARGET_OPTION_FUNCTION_VERSIONS
> +#define TARGET_OPTION_FUNCTION_VERSIONS aarch64_common_function_versions
> +
> +#undef TARGET_COMPARE_VERSION_PRIORITY
> +#define TARGET_COMPARE_VERSION_PRIORITY aarch64_compare_version_priority
> +
> +#undef TARGET_GENERATE_VERSION_DISPATCHER_BODY
> +#define TARGET_GENERATE_VERSION_DISPATCHER_BODY \
> +  aarch64_generate_version_dispatcher_body
> +
> +#undef TARGET_GET_FUNCTION_VERSIONS_DISPATCHER
> +#define TARGET_GET_FUNCTION_VERSIONS_DISPATCHER \
> +  aarch64_get_function_versions_dispatcher
> +
> +#undef TARGET_MANGLE_DECL_ASSEMBLER_NAME
> +#define TARGET_MANGLE_DECL_ASSEMBLER_NAME aarch64_mangle_decl_assembler_name
> +
> +
>  struct gcc_target targetm = TARGET_INITIALIZER;
>  
>  #include "gt-aarch64.h"