Disable a broken multiversioning optimisation

  This patch skips redirect_to_specific clone for aarch64 and riscv,
because the optimisation has two flaws:

1. It checks the value of the "target" attribute, even on targets that
don't use this attribute for multiversioning.

2. The algorithm used is too aggressive, and will eliminate the
indirection in some cases where the runtime choice of callee version
can't be determined statically at compile time.  A correct would need to
verify that:
 - if the current caller version were selected at runtime, then the
   chosen callee version would be eligible for selection.
 - if any higher priority callee version were selected at runtime, then
   a higher priority caller version would have been eligble for
   selection (and hence the current caller version wouldn't have been
   selected).

The current checks only verify a more restrictive version of the first
condition, and don't check the second condition at all.

Fixing the optimisation properly would require implementing target hooks
to check for implications between version attributes, which is too
complicated for this stage.  However, I would like to see this hook
implemented in the future, since it could also help deduplicate other
multiversioning code.

Since this behaviour has existed for x86 and powerpc for a while, I
think it's best to preserve the existing behaviour on those targets,
unless any maintainer for those targets disagrees.

----

There was a previous patch by Yangyu Chen [1] which addressed flaw 1 but
overlooked flaw 2.  I pointed out flaw 2 at the time; as far as I can tell,
there haven't been any further patches proposed for this issue (apologies if I
missed something; I've not checked the mailing lists particularly thoroughly).

I'm currently bootstrap and regression testing this on aarch64. Assuming thes
tests pass, is this ok for master and backport to gcc-14?

[1] https://gcc.gnu.org/pipermail/gcc-patches/2024-November/667527.html

gcc/ChangeLog:

	* multiple_target.cc
	(redirect_to_specific_clone): Assert that "target" attribute is
	used for FMV before checking it.
	(ipa_target_clone): Skip redirect_to_specific_clone on some
	targets.

gcc/testsuite/ChangeLog:

	* g++.target/aarch64/mv-pragma.C: New test.
  

Andrew Carlotti <andrew.carlotti@arm.com> writes:
> This patch skips redirect_to_specific clone for aarch64 and riscv,
> because the optimisation has two flaws:
>
> 1. It checks the value of the "target" attribute, even on targets that
> don't use this attribute for multiversioning.
>
> 2. The algorithm used is too aggressive, and will eliminate the
> indirection in some cases where the runtime choice of callee version
> can't be determined statically at compile time.  A correct would need to
> verify that:
>  - if the current caller version were selected at runtime, then the
>    chosen callee version would be eligible for selection.
>  - if any higher priority callee version were selected at runtime, then
>    a higher priority caller version would have been eligble for
>    selection (and hence the current caller version wouldn't have been
>    selected).
>
> The current checks only verify a more restrictive version of the first
> condition, and don't check the second condition at all.
>
> Fixing the optimisation properly would require implementing target hooks
> to check for implications between version attributes, which is too
> complicated for this stage.  However, I would like to see this hook
> implemented in the future, since it could also help deduplicate other
> multiversioning code.

Yeah, agreed on both counts.  And I think we should try to design the
hooks to minimise cut-&-paste between targets (possibly for target_version
only).

> Since this behaviour has existed for x86 and powerpc for a while, I
> think it's best to preserve the existing behaviour on those targets,
> unless any maintainer for those targets disagrees.
>
> ----
>
> There was a previous patch by Yangyu Chen [1] which addressed flaw 1 but
> overlooked flaw 2.  I pointed out flaw 2 at the time; as far as I can tell,
> there haven't been any further patches proposed for this issue (apologies if I
> missed something; I've not checked the mailing lists particularly thoroughly).
>
> I'm currently bootstrap and regression testing this on aarch64. Assuming thes
> tests pass, is this ok for master and backport to gcc-14?
>
> [1] https://gcc.gnu.org/pipermail/gcc-patches/2024-November/667527.html
>
> gcc/ChangeLog:
>
> 	* multiple_target.cc
> 	(redirect_to_specific_clone): Assert that "target" attribute is
> 	used for FMV before checking it.
> 	(ipa_target_clone): Skip redirect_to_specific_clone on some
> 	targets.
>
> gcc/testsuite/ChangeLog:
>
> 	* g++.target/aarch64/mv-pragma.C: New test.

The patch LGTM apart from nits:

> diff --git a/gcc/multiple_target.cc b/gcc/multiple_target.cc
> index 552b9626aa7173546d3b027a2b1a3d227ad25375..d8becf4d9a9626a4ab3688ab30431487a6db770f 100644
> --- a/gcc/multiple_target.cc
> +++ b/gcc/multiple_target.cc
> @@ -442,7 +442,14 @@ expand_target_clones (struct cgraph_node *node, bool definition)
>  
>  /* When NODE is a target clone, consider all callees and redirect
>     to a clone with equal target attributes.  That prevents multiple
> -   multi-versioning dispatches and a call-chain can be optimized.  */
> +   multi-versioning dispatches and a call-chain can be optimized.
> +
> +   This optimisation might pick the wrong version in some cases, since knowing
> +   that we meet the target requirements for a matching callee version does not
> +   tell us that we won't also meet the target requirements for a higher
> +   priority callee version at runtime.  Since this is longstanding behaviour

"behavior"

> +   for x86 and powerpc, we preserve it for those targets, but skip the optimisation
> +   for targets that use the "target_version" attribute for multi-versioning.  */

two long lines

Please give the RISC-V folks a day or so to comment, but otherwise this
is OK for trunk and GCC 14 with the comments above fixed.

Thanks,
Richard

>  
>  static void
>  redirect_to_specific_clone (cgraph_node *node)
> @@ -451,6 +458,7 @@ redirect_to_specific_clone (cgraph_node *node)
>    if (fv == NULL)
>      return;
>  
> +  gcc_assert (TARGET_HAS_FMV_TARGET_ATTRIBUTE);
>    tree attr_target = lookup_attribute ("target", DECL_ATTRIBUTES (node->decl));
>    if (attr_target == NULL_TREE)
>      return;
> @@ -503,8 +511,9 @@ ipa_target_clone (void)
>    for (unsigned i = 0; i < to_dispatch.length (); i++)
>      create_dispatcher_calls (to_dispatch[i]);
>  
> -  FOR_EACH_FUNCTION (node)
> -    redirect_to_specific_clone (node);
> +  if (TARGET_HAS_FMV_TARGET_ATTRIBUTE)
> +    FOR_EACH_FUNCTION (node)
> +      redirect_to_specific_clone (node);
>  
>    return 0;
>  }
> diff --git a/gcc/testsuite/g++.target/aarch64/mv-pragma.C b/gcc/testsuite/g++.target/aarch64/mv-pragma.C
> new file mode 100644
> index 0000000000000000000000000000000000000000..545d0735438d4774a2cb909f28584ddbf4f54d2a
> --- /dev/null
> +++ b/gcc/testsuite/g++.target/aarch64/mv-pragma.C
> @@ -0,0 +1,31 @@
> +/* { dg-do compile } */
> +/* { dg-require-ifunc "" } */
> +/* { dg-options "-O0" } */
> +
> +#pragma GCC target ("+sve")
> +
> +__attribute__((target_version("default")))
> +int foo ()
> +{
> +  return 1;
> +}
> +
> +__attribute__((target_version("sve2")))
> +int foo ()
> +{
> +  return 2;
> +}
> +
> +__attribute__((target_version("default")))
> +int bar ()
> +{
> +  return foo();
> +}
> +
> +__attribute__((target_version("sha3")))
> +int bar ()
> +{
> +  return foo() + 5;
> +}
> +
> +/* { dg-final { scan-assembler-times "\n\tbl\t_Z3foov\n" 2 } } */
  

On 1/8/25 19:48, Andrew Carlotti wrote:
> This patch skips redirect_to_specific clone for aarch64 and riscv,
> because the optimisation has two flaws:
> 
> 1. It checks the value of the "target" attribute, even on targets that
> don't use this attribute for multiversioning.
> 
> 2. The algorithm used is too aggressive, and will eliminate the
> indirection in some cases where the runtime choice of callee version
> can't be determined statically at compile time.  A correct would need to
> verify that:
>   - if the current caller version were selected at runtime, then the
>     chosen callee version would be eligible for selection.
>   - if any higher priority callee version were selected at runtime, then
>     a higher priority caller version would have been eligble for
>     selection (and hence the current caller version wouldn't have been
>     selected).
> 
> The current checks only verify a more restrictive version of the first
> condition, and don't check the second condition at all.
> 
> Fixing the optimisation properly would require implementing target hooks
> to check for implications between version attributes, which is too
> complicated for this stage.  However, I would like to see this hook
> implemented in the future, since it could also help deduplicate other
> multiversioning code.
> 
> Since this behaviour has existed for x86 and powerpc for a while, I
> think it's best to preserve the existing behaviour on those targets,
> unless any maintainer for those targets disagrees.
> 
> ----
> 
> There was a previous patch by Yangyu Chen [1] which addressed flaw 1 but
> overlooked flaw 2.  I pointed out flaw 2 at the time; as far as I can tell,
> there haven't been any further patches proposed for this issue (apologies if I
> missed something; I've not checked the mailing lists particularly thoroughly).
> 
> I'm currently bootstrap and regression testing this on aarch64. Assuming thes
> tests pass, is this ok for master and backport to gcc-14?
> 
> [1] https://gcc.gnu.org/pipermail/gcc-patches/2024-November/667527.html
> 
> gcc/ChangeLog:
> 
> 	* multiple_target.cc
> 	(redirect_to_specific_clone): Assert that "target" attribute is
> 	used for FMV before checking it.
> 	(ipa_target_clone): Skip redirect_to_specific_clone on some
> 	targets.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.target/aarch64/mv-pragma.C: New test.
> 
> 
> diff --git a/gcc/multiple_target.cc b/gcc/multiple_target.cc
> index 552b9626aa7173546d3b027a2b1a3d227ad25375..d8becf4d9a9626a4ab3688ab30431487a6db770f 100644
> --- a/gcc/multiple_target.cc
> +++ b/gcc/multiple_target.cc
> @@ -442,7 +442,14 @@ expand_target_clones (struct cgraph_node *node, bool definition)
>   
>   /* When NODE is a target clone, consider all callees and redirect
>      to a clone with equal target attributes.  That prevents multiple
> -   multi-versioning dispatches and a call-chain can be optimized.  */
> +   multi-versioning dispatches and a call-chain can be optimized.
> +
> +   This optimisation might pick the wrong version in some cases, since knowing
> +   that we meet the target requirements for a matching callee version does not
> +   tell us that we won't also meet the target requirements for a higher
> +   priority callee version at runtime.  Since this is longstanding behaviour
> +   for x86 and powerpc, we preserve it for those targets, but skip the optimisation
> +   for targets that use the "target_version" attribute for multi-versioning.  */
>   
>   static void
>   redirect_to_specific_clone (cgraph_node *node)
> @@ -451,6 +458,7 @@ redirect_to_specific_clone (cgraph_node *node)
>     if (fv == NULL)
>       return;
>   
> +  gcc_assert (TARGET_HAS_FMV_TARGET_ATTRIBUTE);
>     tree attr_target = lookup_attribute ("target", DECL_ATTRIBUTES (node->decl));
>     if (attr_target == NULL_TREE)
>       return;
> @@ -503,8 +511,9 @@ ipa_target_clone (void)
>     for (unsigned i = 0; i < to_dispatch.length (); i++)
>       create_dispatcher_calls (to_dispatch[i]);
>   
> -  FOR_EACH_FUNCTION (node)
> -    redirect_to_specific_clone (node);
> +  if (TARGET_HAS_FMV_TARGET_ATTRIBUTE)
> +    FOR_EACH_FUNCTION (node)
> +      redirect_to_specific_clone (node);
>   
>     return 0;
>   }
> diff --git a/gcc/testsuite/g++.target/aarch64/mv-pragma.C b/gcc/testsuite/g++.target/aarch64/mv-pragma.C
> new file mode 100644
> index 0000000000000000000000000000000000000000..545d0735438d4774a2cb909f28584ddbf4f54d2a
> --- /dev/null
> +++ b/gcc/testsuite/g++.target/aarch64/mv-pragma.C
> @@ -0,0 +1,31 @@
> +/* { dg-do compile } */
> +/* { dg-require-ifunc "" } */
> +/* { dg-options "-O0" } */
> +
> +#pragma GCC target ("+sve")
> +
> +__attribute__((target_version("default")))
> +int foo ()
> +{
> +  return 1;
> +}
> +
> +__attribute__((target_version("sve2")))
> +int foo ()
> +{
> +  return 2;
> +}
> +
> +__attribute__((target_version("default")))
> +int bar ()
> +{
> +  return foo();
> +}
> +
> +__attribute__((target_version("sha3")))
> +int bar ()
> +{
> +  return foo() + 5;
> +}
> +
> +/* { dg-final { scan-assembler-times "\n\tbl\t_Z3foov\n" 2 } } */

Reviewed-by: Yangyu Chen <cyy@cyyself.name>

The code functional change and test look good to me. Please edit the 
commit message as Richard Sandiford said. I'm willing to implement that 
target hook for RISC-V and aarch64 in 1-2 months to regain this feature 
for better performance.

Thanks,
Yangyu Chen