[clang] Function Multi Versioning supports IFunc lowerings on Darwin platforms

[jroelofs]

No description provided.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Jon Roelofs (jroelofs)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/73688.diff

2 Files Affected:

• (modified) clang/include/clang/Basic/TargetInfo.h (+2)
• (modified) clang/test/CodeGen/attr-target-mv-va-args.c (+19)

diff --git a/clang/include/clang/Basic/TargetInfo.h b/clang/include/clang/Basic/TargetInfo.h
index 41f3c2e403cbef6..1fe2a18cd5dc9cc 100644
--- a/clang/include/clang/Basic/TargetInfo.h
+++ b/clang/include/clang/Basic/TargetInfo.h
@@ -1424,6 +1424,8 @@ class TargetInfo : public TransferrableTargetInfo,
 
   /// Identify whether this target supports IFuncs.
   bool supportsIFunc() const {
+    if (getTriple().isOSBinFormatMachO())
+      return true;
     return getTriple().isOSBinFormatELF() &&
            ((getTriple().isOSLinux() && !getTriple().isMusl()) ||
             getTriple().isOSFreeBSD());
diff --git a/clang/test/CodeGen/attr-target-mv-va-args.c b/clang/test/CodeGen/attr-target-mv-va-args.c
index 96821c610235bdc..dbf5a74205c4c19 100644
--- a/clang/test/CodeGen/attr-target-mv-va-args.c
+++ b/clang/test/CodeGen/attr-target-mv-va-args.c
@@ -3,6 +3,7 @@
 // RUN: %clang_cc1 -triple x86_64-windows-pc -emit-llvm %s -o - | FileCheck %s --check-prefixes=NO-IFUNC,WINDOWS
 // RUN: %clang_cc1 -triple x86_64-linux-musl -emit-llvm %s -o - | FileCheck %s --check-prefixes=NO-IFUNC,NO-IFUNC-ELF
 // RUN: %clang_cc1 -triple x86_64-fuchsia -emit-llvm %s -o - | FileCheck %s --check-prefixes=NO-IFUNC,NO-IFUNC-ELF
+// RUN: %clang_cc1 -triple x86_64-apple-macho -emit-llvm %s -o - | FileCheck %s --check-prefix=IFUNC-MACHO
 int __attribute__((target("sse4.2"))) foo(int i, ...) { return 0; }
 int __attribute__((target("arch=sandybridge"))) foo(int i, ...);
 int __attribute__((target("arch=ivybridge"))) foo(int i, ...) {return 1;}
@@ -30,6 +31,24 @@ int bar(void) {
 // IFUNC-ELF: ret ptr @foo
 // IFUNC-ELF: declare i32 @foo.arch_sandybridge(i32 noundef, ...)
 
+// IFUNC-MACHO: @foo.ifunc = weak_odr ifunc i32 (i32, ...), ptr @foo.resolver
+// IFUNC-MACHO: define{{.*}} i32 @foo.sse4.2(i32 noundef %i, ...)
+// IFUNC-MACHO: ret i32 0
+// IFUNC-MACHO: define{{.*}} i32 @foo.arch_ivybridge(i32 noundef %i, ...)
+// IFUNC-MACHO: ret i32 1
+// IFUNC-MACHO: define{{.*}} i32 @foo(i32 noundef %i, ...)
+// IFUNC-MACHO: ret i32 2
+// IFUNC-MACHO: define{{.*}} i32 @bar()
+// IFUNC-MACHO: call i32 (i32, ...) @foo.ifunc(i32 noundef 1, i32 noundef 97, double
+// IFUNC-MACHO: call i32 (i32, ...) @foo.ifunc(i32 noundef 2, double noundef 2.2{{[0-9Ee+]+}}, ptr noundef
+
+// IFUNC-MACHO: define weak_odr ptr @foo.resolver()
+// IFUNC-MACHO: ret ptr @foo.arch_sandybridge
+// IFUNC-MACHO: ret ptr @foo.arch_ivybridge
+// IFUNC-MACHO: ret ptr @foo.sse4.2
+// IFUNC-MACHO: ret ptr @foo
+// IFUNC-MACHO: declare i32 @foo.arch_sandybridge(i32 noundef, ...)
+
 // NO-IFUNC: define dso_local i32 @foo.sse4.2(i32 noundef %i, ...)
 // NO-IFUNC: ret i32 0
 // NO-IFUNC: define dso_local i32 @foo.arch_ivybridge(i32 noundef %i, ...)

[tahonermann]

I recommend extending testing via additional RUN lines to all of the following tests:

• clang/test/CodeGen/attr-cpuspecific.c
• clang/test/CodeGen/attr-target-clones.c
• clang/test/CodeGen/attr-target-mv-func-ptrs.c
• clang/test/CodeGen/attr-target-mv-va-args.c (already done)
• clang/test/CodeGen/attr-target-mv.c
• clang/test/CodeGenCXX/attr-cpuspecific.cpp
• clang/test/CodeGenCXX/attr-target-mv-diff-ns.cpp
• clang/test/CodeGenCXX/attr-target-mv-inalloca.cpp
• clang/test/CodeGenCXX/attr-target-mv-member-funcs.cpp
• clang/test/CodeGenCXX/attr-target-mv-modules.cpp
• clang/test/CodeGenCXX/attr-target-mv-out-of-line-defs.cpp
• clang/test/CodeGenCXX/attr-target-mv-overloads.cpp
• clang/test/CodeGenCXX/attr-target-clones.cpp

Note that, for many of the above tests, changes planned in #71706 will impact the expected behavior.

[tahonermann]

Added @erichkeane as a reviewer as well.

Will this cause an ABI compatibility issue for any existing use of the function multiversioning attributes on Darwin? I think it might, but I haven't tried to confirm.

[jroelofs]

Will this cause an ABI compatibility issue for any existing use of the function multiversioning attributes on Darwin? I think it might, but I haven't tried to confirm.

I don't think there is an ABI compatibility concern here for a couple of reasons:

• The corresponding builtins are not yet implemented on Darwin until this patch series:
  • [llvm] Support IFuncs on Darwin platforms #73686
  • [clang] Function Multi Versioning supports IFunc lowerings on Darwin platforms #73688
  • [clang] Support __attribute__((ifunc(...))) on Darwin platforms #73687
  • [builtins][arm64] Build __init_cpu_features_resolver on Apple platforms #73685
• On AArch64, the IFunc lowering is required:

  llvm-project/clang/lib/CodeGen/CodeGenFunction.cpp

  Lines 2746 to 2748 in 859338a

  	bool SupportsIFunc = getContext().getTargetInfo().supportsIFunc();
  	assert(SupportsIFunc &&
  	"Multiversion resolver requires target IFUNC support");

[MaskRay]

This PR just updates supportsIFunc and extends clang tests. LGTM

[ilinpv]

LGTM!

[MaskRay]

Is "macosx" a deprecated OS name?