maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/clang/test/CodeGen/attr-target-mv.c

185 lines
6.9 KiB
C
Raw Normal View History

Implement Function Multiversioning for Non-ELF Systems. Similar to how ICC handles CPU-Dispatch on Windows, this patch uses the resolver function directly to forward the call to the proper function. This is not nearly as efficient as IFuncs of course, but is still quite useful for large functions specifically developed for certain processors. This is unfortunately still limited to x86, since it depends on __builtin_cpu_supports and __builtin_cpu_is, which are x86 builtins. The naming for the resolver/forwarding function for cpu-dispatch was taken from ICC's implementation, which uses the unmodified name for this (no mangling additions). This is possible, since cpu-dispatch uses '.A' for the 'default' version. In 'target' multiversioning, this function keeps the '.resolver' extension in order to keep the default function keeping the default mangling. Change-Id: I4731555a39be26c7ad59a2d8fda6fa1a50f73284 Differential Revision: https://reviews.llvm.org/D53586 llvm-svn: 345298
2018-10-26 02:57:19 +08:00
// RUN: %clang_cc1 -triple x86_64-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix=LINUX
// RUN: %clang_cc1 -triple x86_64-windows-pc -emit-llvm %s -o - | FileCheck %s --check-prefix=WINDOWS
Implement Attribute Target MultiVersioning GCC's attribute 'target', in addition to being an optimization hint, also allows function multiversioning. We currently have the former implemented, this is the latter's implementation. This works by enabling functions with the same name/signature to coexist, so that they can all be emitted. Multiversion state is stored in the FunctionDecl itself, and SemaDecl manages the definitions. Note that it ends up having to permit redefinition of functions so that they can all be emitted. Additionally, all versions of the function must be emitted, so this also manages that. Note that this includes some additional rules that GCC does not, since defining something as a MultiVersion function after a usage has been made illegal. The only 'history rewriting' that happens is if a function is emitted before it has been converted to a multiversion'ed function, at which point its name needs to be changed. Function templates and virtual functions are NOT yet supported (not supported in GCC either). Additionally, constructors/destructors are disallowed, but the former is planned. llvm-svn: 322028
2018-01-09 05:34:17 +08:00
int __attribute__((target("sse4.2"))) foo(void) { return 0; }
int __attribute__((target("arch=sandybridge"))) foo(void);
int __attribute__((target("arch=ivybridge"))) foo(void) {return 1;}
int __attribute__((target("default"))) foo(void) { return 2; }
int bar() {
return foo();
}
inline int __attribute__((target("sse4.2"))) foo_inline(void) { return 0; }
inline int __attribute__((target("arch=sandybridge"))) foo_inline(void);
inline int __attribute__((target("arch=ivybridge"))) foo_inline(void) {return 1;}
inline int __attribute__((target("default"))) foo_inline(void) { return 2; }
int bar2() {
return foo_inline();
}
inline __attribute__((target("default"))) void foo_decls(void);
inline __attribute__((target("sse4.2"))) void foo_decls(void);
void bar3() {
foo_decls();
}
inline __attribute__((target("default"))) void foo_decls(void) {}
inline __attribute__((target("sse4.2"))) void foo_decls(void) {}
Implement Function Multiversioning for Non-ELF Systems. Similar to how ICC handles CPU-Dispatch on Windows, this patch uses the resolver function directly to forward the call to the proper function. This is not nearly as efficient as IFuncs of course, but is still quite useful for large functions specifically developed for certain processors. This is unfortunately still limited to x86, since it depends on __builtin_cpu_supports and __builtin_cpu_is, which are x86 builtins. The naming for the resolver/forwarding function for cpu-dispatch was taken from ICC's implementation, which uses the unmodified name for this (no mangling additions). This is possible, since cpu-dispatch uses '.A' for the 'default' version. In 'target' multiversioning, this function keeps the '.resolver' extension in order to keep the default function keeping the default mangling. Change-Id: I4731555a39be26c7ad59a2d8fda6fa1a50f73284 Differential Revision: https://reviews.llvm.org/D53586 llvm-svn: 345298
2018-10-26 02:57:19 +08:00
inline __attribute__((target("default"))) void foo_multi(int i, double d) {}
inline __attribute__((target("avx,sse4.2"))) void foo_multi(int i, double d) {}
inline __attribute__((target("sse4.2,fma4"))) void foo_multi(int i, double d) {}
inline __attribute__((target("arch=ivybridge,fma4,sse4.2"))) void foo_multi(int i, double d) {}
[CodeGen] Fix a crash on mangling multiversioned functions `multiVersionSortPriority` expects features to have no prefix. We currently carry them around in the format "+${feature}". llvm-svn: 322618
2018-01-17 12:46:04 +08:00
void bar4() {
Implement Function Multiversioning for Non-ELF Systems. Similar to how ICC handles CPU-Dispatch on Windows, this patch uses the resolver function directly to forward the call to the proper function. This is not nearly as efficient as IFuncs of course, but is still quite useful for large functions specifically developed for certain processors. This is unfortunately still limited to x86, since it depends on __builtin_cpu_supports and __builtin_cpu_is, which are x86 builtins. The naming for the resolver/forwarding function for cpu-dispatch was taken from ICC's implementation, which uses the unmodified name for this (no mangling additions). This is possible, since cpu-dispatch uses '.A' for the 'default' version. In 'target' multiversioning, this function keeps the '.resolver' extension in order to keep the default function keeping the default mangling. Change-Id: I4731555a39be26c7ad59a2d8fda6fa1a50f73284 Differential Revision: https://reviews.llvm.org/D53586 llvm-svn: 345298
2018-10-26 02:57:19 +08:00
foo_multi(1, 5.0);
[CodeGen] Fix a crash on mangling multiversioned functions `multiVersionSortPriority` expects features to have no prefix. We currently carry them around in the format "+${feature}". llvm-svn: 322618
2018-01-17 12:46:04 +08:00
}
Implement Function Multiversioning for Non-ELF Systems. Similar to how ICC handles CPU-Dispatch on Windows, this patch uses the resolver function directly to forward the call to the proper function. This is not nearly as efficient as IFuncs of course, but is still quite useful for large functions specifically developed for certain processors. This is unfortunately still limited to x86, since it depends on __builtin_cpu_supports and __builtin_cpu_is, which are x86 builtins. The naming for the resolver/forwarding function for cpu-dispatch was taken from ICC's implementation, which uses the unmodified name for this (no mangling additions). This is possible, since cpu-dispatch uses '.A' for the 'default' version. In 'target' multiversioning, this function keeps the '.resolver' extension in order to keep the default function keeping the default mangling. Change-Id: I4731555a39be26c7ad59a2d8fda6fa1a50f73284 Differential Revision: https://reviews.llvm.org/D53586 llvm-svn: 345298
2018-10-26 02:57:19 +08:00
// LINUX: @foo.ifunc = ifunc i32 (), i32 ()* ()* @foo.resolver
// LINUX: @foo_inline.ifunc = ifunc i32 (), i32 ()* ()* @foo_inline.resolver
// LINUX: @foo_decls.ifunc = ifunc void (), void ()* ()* @foo_decls.resolver
// LINUX: @foo_multi.ifunc = ifunc void (i32, double), void (i32, double)* ()* @foo_multi.resolver
// LINUX: define i32 @foo.sse4.2()
// LINUX: ret i32 0
// LINUX: define i32 @foo.arch_ivybridge()
// LINUX: ret i32 1
// LINUX: define i32 @foo()
// LINUX: ret i32 2
// LINUX: define i32 @bar()
// LINUX: call i32 @foo.ifunc()
// WINDOWS: define dso_local i32 @foo.sse4.2()
// WINDOWS: ret i32 0
// WINDOWS: define dso_local i32 @foo.arch_ivybridge()
// WINDOWS: ret i32 1
// WINDOWS: define dso_local i32 @foo()
// WINDOWS: ret i32 2
// WINDOWS: define dso_local i32 @bar()
// WINDOWS: call i32 @foo.resolver()
// LINUX: define i32 ()* @foo.resolver() comdat
// LINUX: call void @__cpu_indicator_init()
// LINUX: ret i32 ()* @foo.arch_sandybridge
// LINUX: ret i32 ()* @foo.arch_ivybridge
// LINUX: ret i32 ()* @foo.sse4.2
// LINUX: ret i32 ()* @foo
// WINDOWS: define dso_local i32 @foo.resolver() comdat
// WINDOWS: call void @__cpu_indicator_init()
// WINDOWS: call i32 @foo.arch_sandybridge
// WINDOWS: call i32 @foo.arch_ivybridge
// WINDOWS: call i32 @foo.sse4.2
// WINDOWS: call i32 @foo
Multiversioning- Ensure all MV functions are emitted. Multiverson function versions are always used (by the resolver), so ensure that they are always emitted. Change-Id: I5d2e0841fddf0d18918b3fb92ae76814add7ee96 llvm-svn: 345839
2018-11-01 23:11:43 +08:00
// LINUX: define linkonce i32 @foo_inline.arch_ivybridge()
// LINUX: ret i32 1
// LINUX: define linkonce i32 @foo_inline()
// LINUX: ret i32 2
// WINDOWS: define linkonce_odr dso_local i32 @foo_inline.arch_ivybridge()
// WINDOWS: ret i32 1
// WINDOWS: define linkonce_odr dso_local i32 @foo_inline()
// WINDOWS: ret i32 2
Implement Function Multiversioning for Non-ELF Systems. Similar to how ICC handles CPU-Dispatch on Windows, this patch uses the resolver function directly to forward the call to the proper function. This is not nearly as efficient as IFuncs of course, but is still quite useful for large functions specifically developed for certain processors. This is unfortunately still limited to x86, since it depends on __builtin_cpu_supports and __builtin_cpu_is, which are x86 builtins. The naming for the resolver/forwarding function for cpu-dispatch was taken from ICC's implementation, which uses the unmodified name for this (no mangling additions). This is possible, since cpu-dispatch uses '.A' for the 'default' version. In 'target' multiversioning, this function keeps the '.resolver' extension in order to keep the default function keeping the default mangling. Change-Id: I4731555a39be26c7ad59a2d8fda6fa1a50f73284 Differential Revision: https://reviews.llvm.org/D53586 llvm-svn: 345298
2018-10-26 02:57:19 +08:00
// LINUX: define i32 @bar2()
// LINUX: call i32 @foo_inline.ifunc()
// WINDOWS: define dso_local i32 @bar2()
// WINDOWS: call i32 @foo_inline.resolver()
// LINUX: define i32 ()* @foo_inline.resolver() comdat
// LINUX: call void @__cpu_indicator_init()
// LINUX: ret i32 ()* @foo_inline.arch_sandybridge
// LINUX: ret i32 ()* @foo_inline.arch_ivybridge
// LINUX: ret i32 ()* @foo_inline.sse4.2
// LINUX: ret i32 ()* @foo_inline
// WINDOWS: define dso_local i32 @foo_inline.resolver() comdat
// WINDOWS: call void @__cpu_indicator_init()
// WINDOWS: call i32 @foo_inline.arch_sandybridge
// WINDOWS: call i32 @foo_inline.arch_ivybridge
// WINDOWS: call i32 @foo_inline.sse4.2
// WINDOWS: call i32 @foo_inline
// LINUX: define void @bar3()
// LINUX: call void @foo_decls.ifunc()
// WINDOWS: define dso_local void @bar3()
// WINDOWS: call void @foo_decls.resolver()
// LINUX: define void ()* @foo_decls.resolver() comdat
// LINUX: ret void ()* @foo_decls.sse4.2
// LINUX: ret void ()* @foo_decls
// WINDOWS: define dso_local void @foo_decls.resolver() comdat
// WINDOWS: call void @foo_decls.sse4.2
// Windows: call void @foo_decls
Multiversioning- Ensure all MV functions are emitted. Multiverson function versions are always used (by the resolver), so ensure that they are always emitted. Change-Id: I5d2e0841fddf0d18918b3fb92ae76814add7ee96 llvm-svn: 345839
2018-11-01 23:11:43 +08:00
// LINUX: define linkonce void @foo_decls()
// LINUX: define linkonce void @foo_decls.sse4.2()
// WINDOWS: define linkonce_odr dso_local void @foo_decls()
// WINDOWS: define linkonce_odr dso_local void @foo_decls.sse4.2()
// LINUX: define linkonce void @foo_multi(i32 %{{[^,]+}}, double %{{[^\)]+}})
// LINUX: define linkonce void @foo_multi.avx_sse4.2(i32 %{{[^,]+}}, double %{{[^\)]+}})
// LINUX: define linkonce void @foo_multi.fma4_sse4.2(i32 %{{[^,]+}}, double %{{[^\)]+}})
// LINUX: define linkonce void @foo_multi.arch_ivybridge_fma4_sse4.2(i32 %{{[^,]+}}, double %{{[^\)]+}})
// WINDOWS: define linkonce_odr dso_local void @foo_multi(i32 %{{[^,]+}}, double %{{[^\)]+}})
// WINDOWS: define linkonce_odr dso_local void @foo_multi.avx_sse4.2(i32 %{{[^,]+}}, double %{{[^\)]+}})
// WINDOWS: define linkonce_odr dso_local void @foo_multi.fma4_sse4.2(i32 %{{[^,]+}}, double %{{[^\)]+}})
// WINDOWS: define linkonce_odr dso_local void @foo_multi.arch_ivybridge_fma4_sse4.2(i32 %{{[^,]+}}, double %{{[^\)]+}})
Implement Function Multiversioning for Non-ELF Systems. Similar to how ICC handles CPU-Dispatch on Windows, this patch uses the resolver function directly to forward the call to the proper function. This is not nearly as efficient as IFuncs of course, but is still quite useful for large functions specifically developed for certain processors. This is unfortunately still limited to x86, since it depends on __builtin_cpu_supports and __builtin_cpu_is, which are x86 builtins. The naming for the resolver/forwarding function for cpu-dispatch was taken from ICC's implementation, which uses the unmodified name for this (no mangling additions). This is possible, since cpu-dispatch uses '.A' for the 'default' version. In 'target' multiversioning, this function keeps the '.resolver' extension in order to keep the default function keeping the default mangling. Change-Id: I4731555a39be26c7ad59a2d8fda6fa1a50f73284 Differential Revision: https://reviews.llvm.org/D53586 llvm-svn: 345298
2018-10-26 02:57:19 +08:00
// LINUX: define void @bar4()
// LINUX: call void @foo_multi.ifunc(i32 1, double 5.{{[0+e]*}})
// WINDOWS: define dso_local void @bar4()
// WINDOWS: call void @foo_multi.resolver(i32 1, double 5.{{[0+e]*}})
// LINUX: define void (i32, double)* @foo_multi.resolver() comdat
// LINUX: and i32 %{{.*}}, 4352
// LINUX: icmp eq i32 %{{.*}}, 4352
// LINUX: ret void (i32, double)* @foo_multi.fma4_sse4.2
// LINUX: icmp eq i32 %{{.*}}, 12
// LINUX: and i32 %{{.*}}, 4352
// LINUX: icmp eq i32 %{{.*}}, 4352
// LINUX: ret void (i32, double)* @foo_multi.arch_ivybridge_fma4_sse4.2
// LINUX: and i32 %{{.*}}, 768
// LINUX: icmp eq i32 %{{.*}}, 768
// LINUX: ret void (i32, double)* @foo_multi.avx_sse4.2
// LINUX: ret void (i32, double)* @foo_multi
// WINDOWS: define dso_local void @foo_multi.resolver(i32, double) comdat
// WINDOWS: and i32 %{{.*}}, 4352
// WINDOWS: icmp eq i32 %{{.*}}, 4352
// WINDOWS: call void @foo_multi.fma4_sse4.2(i32 %0, double %1)
// WINDOWS-NEXT: ret void
// WINDOWS: icmp eq i32 %{{.*}}, 12
// WINDOWS: and i32 %{{.*}}, 4352
// WINDOWS: icmp eq i32 %{{.*}}, 4352
// WINDOWS: call void @foo_multi.arch_ivybridge_fma4_sse4.2(i32 %0, double %1)
// WINDOWS-NEXT: ret void
// WINDOWS: and i32 %{{.*}}, 768
// WINDOWS: icmp eq i32 %{{.*}}, 768
// WINDOWS: call void @foo_multi.avx_sse4.2(i32 %0, double %1)
// WINDOWS-NEXT: ret void
// WINDOWS: call void @foo_multi(i32 %0, double %1)
// WINDOWS-NEXT: ret void
// LINUX: declare i32 @foo.arch_sandybridge()
// WINDOWS: declare dso_local i32 @foo.arch_sandybridge()
// LINUX: define linkonce i32 @foo_inline.sse4.2()
// LINUX: ret i32 0
// WINDOWS: define linkonce_odr dso_local i32 @foo_inline.sse4.2()
// WINDOWS: ret i32 0
// LINUX: declare i32 @foo_inline.arch_sandybridge()
// WINDOWS: declare dso_local i32 @foo_inline.arch_sandybridge()