================
@@ -134,6 +135,106 @@ foo_priority(int x) { return x & (x - 1); }
// CHECK: ret ptr @foo_priority.default
+// Test feature-based target_clones
+int __attribute__((target_clones("altivec", "default")))
+foo_altivec(void) { return 0; }
+// CHECK: define internal {{.*}}i32 @foo_altivec.altivec()
+// CHECK: define internal {{.*}}i32 @foo_altivec.default()
+// CHECK: define internal ptr @foo_altivec.resolver()
+// if (__builtin_cpu_supports("altivec")) return &foo_altivec.altivec;
+// CHECK: %[[#ALTIVEC:]] = load i32, ptr getelementptr inbounds nuw (i8, ptr
@_system_configuration, {{i32|i64}} 204)
+// CHECK-NEXT: icmp ugt i32 %[[#ALTIVEC]], 0
+// CHECK: ret ptr @foo_altivec.altivec
+// CHECK: ret ptr @foo_altivec.default
+
+int __attribute__((target_clones("vsx", "default")))
+foo_vsx(void) { return 0; }
+// CHECK: define internal {{.*}}i32 @foo_vsx.vsx()
+// CHECK: define internal {{.*}}i32 @foo_vsx.default()
+// CHECK: define internal ptr @foo_vsx.resolver()
+// if (__builtin_cpu_supports("vsx")) return &foo_vsx.vsx;
+// CHECK: %[[#VSX:]] = load i32, ptr getelementptr inbounds nuw (i8, ptr
@_system_configuration, {{i32|i64}} 204)
+// CHECK-NEXT: icmp ugt i32 %[[#VSX]], 1
+// CHECK: ret ptr @foo_vsx.vsx
+// CHECK: ret ptr @foo_vsx.default
+
+int __attribute__((target_clones("htm", "default")))
+foo_htm(void) { return 0; }
+// CHECK: define internal {{.*}}i32 @foo_htm.htm()
+// CHECK: define internal {{.*}}i32 @foo_htm.default()
+// CHECK: define internal ptr @foo_htm.resolver()
+// if (__builtin_cpu_supports("htm")) return &foo_htm.htm;
+// CHECK: %[[#HTM:]] = call i64 @getsystemcfg(i32 59)
+// CHECK-NEXT: icmp ugt i64 %[[#HTM]], 0
+// CHECK: ret ptr @foo_htm.htm
+// CHECK: ret ptr @foo_htm.default
+
+#ifdef _ARCH_PWR10
+int __attribute__((target_clones("mma", "default")))
+foo_mma(void) { return 0; }
+#endif
+// CHECK-P10: define internal {{.*}}i32 @foo_mma.mma()
+// CHECK-P10: define internal {{.*}}i32 @foo_mma.default()
+// CHECK-P10: define internal ptr @foo_mma.resolver()
+// if (__builtin_cpu_supports("mma")) return &foo_mma.mma;
+// CHECK-P10: %[[#MMA:]] = call i64 @getsystemcfg(i32 62)
+// CHECK-P10-NEXT: icmp ugt i64 %[[#MMA]], 0
+// CHECK-P10: ret ptr @foo_mma.mma
+// CHECK-P10: ret ptr @foo_mma.default
+
+// Test multiple features with priority ordering (random source order)
+// Source order: altivec, power8-vector, cpu=pwr11, vsx, power9-vector, default
+// Resolver order by priority: cpu=pwr11 (500) > power9-vector (311) >
power8-vector (212) > vsx (111) > altivec (50) > default (0)
+int __attribute__((target_clones("altivec", "power8-vector", "cpu=pwr11",
"vsx", "power9-vector", "default")))
+foo_multi_features(void) { return 0; }
+// CHECK: define internal {{.*}}i32 @foo_multi_features.altivec()
+// CHECK: define internal {{.*}}i32 @foo_multi_features.power8_vector()
+// CHECK: define internal {{.*}}i32 @foo_multi_features.cpu_pwr11()
#[[#ATTR_P11]]
+// CHECK: define internal {{.*}}i32 @foo_multi_features.vsx()
+// CHECK: define internal {{.*}}i32 @foo_multi_features.power9_vector()
+// CHECK: define internal {{.*}}i32 @foo_multi_features.default()
+// CHECK: define internal ptr @foo_multi_features.resolver()
+// Resolver checks in priority order (highest first):
+// if (__builtin_cpu_supports("arch_3_1")) return
&foo_multi_features.cpu_pwr11;
+// CHECK: load i32, ptr getelementptr inbounds nuw (i8, ptr
@_system_configuration, {{i32|i64}} 4)
+// CHECK-NEXT: icmp uge i32 {{.*}}, 262144
+// CHECK: ret ptr @foo_multi_features.cpu_pwr11
+// if (__builtin_cpu_supports("arch_3_00")) return
&foo_multi_features.power9_vector;
+// CHECK: load i32, ptr getelementptr inbounds nuw (i8, ptr
@_system_configuration, {{i32|i64}} 4)
+// CHECK-NEXT: icmp uge i32 {{.*}}, 131072
+// CHECK: ret ptr @foo_multi_features.power9_vector
+// if (__builtin_cpu_supports("arch_2_07")) return
&foo_multi_features.power8_vector;
----------------
w2yehia wrote:
I agree we might be able to elide generating the clone in the frontend.
However, the clones have internal linkage, and at `-O1` or higher the "dead"
function will get optimized away (we first optimize away the reference to it in
the resolver function, then `GlobalOptPass` deletes it).
I considered mapping some features into CPUs, i.e treat
`target_clones("crypto", "power8-vector", "default")` as
`target_clones("cpu=power8", "default")`
GCC (on linux) doesn't do that; it encodes the feature string in the clone's
name so I decided to follow GCC's precedence there.
https://github.com/llvm/llvm-project/pull/206786
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
