Internal change

tcmalloc/allocation_sampling.h

@@ -50,11 +50,9 @@ std::unique_ptr<const ProfileBase> DumpHeapProfile(Static& state);
 // avoid relying on the percpu code.
 extern "C" ABSL_CONST_INIT thread_local Sampler tcmalloc_sampler
     ABSL_ATTRIBUTE_INITIAL_EXEC;
-
-#endif
-
 ABSL_CONST_INIT ABSL_ATTRIBUTE_WEAK thread_local Sampler tcmalloc_sampler
     ABSL_ATTRIBUTE_INITIAL_EXEC;
+#endif
 
 inline Sampler& GetThreadSampler() {
   static_assert(sizeof(Sampler) == TCMALLOC_SAMPLER_SIZE,

tcmalloc/internal/percpu.cc

@@ -144,6 +144,9 @@ static void InitPerCpu() {
     // objects can overlap.
     volatile auto slabs_addr = reinterpret_cast<uintptr_t>(&tcmalloc_slabs);
     auto rseq_abi_addr = reinterpret_cast<uintptr_t>(&__rseq_abi);
+    volatile auto tcmalloc_rseq_layout_addr =
+        reinterpret_cast<uintptr_t>(&tcmalloc_rseq_layout);
+    TC_CHECK_EQ(rseq_abi_addr, tcmalloc_rseq_layout_addr);
     //  Ensure __rseq_abi alignment required by ABI.
     TC_CHECK_EQ(rseq_abi_addr % 32, 0);
     // Ensure that all our TLS data is in a single cache line.

tcmalloc/internal/percpu.h

@@ -189,6 +189,15 @@ ABSL_CONST_INIT thread_local volatile kernel_rseq __rseq_abi
 };
 ABSL_CONST_INIT thread_local volatile int tcmalloc_cached_vcpu
     ABSL_ATTRIBUTE_WEAK = kCpuIdUninitialized;
+ABSL_CONST_INIT thread_local char tcmalloc_sampler ABSL_ATTRIBUTE_WEAK = 0;
+
+// Because of the weak definitions we also add this non-weak symbol.
+// It's not used other than ensuring that our asm file with the real TLS layout
+// is actually linked. Without it linker manages to not link the asm file
+// in some configurations since all symbols it provides are already defined
+// (as weak), even though the asm file is in the same static library as other
+// linked in files.
+extern "C" ABSL_CONST_INIT thread_local char tcmalloc_rseq_layout;
 
 inline int GetRealCpuUnsafe() { return __rseq_abi.cpu_id; }
 #else  // !TCMALLOC_INTERNAL_PERCPU_USE_RSEQ

tcmalloc/internal/percpu_rseq_asm.S

@@ -36,11 +36,13 @@
 .type tcmalloc_cached_vcpu, @object
 .type tcmalloc_slabs, @object
 .type __rseq_abi, @object
+.type tcmalloc_rseq_layout, @object
 .section .tdata, "awT", @progbits
 .globl tcmalloc_sampler
 .globl tcmalloc_cached_vcpu
 .globl tcmalloc_slabs
 .globl __rseq_abi
+.globl tcmalloc_rseq_layout
 .p2align 6
 .zero 64 + 32 - TCMALLOC_SAMPLER_SIZE - 8
 tcmalloc_sampler:
@@ -49,6 +51,7 @@ tcmalloc_cached_vcpu:
 .long 0xffffffff  // cpu_id (kCpuIdUninitialized)
 tcmalloc_slabs:
 .long 0
+tcmalloc_rseq_layout:
 __rseq_abi:
 .long 0           // cpu_id_start
 .long 0xffffffff  // cpu_id (kCpuIdUninitialized)