compiler-rt: fix compiler rt with CROSS_TOOLCHAIN

[brooks]

Summary

Always fall back to the locked version of atomic_fetch_nand when the
__c11_atomic_fetch_nand builtin isn't available (LLVM prior to 14).

[brooks]

This seems a bit problematic in that the set of available symbols depends on the compiler building the tree. It may be that we want to generate them and emit the function without the if (lockfree(ptr) path instead.

[arichardson]

Adding downstream patches is not great but I think in this case it's reasonable. Although maybe upstream would be happy with this too, I'm not sure what the supported compilers for compiler-rt are.

[brooks]

In D35311#800371, @arichardson wrote:

Adding downstream patches is not great but I think in this case it's reasonable. Although maybe upstream would be happy with this too, I'm not sure what the supported compilers for compiler-rt are.

I can submit what ever we decide is the right approach upstream.

[brooks]

In D35311#800370, @brooks wrote:

This seems a bit problematic in that the set of available symbols depends on the compiler building the tree. It may be that we want to generate them and emit the function without the if (lockfree(ptr) path instead.

An alternative version also compiles:

diff --git a/contrib/llvm-project/compiler-rt/lib/builtins/atomic.c b/contrib/llvm-project/compiler-rt/lib/builtins/atomic.c
index b8c7f4b0511c..333429751f24 100644
--- a/contrib/llvm-project/compiler-rt/lib/builtins/atomic.c
+++ b/contrib/llvm-project/compiler-rt/lib/builtins/atomic.c
@@ -336,6 +336,7 @@ OPTIMISED_CASES
     return tmp;                                                                \
   }
 
+#if __has_builtin(__c11_atomic_fetch_nand)
 #define ATOMIC_RMW_NAND(n, lockfree, type)                                     \
   type __atomic_fetch_nand_##n(type *ptr, type val, int model) {               \
     if (lockfree(ptr))                                                         \
@@ -347,6 +348,17 @@ OPTIMISED_CASES
     unlock(l);                                                                 \
     return tmp;                                                                \
   }
+#else
+#define ATOMIC_RMW_NAND(n, lockfree, type)                                     \
+  type __atomic_fetch_nand_##n(type *ptr, type val, int model) {               \
+    Lock *l = lock_for_pointer(ptr);                                           \
+    lock(l);                                                                   \
+    type tmp = *ptr;                                                           \
+    *ptr = ~(tmp & val);                                                       \
+    unlock(l);                                                                 \
+    return tmp;                                                                \
+  }
+#endif
 
 #define OPTIMISED_CASE(n, lockfree, type) ATOMIC_RMW(n, lockfree, type, add, +)
 OPTIMISED_CASES

[brooks]

• Switch to building the locked version

[brooks]

https://reviews.llvm.org/D126395

[emaste]

Seems OK to me. I was wondering about trying to reuse the lock version body in the other but I guess the duplication is actually clearer/more maintainable.

[dim]

This LGTM, provided upstream accepts the same patch. :)

[brooks]

Upstream accepted the first version of this patch as https://reviews.llvm.org/D126710