[PATCH] D125350: [RFC][Clang] Add a check if the target supports atomicrmw instruction with specific operator and type

[Shilei Tian via Phabricator via cfe-commits]

tianshilei1992 added a comment.

As for the test, here is a simple code snippet:

  #pragma omp begin declare target device_type(nohost)
  void foo() {
    double x, y;
  #pragma omp atomic update
    x += y;
  }
  #pragma omp end declare target

`sm_35` doesn't support `FAdd` with `double` type, then it emits:

  define protected void @foo() #0 {
  entry:
    %x = alloca double, align 8
    %y = alloca double, align 8
    %atomic-temp = alloca double, align 8
    %0 = load double, ptr %y, align 8
    %atomic-load = load atomic i64, ptr %x monotonic, align 8
    br label %atomic_cont
  
  atomic_cont:                                      ; preds = %atomic_cont, %entry
    %1 = phi i64 [ %atomic-load, %entry ], [ %5, %atomic_cont ]
    %2 = bitcast i64 %1 to double
    %add = fadd double %2, %0
    store double %add, ptr %atomic-temp, align 8
    %3 = load i64, ptr %atomic-temp, align 8
    %4 = cmpxchg ptr %x, i64 %1, i64 %3 monotonic monotonic, align 8
    %5 = extractvalue { i64, i1 } %4, 0
    %6 = extractvalue { i64, i1 } %4, 1
    br i1 %6, label %atomic_exit, label %atomic_cont
  
  atomic_exit:                                      ; preds = %atomic_cont
    ret void
  }

Starting from `sm_60`, it is supported, now we emits:

  define protected void @foo() #0 {
  entry:
    %x = alloca double, align 8
    %y = alloca double, align 8
    %0 = load double, ptr %y, align 8
    %1 = atomicrmw fadd ptr %x, double %0 monotonic, align 8
    ret void
  }


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D125350/new/

https://reviews.llvm.org/D125350