================
@@ -126,3 +138,23 @@ define void @f9(ptr %cptr, ptr %aptr, ptr %bptr) {
store fp128 %c, ptr %cptr
ret void
}
+
+; Test f16 copies in which the sign comes from an f16.
+define half @f10(half %a, half %b) {
+; CHECK-LABEL: f10:
+; CHECK: brasl %r14, copysignh@PLT
----------------
JonPsson1 wrote:
I have come to realize that these tests actually shouldn't involve copysignh at
all, as here *libcalls* are tested.
On the other hand, if I compile a C program with copysign() using _Float16,
Clang emits calls to the *intrinsic*:
```
%0 = load half, ptr %v0, align 2, !tbaa !4
%conv = fpext half %0 to double
%1 = load half, ptr %v1, align 2, !tbaa !4
%conv2 = fpext half %1 to double
%2 = call double @llvm.copysign.f64(double %conv, double %conv2)
```
However the DAGCombiner will remove FPEXTs of the operands, so the backend
still has to handle FCOPYSIGNs with one or two half operands (the
ISD::FCOPYSIGN node can have the second argument of a different type).
This is now implemented in lowerFCOPYSIGN() and tested in fp-copysign-03.ll.
https://github.com/llvm/llvm-project/pull/109164
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
