AMDGPU: Correctly expand f64 sqrt intrinsic

rocm-device-libs and llpc were avoiding using f64 sqrt
intrinsics in favor of their own expansions. Port the
expansion into the backend. Both of these users should be
updated to call the intrinsic instead.

The library and llpc expansions are slightly different.
llpc uses an ldexp to do the scale; the library uses a multiply.

Use ldexp to do the scale instead of the multiply.
I believe v_ldexp_f64 and v_mul_f64 are always the same number of
cycles, but it's cheaper to materialize the 32-bit integer constant
than the 64-bit double constant.

The libraries have another fast version of sqrt which will
be handled separately.

I am tempted to do this in an IR expansion instead. In the IR
we could take advantage of computeKnownFPClass to avoid
the 0-or-inf argument check.

1 files changed, 7 insertions, 0 deletions

diff --git a/llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h b/llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h
index a1ff764..5341b57 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/MachineIRBuilder.h
@@ -1181,6 +1181,13 @@ public:
                                 const SrcOp &Op0, const SrcOp &Op1,
                                 std::optional<unsigned> Flags = std::nullopt);
 
+  /// Build and insert a \p Res = G_IS_FPCLASS \p Pred\p Src, \p Mask
+  MachineInstrBuilder buildIsFPClass(const DstOp &Res, const SrcOp &Src,
+                                     unsigned Mask) {
+    return buildInstr(TargetOpcode::G_IS_FPCLASS, {Res},
+                      {Src, SrcOp(static_cast<int64_t>(Mask))});
+  }
+
   /// Build and insert a \p Res = G_SELECT \p Tst, \p Op0, \p Op1
   ///
   /// \pre setBasicBlock or setMI must have been called.