[SelectionDAG] Fixed f16-from-vector promotion problem

Summary:
In the case of an fp_extend of v1f16 to v1f32 where the v1f16 is the
result of a bitcast from i16, avoid creating an illegal fp16_to_fp where
the input is not a vector and the result is a v1f32.

V2: The fix is now to avoid vector scalarization creating a v1->scalar
bitcast.

Reviewers: srhines, t.p.northover

Subscribers: nhaehnle, llvm-commits, dstuttard, t-tye, yaxunl, wdng, kzhuravl, arsenm

Differential Revision: https://reviews.llvm.org/D41126

llvm-svn: 322120

1 files changed, 26 insertions, 0 deletions

diff --git a/llvm/test/CodeGen/AMDGPU/unpack-half.ll b/llvm/test/CodeGen/AMDGPU/unpack-half.ll
new file mode 100644
index 00000000000..b2133986ba5
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/unpack-half.ll
@@ -0,0 +1,26 @@
+; RUN: llc -march=amdgcn -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -march=amdgcn -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck %s
+
+; On gfx6 and gfx7, this test shows a bug in SelectionDAG where scalarizing the
+; extension of a vector of f16 generates an illegal node that errors later.
+
+; CHECK-LABEL: {{^}}main:
+; CHECK: v_cvt_f32_f16
+
+define amdgpu_gs void @main(i32 inreg %arg) local_unnamed_addr #0 {
+.entry:
+  %tmp = load volatile float, float addrspace(1)* undef
+  %tmp1 = bitcast float %tmp to i32
+  %im0.i = lshr i32 %tmp1, 16
+  %tmp2 = insertelement <2 x i32> undef, i32 %im0.i, i32 1
+  %tmp3 = trunc <2 x i32> %tmp2 to <2 x i16>
+  %tmp4 = bitcast <2 x i16> %tmp3 to <2 x half>
+  %tmp5 = fpext <2 x half> %tmp4 to <2 x float>
+  %bc = bitcast <2 x float> %tmp5 to <2 x i32>
+  %tmp6 = extractelement <2 x i32> %bc, i32 1
+  store volatile i32 %tmp6, i32 addrspace(1)* undef
+  ret void
+}
+
+attributes #0 = { nounwind }
+