[PATCH][AVX512] Handle unsigned v4i32->v4f64 convert

[Cameron McInally]

Hey guys,

Here is a pattern to convert 4 packed unsigned doubleword integers to
4 packed doubles using the AVX512 VCVTUDQ2PD instruction.

Tia,
Cameron
-------------- next part --------------
Index: test/CodeGen/X86/avx512-cvt.ll
===================================================================
--- test/CodeGen/X86/avx512-cvt.ll	(revision 211134)
+++ test/CodeGen/X86/avx512-cvt.ll	(working copy)
@@ -192,6 +192,14 @@
   ret <16 x double> %b
 }
 
+; CHECK-LABEL: uitof64_256
+; CHECK: vcvtudq2pd
+; CHECK: ret
+define <4 x double> @uitof64_256(<4 x i32> %a) nounwind {
+  %b = uitofp <4 x i32> %a to <4 x double>
+  ret <4 x double> %b
+}
+
 ; CHECK-LABEL: uitof32
 ; CHECK: vcvtudq2ps
 ; CHECK: ret
Index: lib/Target/X86/X86InstrAVX512.td
===================================================================
--- lib/Target/X86/X86InstrAVX512.td	(revision 211134)
+++ lib/Target/X86/X86InstrAVX512.td	(working copy)
@@ -3174,6 +3174,10 @@
           (EXTRACT_SUBREG (v16f32 (VCVTUDQ2PSZrr
            (v16i32 (SUBREG_TO_REG (i32 0), VR128X:$src1, sub_xmm)))), sub_xmm)>;
 
+def : Pat<(v4f64 (uint_to_fp (v4i32 VR128X:$src1))),
+          (EXTRACT_SUBREG (v8f64 (VCVTUDQ2PDZrr
+           (v8i32 (SUBREG_TO_REG (i32 0), VR128X:$src1, sub_xmm)))), sub_ymm)>;
+
 def : Pat<(v16f32 (int_x86_avx512_mask_cvtdq2ps_512 (v16i32 VR512:$src),
                    (bc_v16f32 (v16i32 immAllZerosV)), (i16 -1), imm:$rc)),
           (VCVTDQ2PSZrrb VR512:$src, imm:$rc)>;