From: Adam Nemet <anemet@apple.com>
Date: Wed, 2 Jul 2014 21:26:01 +0000 (+0000)
Subject: [X86] AVX512: Allow writemask argument in vpermt* intrinsics
X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=commitdiff_plain;h=998743e1855324c274561abbed15e60373025313

[X86] AVX512: Allow writemask argument in vpermt* intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@212223 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index dd1a52facf0..c8b0ecdbe49 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -693,22 +693,32 @@ defm VPERMI2PD : avx512_perm_3src<0x77, "vpermi2pd",  VR512, memopv8f64,
 
 multiclass avx512_perm_table_3src<bits<8> opc, string Suffix, RegisterClass RC,
                           PatFrag mem_frag, X86MemOperand x86memop,
-                          SDNode OpNode, ValueType OpVT, RegisterClass KRC> :
+                          SDNode OpNode, ValueType OpVT, RegisterClass KRC,
+                          ValueType MaskVT, RegisterClass MRC> :
         avx512_perm_3src<opc, "vpermt2"##Suffix, RC, mem_frag, x86memop, OpNode,
                          OpVT, KRC> {
   def : Pat<(OpVT (!cast<Intrinsic>("int_x86_avx512_mask_vpermt_"##Suffix##"_512")
                      VR512:$idx, VR512:$src1, VR512:$src2, -1)),
             (!cast<Instruction>(NAME#rr) VR512:$src1, VR512:$idx, VR512:$src2)>;
+
+  def : Pat<(OpVT (!cast<Intrinsic>("int_x86_avx512_mask_vpermt_"##Suffix##"_512")
+                     VR512:$idx, VR512:$src1, VR512:$src2, MRC:$mask)),
+            (!cast<Instruction>(NAME#rrk) VR512:$src1,
+              (MaskVT (COPY_TO_REGCLASS MRC:$mask, KRC)), VR512:$idx, VR512:$src2)>;
 }
 
 defm VPERMT2D  : avx512_perm_table_3src<0x7E, "d",  VR512, memopv16i32, i512mem,
-                               X86VPermv3, v16i32, VK16WM>, EVEX_V512, EVEX_CD8<32, CD8VF>;
+                               X86VPermv3, v16i32, VK16WM, v16i1, GR16>,
+                 EVEX_V512, EVEX_CD8<32, CD8VF>;
 defm VPERMT2Q  : avx512_perm_table_3src<0x7E, "q",  VR512, memopv8i64, i512mem,
-                               X86VPermv3, v8i64, VK8WM>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+                               X86VPermv3, v8i64, VK8WM, v8i1, GR8>,
+                 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
 defm VPERMT2PS : avx512_perm_table_3src<0x7F, "ps",  VR512, memopv16f32, i512mem,
-                               X86VPermv3, v16f32, VK16WM>, EVEX_V512, EVEX_CD8<32, CD8VF>;
+                               X86VPermv3, v16f32, VK16WM, v16i1, GR16>,
+                 EVEX_V512, EVEX_CD8<32, CD8VF>;
 defm VPERMT2PD : avx512_perm_table_3src<0x7F, "pd",  VR512, memopv8f64, i512mem,
-                               X86VPermv3, v8f64, VK8WM>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+                               X86VPermv3, v8f64, VK8WM, v8i1, GR8>,
+                 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
 
 //===----------------------------------------------------------------------===//
 // AVX-512 - BLEND using mask
diff --git a/test/CodeGen/X86/avx512-intrinsics.ll b/test/CodeGen/X86/avx512-intrinsics.ll
index 51175390a25..18cfcfe78b0 100644
--- a/test/CodeGen/X86/avx512-intrinsics.ll
+++ b/test/CodeGen/X86/avx512-intrinsics.ll
@@ -594,6 +594,13 @@ define <16 x float> @test_vpermt2ps(<16 x float>%x, <16 x float>%y, <16 x i32>%p
   ret <16 x float> %res
 }
 
+define <16 x float> @test_vpermt2ps_mask(<16 x float>%x, <16 x float>%y, <16 x i32>%perm, i16 %mask) {
+; CHECK-LABEL: test_vpermt2ps_mask:
+; CHECK: vpermt2ps %zmm1, %zmm2, %zmm0 {%k1} ## encoding: [0x62,0xf2,0x6d,0x49,0x7f,0xc1]
+  %res = call <16 x float> @llvm.x86.avx512.mask.vpermt.ps.512(<16 x i32>%perm, <16 x float>%x, <16 x float>%y, i16 %mask)
+  ret <16 x float> %res
+}
+
 declare <16 x float> @llvm.x86.avx512.mask.vpermt.ps.512(<16 x i32>, <16 x float>, <16 x float>, i16)
 
 define <8 x i64> @test_vmovntdqa(i8 *%x) {