From 906f3971374af132ff052af3a3d6cc56b949738f Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@gmail.com>
Date: Sun, 8 May 2016 23:08:40 +0000
Subject: [PATCH] [AVX512] Change predicates on some vXi16/vXi8 AVX store
 patterns so they stay enabled unless VLX and BWI instructions are supported."

Without this we could fail instruction selection if VLX was enabled, but BWI wasn't.

llvm-svn: 268885
---
 llvm/lib/Target/X86/X86InstrSSE.td | 38 +++++++++++++++++-------------
 1 file changed, 22 insertions(+), 16 deletions(-)

diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index 814247899cee..52b288926ef0 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -965,18 +965,10 @@ let Predicates = [HasAVX, NoVLX] in {
             (VMOVAPSmr addr:$dst, VR128:$src)>;
   def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
             (VMOVAPSmr addr:$dst, VR128:$src)>;
-  def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
-            (VMOVAPSmr addr:$dst, VR128:$src)>;
-  def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
-            (VMOVAPSmr addr:$dst, VR128:$src)>;
   def : Pat<(store (v2i64 VR128:$src), addr:$dst),
             (VMOVUPSmr addr:$dst, VR128:$src)>;
   def : Pat<(store (v4i32 VR128:$src), addr:$dst),
             (VMOVUPSmr addr:$dst, VR128:$src)>;
-  def : Pat<(store (v8i16 VR128:$src), addr:$dst),
-            (VMOVUPSmr addr:$dst, VR128:$src)>;
-  def : Pat<(store (v16i8 VR128:$src), addr:$dst),
-            (VMOVUPSmr addr:$dst, VR128:$src)>;
 
   // 256-bit load/store
   def : Pat<(alignedloadv4i64 addr:$src),
@@ -987,18 +979,10 @@ let Predicates = [HasAVX, NoVLX] in {
             (VMOVAPSYmr addr:$dst, VR256:$src)>;
   def : Pat<(alignedstore256 (v8i32 VR256:$src), addr:$dst),
             (VMOVAPSYmr addr:$dst, VR256:$src)>;
-  def : Pat<(alignedstore256 (v16i16 VR256:$src), addr:$dst),
-            (VMOVAPSYmr addr:$dst, VR256:$src)>;
-  def : Pat<(alignedstore256 (v32i8 VR256:$src), addr:$dst),
-            (VMOVAPSYmr addr:$dst, VR256:$src)>;
   def : Pat<(store (v4i64 VR256:$src), addr:$dst),
             (VMOVUPSYmr addr:$dst, VR256:$src)>;
   def : Pat<(store (v8i32 VR256:$src), addr:$dst),
             (VMOVUPSYmr addr:$dst, VR256:$src)>;
-  def : Pat<(store (v16i16 VR256:$src), addr:$dst),
-            (VMOVUPSYmr addr:$dst, VR256:$src)>;
-  def : Pat<(store (v32i8 VR256:$src), addr:$dst),
-            (VMOVUPSYmr addr:$dst, VR256:$src)>;
 
   // Special patterns for storing subvector extracts of lower 128-bits
   // Its cheaper to just use VMOVAPS/VMOVUPS instead of VEXTRACTF128mr
@@ -1041,6 +1025,28 @@ let Predicates = [HasAVX, NoVLX] in {
             (VMOVUPSmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>;
 }
 
+let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
+  // 128-bit load/store
+  def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
+            (VMOVAPSmr addr:$dst, VR128:$src)>;
+  def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
+            (VMOVAPSmr addr:$dst, VR128:$src)>;
+  def : Pat<(store (v8i16 VR128:$src), addr:$dst),
+            (VMOVUPSmr addr:$dst, VR128:$src)>;
+  def : Pat<(store (v16i8 VR128:$src), addr:$dst),
+            (VMOVUPSmr addr:$dst, VR128:$src)>;
+
+  // 256-bit load/store
+  def : Pat<(alignedstore256 (v16i16 VR256:$src), addr:$dst),
+            (VMOVAPSYmr addr:$dst, VR256:$src)>;
+  def : Pat<(alignedstore256 (v32i8 VR256:$src), addr:$dst),
+            (VMOVAPSYmr addr:$dst, VR256:$src)>;
+  def : Pat<(store (v16i16 VR256:$src), addr:$dst),
+            (VMOVUPSYmr addr:$dst, VR256:$src)>;
+  def : Pat<(store (v32i8 VR256:$src), addr:$dst),
+            (VMOVUPSYmr addr:$dst, VR256:$src)>;
+}
+
 // Use movaps / movups for SSE integer load / store (one byte shorter).
 // The instructions selected below are then converted to MOVDQA/MOVDQU
 // during the SSE domain pass.