[llvm] r362915 - [X86] Use EVEX instructions for f128 FAND/FOR/FXOR when avx512vl is enabled.

[Craig Topper via llvm-commits]

Author: ctopper
Date: Sun Jun  9 18:18:55 2019
New Revision: 362915

URL: http://llvm.org/viewvc/llvm-project?rev=362915&view=rev
Log:
[X86] Use EVEX instructions for f128 FAND/FOR/FXOR when avx512vl is enabled.

Modified:
    llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td

Modified: llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td?rev=362915&r1=362914&r2=362915&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td (original)
+++ llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td Sun Jun  9 18:18:55 2019
@@ -416,7 +416,7 @@ def : Pat<(f128 (X86fxor VR128:$src1, VR
           (XORPSrr VR128:$src1, VR128:$src2)>;
 }
 
-let Predicates = [HasAVX] in {
+let Predicates = [HasAVX, NoVLX] in {
 // andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
 def : Pat<(f128 (X86fand VR128:$src1, (loadf128 addr:$src2))),
           (VANDPSrm VR128:$src1, f128mem:$src2)>;
@@ -436,3 +436,24 @@ def : Pat<(f128 (X86fxor VR128:$src1, (l
 def : Pat<(f128 (X86fxor VR128:$src1, VR128:$src2)),
           (VXORPSrr VR128:$src1, VR128:$src2)>;
 }
+
+let Predicates = [HasVLX] in {
+// andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
+def : Pat<(f128 (X86fand VR128X:$src1, (loadf128 addr:$src2))),
+          (VANDPSZ128rm VR128X:$src1, f128mem:$src2)>;
+
+def : Pat<(f128 (X86fand VR128X:$src1, VR128X:$src2)),
+          (VANDPSZ128rr VR128X:$src1, VR128X:$src2)>;
+
+def : Pat<(f128 (X86for VR128X:$src1, (loadf128 addr:$src2))),
+          (VORPSZ128rm VR128X:$src1, f128mem:$src2)>;
+
+def : Pat<(f128 (X86for VR128X:$src1, VR128X:$src2)),
+          (VORPSZ128rr VR128X:$src1, VR128X:$src2)>;
+
+def : Pat<(f128 (X86fxor VR128X:$src1, (loadf128 addr:$src2))),
+          (VXORPSZ128rm VR128X:$src1, f128mem:$src2)>;
+
+def : Pat<(f128 (X86fxor VR128X:$src1, VR128X:$src2)),
+          (VXORPSZ128rr VR128X:$src1, VR128X:$src2)>;
+}