[X86, AVX] fix zero-extending integer operand load patterns to use integer instructions

This is a follow-on to r233704 and another partial fix for PR22685:
https://llvm.org/bugs/show_bug.cgi?id=22685

llvm-svn: 233724

1 files changed, 9 insertions, 8 deletions

diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index 9b134d7a690..2697a543df6 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -643,9 +643,6 @@ let Predicates = [UseAVX] in {
 
   // Represent the same patterns above but in the form they appear for
   // 256-bit types
-  def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
-                   (v4i32 (scalar_to_vector (loadi32 addr:$src))), (iPTR 0)))),
-            (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
   def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
                    (v4f32 (scalar_to_vector (loadf32 addr:$src))), (iPTR 0)))),
             (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_xmm)>;
@@ -653,9 +650,6 @@ let Predicates = [UseAVX] in {
                    (v2f64 (scalar_to_vector (loadf64 addr:$src))), (iPTR 0)))),
             (SUBREG_TO_REG (i32 0), (VMOVSDrm addr:$src), sub_xmm)>;
   }
-  def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
-                   (v2i64 (scalar_to_vector (loadi64 addr:$src))), (iPTR 0)))),
-            (SUBREG_TO_REG (i64 0), (VMOVSDrm addr:$src), sub_xmm)>;
 
   // Extract and store.
   def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
@@ -793,7 +787,7 @@ let Predicates = [UseSSE2] in {
             (MOVSDrr VR128:$src1, (COPY_TO_REGCLASS VR128:$src2, FR64))>;
 
   // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
-  // is during lowering, where it's not possible to recognize the fold cause
+  // is during lowering, where it's not possible to recognize the fold because
   // it has two uses through a bitcast. One use disappears at isel time and the
   // fold opportunity reappears.
   def : Pat<(v2f64 (X86Movlpd VR128:$src1, VR128:$src2)),
@@ -4907,7 +4901,8 @@ let Predicates = [UseAVX] in {
     def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
               (VMOVDI2PDIrr GR32:$src)>;
 
-  // AVX 128-bit movd/movq instruction write zeros in the high 128-bit part.
+  // AVX 128-bit movd/movq instructions write zeros in the high 128-bit part.
+  // These instructions also write zeros in the high part of a 256-bit register.
   let AddedComplexity = 20 in {
     def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
               (VMOVDI2PDIrm addr:$src)>;
@@ -4915,6 +4910,9 @@ let Predicates = [UseAVX] in {
               (VMOVDI2PDIrm addr:$src)>;
     def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
               (VMOVDI2PDIrm addr:$src)>;
+    def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
+                (v4i32 (scalar_to_vector (loadi32 addr:$src))), (iPTR 0)))),
+              (SUBREG_TO_REG (i32 0), (VMOVDI2PDIrm addr:$src), sub_xmm)>;
   }
   // Use regular 128-bit instructions to match 256-bit scalar_to_vec+zext.
   def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
@@ -5033,6 +5031,9 @@ let Predicates = [UseAVX], AddedComplexity = 20 in {
             (VMOVZQI2PQIrm addr:$src)>;
   def : Pat<(v2i64 (X86vzload addr:$src)),
             (VMOVZQI2PQIrm addr:$src)>;
+  def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
+              (v2i64 (scalar_to_vector (loadi64 addr:$src))), (iPTR 0)))),
+            (SUBREG_TO_REG (i64 0), (VMOVZQI2PQIrm addr:$src), sub_xmm)>;
 }
 
 let Predicates = [UseSSE2], AddedComplexity = 20 in {