[PATCH] D55936: [x86] add load fold patterns for movddup with vzext_load

[Sanjay Patel via Phabricator via llvm-commits]

spatel updated this revision to Diff 179292.
spatel added a comment.

Patch updated:
Add pattern for AVX512VL. I moved the dedicated test for this bug (rL349920 <https://reviews.llvm.org/rL349920>), so we'd have a single file with the relevant RUN lines without adding a bunch of noise.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D55936/new/

https://reviews.llvm.org/D55936

Files:
  lib/Target/X86/X86InstrAVX512.td
  lib/Target/X86/X86InstrSSE.td
  test/CodeGen/X86/build-vector-128.ll
  test/CodeGen/X86/movddup-load-fold.ll
  test/CodeGen/X86/vector-shuffle-combining-xop.ll


Index: test/CodeGen/X86/vector-shuffle-combining-xop.ll
===================================================================
--- test/CodeGen/X86/vector-shuffle-combining-xop.ll
+++ test/CodeGen/X86/vector-shuffle-combining-xop.ll
@@ -332,8 +332,7 @@
 ; X86AVX2-LABEL: buildvector_v4f32_0404:
 ; X86AVX2:       # %bb.0:
 ; X86AVX2-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86AVX2-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X86AVX2-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
+; X86AVX2-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
 ; X86AVX2-NEXT:    vmovapd %xmm0, (%eax)
 ; X86AVX2-NEXT:    retl
 ;
Index: test/CodeGen/X86/movddup-load-fold.ll
===================================================================
--- test/CodeGen/X86/movddup-load-fold.ll
+++ test/CodeGen/X86/movddup-load-fold.ll
@@ -9,14 +9,12 @@
 define <4 x float> @movddup_load_fold(float %x, float %y) {
 ; SSE-LABEL: movddup_load_fold:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
-; SSE-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
+; SSE-NEXT:    movddup {{.*#+}} xmm0 = mem[0,0]
 ; SSE-NEXT:    retl
 ;
 ; AVX-LABEL: movddup_load_fold:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
+; AVX-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
 ; AVX-NEXT:    retl
   %i0 = insertelement <4 x float> zeroinitializer, float %x, i32 0
   %i1 = insertelement <4 x float> %i0, float %y, i32 1
Index: test/CodeGen/X86/build-vector-128.ll
===================================================================
--- test/CodeGen/X86/build-vector-128.ll
+++ test/CodeGen/X86/build-vector-128.ll
@@ -556,8 +556,7 @@
 ;
 ; AVX2-32-LABEL: PR37502:
 ; AVX2-32:       # %bb.0:
-; AVX2-32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX2-32-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
+; AVX2-32-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
 ; AVX2-32-NEXT:    retl
 ;
 ; AVX2-64-LABEL: PR37502:
Index: lib/Target/X86/X86InstrSSE.td
===================================================================
--- lib/Target/X86/X86InstrSSE.td
+++ lib/Target/X86/X86InstrSSE.td
@@ -4669,12 +4669,16 @@
 let Predicates = [HasAVX, NoVLX] in {
   def : Pat<(X86Movddup (loadv2f64 addr:$src)),
             (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
+  def : Pat<(X86Movddup (v2f64 (X86vzload addr:$src))),
+            (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
 }
 
 let Predicates = [UseSSE3] in {
   // No need for aligned memory as this only loads 64-bits.
   def : Pat<(X86Movddup (loadv2f64 addr:$src)),
             (MOVDDUPrm addr:$src)>;
+  def : Pat<(X86Movddup (v2f64 (X86vzload addr:$src))),
+            (MOVDDUPrm addr:$src)>;
 }
 
 //===---------------------------------------------------------------------===//
@@ -8034,6 +8038,8 @@
             (VMOVDDUPrr VR128:$src)>;
   def : Pat<(v2f64 (X86VBroadcast (loadv2f64 addr:$src))),
             (VMOVDDUPrm addr:$src)>;
+  def : Pat<(v2f64 (X86VBroadcast (v2f64 (X86vzload addr:$src)))),
+            (VMOVDDUPrm addr:$src)>;
 }
 
 let Predicates = [HasAVX1Only] in {
Index: lib/Target/X86/X86InstrAVX512.td
===================================================================
--- lib/Target/X86/X86InstrAVX512.td
+++ lib/Target/X86/X86InstrAVX512.td
@@ -11217,6 +11217,8 @@
           (VMOVDDUPZ128rr (v2f64 (COPY_TO_REGCLASS FR64X:$src, VR128X)))>;
 def : Pat<(v2f64 (X86VBroadcast (loadv2f64 addr:$src))),
           (VMOVDDUPZ128rm addr:$src)>;
+def : Pat<(v2f64 (X86VBroadcast (v2f64 (X86vzload addr:$src)))),
+          (VMOVDDUPZ128rm addr:$src)>;
 
 def : Pat<(vselect (v2i1 VK2WM:$mask), (v2f64 (X86VBroadcast f64:$src)),
                    (v2f64 VR128X:$src0)),


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D55936.179292.patch
Type: text/x-patch
Size: 3684 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20181221/e523d050/attachment.bin>