From 52c02d70e276aa1e138b9f43988de60838aacb8c Mon Sep 17 00:00:00 2001
From: Sanjay Patel <spatel@rotateright.com>
Date: Sat, 22 Dec 2018 16:59:02 +0000
Subject: [PATCH] [x86] add load fold patterns for movddup with vzext_load

The missed load folding noticed in D55898 is visible independent of that change
either with an adjusted IR pattern to start or with AVX2/AVX512 (where the build
vector becomes a broadcast first; movddup is not produced until we get into isel
via tablegen patterns).

Differential Revision: https://reviews.llvm.org/D55936

llvm-svn: 350005
---
 llvm/lib/Target/X86/X86InstrAVX512.td                 | 2 ++
 llvm/lib/Target/X86/X86InstrSSE.td                    | 6 ++++++
 llvm/test/CodeGen/X86/build-vector-128.ll             | 6 ++----
 llvm/test/CodeGen/X86/movddup-load-fold.ll            | 6 ++----
 llvm/test/CodeGen/X86/vector-shuffle-combining-xop.ll | 3 +--
 5 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 7e60b9c..105ca2e 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -11217,6 +11217,8 @@ def : Pat<(v2f64 (X86VBroadcast f64:$src)),
           (VMOVDDUPZ128rr (v2f64 (COPY_TO_REGCLASS FR64X:$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:$src0)),
diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index 94cd5a6..e2bcd18 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -4669,12 +4669,16 @@ defm MOVDDUP : sse3_replicate_dfp<"movddup", SchedWriteFShuffle>;
 let Predicates = [HasAVX, NoVLX] in {
   def : Pat<(X86Movddup (loadv2f64 addr:$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 @@ let Predicates = [HasAVX, NoVLX] in {
             (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 {
diff --git a/llvm/test/CodeGen/X86/build-vector-128.ll b/llvm/test/CodeGen/X86/build-vector-128.ll
index 1c1dab6..b80f6fa 100644
--- a/llvm/test/CodeGen/X86/build-vector-128.ll
+++ b/llvm/test/CodeGen/X86/build-vector-128.ll
@@ -526,8 +526,7 @@ define <4 x float> @PR37502(float %x, float %y) {
 ;
 ; SSE41-32-LABEL: PR37502:
 ; SSE41-32:       # %bb.0:
-; SSE41-32-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
-; SSE41-32-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
+; SSE41-32-NEXT:    movddup {{.*#+}} xmm0 = mem[0,0]
 ; SSE41-32-NEXT:    retl
 ;
 ; SSE41-64-LABEL: PR37502:
@@ -538,8 +537,7 @@ define <4 x float> @PR37502(float %x, float %y) {
 ;
 ; AVX-32-LABEL: PR37502:
 ; AVX-32:       # %bb.0:
-; AVX-32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX-32-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
+; AVX-32-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
 ; AVX-32-NEXT:    retl
 ;
 ; AVX1-64-LABEL: PR37502:
diff --git a/llvm/test/CodeGen/X86/movddup-load-fold.ll b/llvm/test/CodeGen/X86/movddup-load-fold.ll
index a0e65fb..f1af6e9 100644
--- a/llvm/test/CodeGen/X86/movddup-load-fold.ll
+++ b/llvm/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
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-xop.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-xop.ll
index 5fe0a2b..37d13d1 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-xop.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-xop.ll
@@ -332,8 +332,7 @@ define void @buildvector_v4f32_0404(float %a, float %b, <4 x float>* %ptr) {
 ; 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
 ;
-- 
2.7.4