From 5f96ea6159852c07b5652ad38ace6b0aab375441 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Tue, 17 Jul 2018 20:16:18 +0000
Subject: [PATCH] [X86] Add patterns for folding full vector load into MOVHPS
 and MOVLPS with SSE1 only.

llvm-svn: 337320
---
 lib/Target/X86/X86InstrAVX512.td        |  4 ++-
 lib/Target/X86/X86InstrSSE.td           | 37 +++++++++++++++----------
 test/CodeGen/X86/vector-shuffle-sse1.ll |  4 +--
 3 files changed, 26 insertions(+), 19 deletions(-)

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 3a1f840139e..2035e49720f 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -6452,7 +6452,9 @@ multiclass avx512_mov_hilo_packed<bits<8> opc, string OpcodeStr,
                   Sched<[SchedWriteFShuffle.XMM.Folded, ReadAfterLd]>, EVEX_4V;
 }
 
-defm VMOVHPSZ128 : avx512_mov_hilo_packed<0x16, "vmovhps", X86Movlhps,
+// No patterns for MOVLPS/MOVHPS as the Movlhps node should only be created in
+// SSE1. And MOVLPS pattern is even more complex.
+defm VMOVHPSZ128 : avx512_mov_hilo_packed<0x16, "vmovhps", null_frag,
                                   v4f32x_info>, EVEX_CD8<32, CD8VT2>, PS;
 defm VMOVHPDZ128 : avx512_mov_hilo_packed<0x16, "vmovhpd", X86Unpckl,
                                   v2f64x_info>, EVEX_CD8<64, CD8VT1>, PD, VEX_W;
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 448be0eda0e..3797d91fb31 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -661,19 +661,16 @@ let Predicates = [UseSSE1] in {
 // SSE 1 & 2 - Move Low packed FP Instructions
 //===----------------------------------------------------------------------===//
 
-multiclass sse12_mov_hilo_packed_base<bits<8>opc, SDNode psnode, SDNode pdnode,
+multiclass sse12_mov_hilo_packed_base<bits<8>opc,  SDNode pdnode,
                                       string base_opc, string asm_opr> {
+  // No pattern as they need be special cased between high and low.
   let hasSideEffects = 0, mayLoad = 1 in
   def PSrm : PI<opc, MRMSrcMem,
-         (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
-         !strconcat(base_opc, "s", asm_opr),
-     [(set VR128:$dst,
-       (psnode VR128:$src1,
-              (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))],
-              SSEPackedSingle>, PS,
-     Sched<[SchedWriteFShuffle.XMM.Folded, ReadAfterLd]>;
+                (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
+                !strconcat(base_opc, "s", asm_opr),
+                [], SSEPackedSingle>, PS,
+                Sched<[SchedWriteFShuffle.XMM.Folded, ReadAfterLd]>;
 
-  let hasSideEffects = 0, mayLoad = 1 in
   def PDrm : PI<opc, MRMSrcMem,
          (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
          !strconcat(base_opc, "d", asm_opr),
@@ -683,19 +680,19 @@ multiclass sse12_mov_hilo_packed_base<bits<8>opc, SDNode psnode, SDNode pdnode,
      Sched<[SchedWriteFShuffle.XMM.Folded, ReadAfterLd]>;
 }
 
-multiclass sse12_mov_hilo_packed<bits<8>opc, SDPatternOperator psnode,
-                                 SDPatternOperator pdnode, string base_opc> {
+multiclass sse12_mov_hilo_packed<bits<8>opc, SDPatternOperator pdnode,
+                                 string base_opc> {
   let Predicates = [UseAVX] in
-    defm V#NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
+    defm V#NAME : sse12_mov_hilo_packed_base<opc, pdnode, base_opc,
                                     "\t{$src2, $src1, $dst|$dst, $src1, $src2}">,
                                     VEX_4V, VEX_WIG;
 
   let Constraints = "$src1 = $dst" in
-    defm NAME : sse12_mov_hilo_packed_base<opc, psnode, pdnode, base_opc,
+    defm NAME : sse12_mov_hilo_packed_base<opc,  pdnode, base_opc,
                                     "\t{$src2, $dst|$dst, $src2}">;
 }
 
-defm MOVL : sse12_mov_hilo_packed<0x12, null_frag, X86Movsd, "movlp">;
+defm MOVL : sse12_mov_hilo_packed<0x12, X86Movsd, "movlp">;
 
 let SchedRW = [WriteFStore] in {
 let Predicates = [UseAVX] in {
@@ -725,13 +722,18 @@ let Predicates = [UseSSE1] in {
   def : Pat<(store (i64 (extractelt (bc_v2i64 (v4f32 VR128:$src2)),
                                  (iPTR 0))), addr:$src1),
             (MOVLPSmr addr:$src1, VR128:$src2)>;
+
+  // This pattern helps select MOVLPS on SSE1 only targets. With SSE2 we'll
+  // end up with a movsd or bleand instead of shufp.
+  def : Pat<(X86Shufp (memopv4f32 addr:$src2), VR128:$src1, (i8 -28)),
+            (MOVLPSrm VR128:$src1, addr:$src2)>;
 }
 
 //===----------------------------------------------------------------------===//
 // SSE 1 & 2 - Move Hi packed FP Instructions
 //===----------------------------------------------------------------------===//
 
-defm MOVH : sse12_mov_hilo_packed<0x16, X86Movlhps, X86Unpckl, "movhp">;
+defm MOVH : sse12_mov_hilo_packed<0x16, X86Unpckl, "movhp">;
 
 let SchedRW = [WriteFStore] in {
 // v2f64 extract element 1 is always custom lowered to unpack high to low
@@ -796,6 +798,11 @@ let Predicates = [UseSSE1] in {
   def : Pat<(X86Movlhps VR128:$src1,
                  (bc_v4f32 (v2i64 (X86vzload addr:$src2)))),
             (MOVHPSrm VR128:$src1, addr:$src2)>;
+
+  // This pattern helps select MOVHPS on SSE1 only targets. With SSE2 we'll
+  // end up with a movsd or bleand instead of shufp.
+  def : Pat<(X86Movlhps VR128:$src1, (memopv4f32 addr:$src2)),
+            (MOVHPSrm VR128:$src1, addr:$src2)>;
 }
 
 let Predicates = [UseSSE2] in {
diff --git a/test/CodeGen/X86/vector-shuffle-sse1.ll b/test/CodeGen/X86/vector-shuffle-sse1.ll
index 5f8c21bd852..dda46e062d5 100644
--- a/test/CodeGen/X86/vector-shuffle-sse1.ll
+++ b/test/CodeGen/X86/vector-shuffle-sse1.ll
@@ -298,9 +298,7 @@ define <4 x float> @shuffle_mem_v4f32_6723(<4 x float> %a, <4 x float>* %pb) {
 define <4 x float> @shuffle_mem_v4f32_4523(<4 x float> %a, <4 x float>* %pb) {
 ; SSE1-LABEL: shuffle_mem_v4f32_4523:
 ; SSE1:       # %bb.0:
-; SSE1-NEXT:    movaps (%rdi), %xmm1
-; SSE1-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,3]
-; SSE1-NEXT:    movaps %xmm1, %xmm0
+; SSE1-NEXT:    movlps {{.*#+}} xmm0 = mem[0,1],xmm0[2,3]
 ; SSE1-NEXT:    retq
   %b = load <4 x float>, <4 x float>* %pb, align 16
   %shuffle = shufflevector <4 x float> %a, <4 x float> %b, <4 x i32> <i32 4, i32 5, i32 2, i32 3>