[X86][AVX2] Fix VPBROADCASTQ folding on 32-bit targets.

As i64 isn't a value type on 32-bit targets, we need to fold the VZEXT_LOAD into VPBROADCASTQ. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295733 91177308-0d34-0410-b5e6-96231b3b80d8
stryku · Feb 21, 2017 · 3f5e9f4 · 3f5e9f4
1 parent eb973f7
commit 3f5e9f4
Show file tree

Hide file tree

Showing 3 changed files with 18 additions and 4 deletions.
diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
@@ -1030,7 +1030,18 @@ multiclass avx512_subvec_broadcast_rm<bits<8> opc, string OpcodeStr,
                             AVX5128IBase, EVEX;
 }
 
+let Predicates = [HasAVX512] in {
+  // 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
+  def : Pat<(v8i64 (X86VBroadcast (v8i64 (X86vzload addr:$src)))),
+            (VPBROADCASTQZm addr:$src)>;
+}
+
 let Predicates = [HasVLX, HasBWI] in {
+  // 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
+  def : Pat<(v2i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
+            (VPBROADCASTQZ128m addr:$src)>;
+  def : Pat<(v4i64 (X86VBroadcast (v4i64 (X86vzload addr:$src)))),
+            (VPBROADCASTQZ256m addr:$src)>;
   // loadi16 is tricky to fold, because !isTypeDesirableForOp, justifiably.
   // This means we'll encounter truncated i32 loads; match that here.
   def : Pat<(v8i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
@@ -8265,6 +8265,11 @@ defm VPBROADCASTQ  : avx2_broadcast<0x59, "vpbroadcastq", i64mem, loadi64,
                                     v2i64, v4i64, NoVLX>;
 
 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
+  // 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
+  def : Pat<(v2i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
+            (VPBROADCASTQrm addr:$src)>;
+  def : Pat<(v4i64 (X86VBroadcast (v4i64 (X86vzload addr:$src)))),
+            (VPBROADCASTQYrm addr:$src)>;
   // loadi16 is tricky to fold, because !isTypeDesirableForOp, justifiably.
   // This means we'll encounter truncated i32 loads; match that here.
   def : Pat<(v8i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),

diff --git a/test/CodeGen/X86/vector-shuffle-combining-avx2.ll b/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
@@ -803,8 +803,7 @@ define <32 x i8> @combine_unpack_unpack_pshufb(<32 x i8> %a0) {
 define <16 x i8> @combine_broadcast_pshufb_insertion_v2i64(i64 %a0) {
 ; X32-LABEL: combine_broadcast_pshufb_insertion_v2i64:
 ; X32:       # BB#0:
-; X32-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
-; X32-NEXT:    vpbroadcastq %xmm0, %xmm0
+; X32-NEXT:    vpbroadcastq {{[0-9]+}}(%esp), %xmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: combine_broadcast_pshufb_insertion_v2i64:
@@ -821,8 +820,7 @@ define <16 x i8> @combine_broadcast_pshufb_insertion_v2i64(i64 %a0) {
 define <8 x i32> @combine_broadcast_permd_insertion_v4i64(i64 %a0) {
 ; X32-LABEL: combine_broadcast_permd_insertion_v4i64:
 ; X32:       # BB#0:
-; X32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X32-NEXT:    vbroadcastsd %xmm0, %ymm0
+; X32-NEXT:    vbroadcastsd {{[0-9]+}}(%esp), %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: combine_broadcast_permd_insertion_v4i64: