[X86][AVX] Update VBROADCAST folds to always use v2i64 X86vzload

The VBROADCAST combines and SimplifyDemandedVectorElts improvements mean that we now more consistently use shorter (128-bit) X86vzload input operands.

Follow up to D58053

llvm-svn: 354346

5 files changed, 6 insertions, 9 deletions

diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 0fba8cb4d09..c82d1703a0d 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -1378,7 +1378,7 @@ multiclass avx512_subvec_broadcast_rm_dq<bits<8> opc, string OpcodeStr,
 
 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:$src)))),
+  def : Pat<(v8i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
             (VPBROADCASTQZm addr:$src)>;
 }
 
@@ -1386,7 +1386,7 @@ let Predicates = [HasVLX] 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:$src)))),
+  def : Pat<(v4i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
             (VPBROADCASTQZ256m addr:$src)>;
 }
 let Predicates = [HasVLX, HasBWI] in {
diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index bc2622446bc..a55b1489e9e 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -7850,7 +7850,7 @@ let Predicates = [HasAVX2, NoVLX] in {
   // 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
   def : Pat<(v2i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
             (VPBROADCASTQrm addr:$src)>;
-  def : Pat<(v4i64 (X86VBroadcast (v4i64 (X86vzload addr:$src)))),
+  def : Pat<(v4i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
             (VPBROADCASTQYrm addr:$src)>;
 
   def : Pat<(v4i32 (X86VBroadcast (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
diff --git a/llvm/test/CodeGen/X86/insertelement-shuffle.ll b/llvm/test/CodeGen/X86/insertelement-shuffle.ll
index 8880dda4984..5b44337785e 100644
--- a/llvm/test/CodeGen/X86/insertelement-shuffle.ll
+++ b/llvm/test/CodeGen/X86/insertelement-shuffle.ll
@@ -95,8 +95,7 @@ define <8 x i64> @insert_subvector_into_undef(i32 %x0, i32 %x1) nounwind {
 ;
 ; X86_AVX512-LABEL: insert_subvector_into_undef:
 ; X86_AVX512:       # %bb.0:
-; X86_AVX512-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X86_AVX512-NEXT:    vbroadcastsd %xmm0, %zmm0
+; X86_AVX512-NEXT:    vbroadcastsd {{[0-9]+}}(%esp), %zmm0
 ; X86_AVX512-NEXT:    retl
 ;
 ; X64_AVX512-LABEL: insert_subvector_into_undef:
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
index 963fb98f56a..bedf6823277 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll
@@ -655,8 +655,7 @@ define <16 x i8> @combine_pshufb_insertion_as_broadcast_v2i64(i64 %a0) {
 define <8 x i32> @combine_permd_insertion_as_broadcast_v4i64(i64 %a0) {
 ; X86-LABEL: combine_permd_insertion_as_broadcast_v4i64:
 ; X86:       # %bb.0:
-; X86-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X86-NEXT:    vbroadcastsd %xmm0, %ymm0
+; X86-NEXT:    vbroadcastsd {{[0-9]+}}(%esp), %ymm0
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: combine_permd_insertion_as_broadcast_v4i64:
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512bw.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512bw.ll
index 6a295ba8cc5..b8efb26d6fc 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512bw.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512bw.ll
@@ -975,8 +975,7 @@ define <16 x float> @combine_vpermi2var_vpermvar_16f32_as_vperm2_zero(<16 x floa
 define <8 x i64> @combine_vpermvar_insertion_as_broadcast_v8i64(i64 %a0) {
 ; X86-LABEL: combine_vpermvar_insertion_as_broadcast_v8i64:
 ; X86:       # %bb.0:
-; X86-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X86-NEXT:    vbroadcastsd %xmm0, %zmm0
+; X86-NEXT:    vbroadcastsd {{[0-9]+}}(%esp), %zmm0
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: combine_vpermvar_insertion_as_broadcast_v8i64: