[AArch64][GlobalISel] Split vector stores of zero.

This results in a very minor improvement in most cases, generating
stores of xzr instead of moving zero to a vector register.

Differential Revision: https://reviews.llvm.org/D115479

diff --git a/llvm/lib/Target/AArch64/AArch64Combine.td b/llvm/lib/Target/AArch64/AArch64Combine.td
index d2097f7..1994e0e 100644 (file)
--- a/llvm/lib/Target/AArch64/AArch64Combine.td
+++ b/llvm/lib/Target/AArch64/AArch64Combine.td
@@ -196,6 +196,13 @@ def mutate_anyext_to_zext : GICombineRule<
   (apply [{ applyMutateAnyExtToZExt(*${d}, MRI, B, Observer); }])
 >;
 
+def split_store_zero_128 : GICombineRule<
+  (defs root:$d),
+  (match (wip_match_opcode G_STORE):$d,
+          [{ return matchSplitStoreZero128(*${d}, MRI); }]),
+  (apply [{ applySplitStoreZero128(*${d}, MRI, B, Observer); }])
+>;
+
 // Post-legalization combines which should happen at all optimization levels.
 // (E.g. ones that facilitate matching for the selector) For example, matching
 // pseudos.
@@ -220,6 +227,7 @@ def AArch64PostLegalizerCombinerHelper
                         icmp_to_true_false_known_bits, merge_unmerge,
                         select_combines, fold_merge_to_zext,
                         constant_fold, identity_combines,
-                        ptr_add_immed_chain, overlapping_and]> {
+                        ptr_add_immed_chain, overlapping_and,
+                        split_store_zero_128]> {
   let DisableRuleOption = "aarch64postlegalizercombiner-disable-rule";
 }

diff --git a/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerCombiner.cpp b/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerCombiner.cpp
index a9b3792..5316a01 100644 (file)
--- a/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerCombiner.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerCombiner.cpp
@@ -289,6 +289,44 @@ static void applyMutateAnyExtToZExt(MachineInstr &MI, MachineRegisterInfo &MRI,
   Observer.changedInstr(MI);
 }
 
+/// Match a 128b store of zero and split it into two 64 bit stores, for
+/// size/performance reasons.
+static bool matchSplitStoreZero128(MachineInstr &MI, MachineRegisterInfo &MRI) {
+  GStore &Store = cast<GStore>(MI);
+  if (!Store.isSimple())
+    return false;
+  LLT ValTy = MRI.getType(Store.getValueReg());
+  if (!ValTy.isVector() || ValTy.getSizeInBits() != 128)
+    return false;
+  if (ValTy.getSizeInBits() != Store.getMemSizeInBits())
+    return false; // Don't split truncating stores.
+  if (!MRI.hasOneNonDBGUse(Store.getValueReg()))
+    return false;
+  auto MaybeCst = isConstantOrConstantSplatVector(
+      *MRI.getVRegDef(Store.getValueReg()), MRI);
+  return MaybeCst && MaybeCst->isZero();
+}
+
+static void applySplitStoreZero128(MachineInstr &MI, MachineRegisterInfo &MRI,
+                                   MachineIRBuilder &B,
+                                   GISelChangeObserver &Observer) {
+  B.setInstrAndDebugLoc(MI);
+  GStore &Store = cast<GStore>(MI);
+  LLT ValTy = MRI.getType(Store.getValueReg());
+  assert(ValTy.isVector() && "Expected a vector store value");
+  LLT NewTy = LLT::scalar(64);
+  Register PtrReg = Store.getPointerReg();
+  auto Zero = B.buildConstant(NewTy, 0);
+  auto HighPtr = B.buildPtrAdd(MRI.getType(PtrReg), PtrReg,
+                               B.buildConstant(LLT::scalar(64), 8));
+  auto &MF = *MI.getMF();
+  auto *LowMMO = MF.getMachineMemOperand(&Store.getMMO(), 0, NewTy);
+  auto *HighMMO = MF.getMachineMemOperand(&Store.getMMO(), 8, NewTy);
+  B.buildStore(Zero, PtrReg, *LowMMO);
+  B.buildStore(Zero, HighPtr, *HighMMO);
+  Store.eraseFromParent();
+}
+
 #define AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_DEPS
 #include "AArch64GenPostLegalizeGICombiner.inc"
 #undef AARCH64POSTLEGALIZERCOMBINERHELPER_GENCOMBINERHELPER_DEPS

diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/postlegalizer-combiner-split-zero-stores.mir b/llvm/test/CodeGen/AArch64/GlobalISel/postlegalizer-combiner-split-zero-stores.mir
new file mode 100644 (file)
index 0000000..b4c856c
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/postlegalizer-combiner-split-zero-stores.mir
@@ -0,0 +1,200 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple aarch64 -run-pass=aarch64-postlegalizer-combiner -verify-machineinstrs %s -o - | FileCheck %s
+
+...
+---
+name:            v2s64_split
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $x0
+
+    ; Split a store of <2 x i64> into two scalar stores.
+
+    ; CHECK-LABEL: name: v2s64_split
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CHECK-NEXT: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; CHECK-NEXT: G_STORE %zero(s64), [[COPY]](p0) :: (store (s64), align 16)
+    ; CHECK-NEXT: G_STORE %zero(s64), [[PTR_ADD]](p0) :: (store (s64) into unknown-address + 8)
+    ; CHECK-NEXT: RET_ReallyLR
+    %0:_(p0) = COPY $x0
+    %zero:_(s64) = G_CONSTANT i64 0
+    %zerovec:_(<2 x s64>) = G_BUILD_VECTOR %zero, %zero
+    G_STORE %zerovec(<2 x s64>), %0(p0) :: (store (<2 x s64>))
+    RET_ReallyLR
+
+...
+---
+name:            v4i32_split
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: v4i32_split
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CHECK-NEXT: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C1]](s64)
+    ; CHECK-NEXT: G_STORE [[C]](s64), [[COPY]](p0) :: (store (s64), align 16)
+    ; CHECK-NEXT: G_STORE [[C]](s64), [[PTR_ADD]](p0) :: (store (s64) into unknown-address + 8)
+    ; CHECK-NEXT: RET_ReallyLR
+    %0:_(p0) = COPY $x0
+    %zero:_(s32) = G_CONSTANT i32 0
+    %zerovec:_(<4 x s32>) = G_BUILD_VECTOR %zero, %zero, %zero, %zero
+    G_STORE %zerovec(<4 x s32>), %0(p0) :: (store (<4 x s32>))
+    RET_ReallyLR
+
+...
+---
+name:            v8i16_split
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: v8i16_split
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+    ; CHECK-NEXT: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C1]](s64)
+    ; CHECK-NEXT: G_STORE [[C]](s64), [[COPY]](p0) :: (store (s64), align 16)
+    ; CHECK-NEXT: G_STORE [[C]](s64), [[PTR_ADD]](p0) :: (store (s64) into unknown-address + 8)
+    ; CHECK-NEXT: RET_ReallyLR
+    %0:_(p0) = COPY $x0
+    %zero:_(s16) = G_CONSTANT i16 0
+    %zerovec:_(<8 x s16>) = G_BUILD_VECTOR %zero, %zero, %zero, %zero, %zero, %zero, %zero, %zero
+    G_STORE %zerovec(<8 x s16>), %0(p0) :: (store (<8 x s16>))
+    RET_ReallyLR
+
+...
+
+# Negative tests
+---
+name:            v2i32_nosplit
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: v2i32_nosplit
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+    ; CHECK-NEXT: %zerovec:_(<2 x s32>) = G_BUILD_VECTOR %zero(s32), %zero(s32)
+    ; CHECK-NEXT: G_STORE %zerovec(<2 x s32>), [[COPY]](p0) :: (store (<2 x s32>))
+    ; CHECK-NEXT: RET_ReallyLR
+    %0:_(p0) = COPY $x0
+    %zero:_(s32) = G_CONSTANT i32 0
+    %zerovec:_(<2 x s32>) = G_BUILD_VECTOR %zero, %zero
+    G_STORE %zerovec(<2 x s32>), %0(p0) :: (store (<2 x s32>))
+    RET_ReallyLR
+
+...
+---
+name:            multiple_uses
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: multiple_uses
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: %zerovec:_(<2 x s64>) = G_BUILD_VECTOR %zero(s64), %zero(s64)
+    ; CHECK-NEXT: G_STORE %zerovec(<2 x s64>), [[COPY]](p0) :: (store (<2 x s64>))
+    ; CHECK-NEXT: $q0 = COPY %zerovec(<2 x s64>)
+    ; CHECK-NEXT: RET_ReallyLR
+    %0:_(p0) = COPY $x0
+    %zero:_(s64) = G_CONSTANT i64 0
+    %zerovec:_(<2 x s64>) = G_BUILD_VECTOR %zero, %zero
+    G_STORE %zerovec(<2 x s64>), %0(p0) :: (store (<2 x s64>))
+    $q0 = COPY %zerovec
+    RET_ReallyLR
+
+...
+---
+name:            truncating
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: truncating
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: %zerovec:_(<2 x s64>) = G_BUILD_VECTOR %zero(s64), %zero(s64)
+    ; CHECK-NEXT: G_STORE %zerovec(<2 x s64>), [[COPY]](p0) :: (store (<2 x s32>))
+    ; CHECK-NEXT: RET_ReallyLR
+    %0:_(p0) = COPY $x0
+    %zero:_(s64) = G_CONSTANT i64 0
+    %zerovec:_(<2 x s64>) = G_BUILD_VECTOR %zero, %zero
+    G_STORE %zerovec(<2 x s64>), %0(p0) :: (store (<2 x s32>))
+    RET_ReallyLR
+
+...
+---
+name:            volatile
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: volatile
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: %zerovec:_(<2 x s64>) = G_BUILD_VECTOR %zero(s64), %zero(s64)
+    ; CHECK-NEXT: G_STORE %zerovec(<2 x s64>), [[COPY]](p0) :: (volatile store (<4 x s32>))
+    ; CHECK-NEXT: RET_ReallyLR
+    %0:_(p0) = COPY $x0
+    %zero:_(s64) = G_CONSTANT i64 0
+    %zerovec:_(<2 x s64>) = G_BUILD_VECTOR %zero, %zero
+    G_STORE %zerovec(<2 x s64>), %0(p0) :: (volatile store (<4 x s32>))
+    RET_ReallyLR
+
+...
+---
+name:            s128_scalar
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.1.entry:
+    liveins: $x0
+
+    ; Split a store of <2 x i64> into two scalar stores.
+
+    ; CHECK-LABEL: name: s128_scalar
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK-NEXT: %zero:_(s128) = G_CONSTANT i128 0
+    ; CHECK-NEXT: G_STORE %zero(s128), [[COPY]](p0) :: (store (s128))
+    ; CHECK-NEXT: RET_ReallyLR
+    %0:_(p0) = COPY $x0
+    %zero:_(s128) = G_CONSTANT i128 0
+    G_STORE %zero(s128), %0(p0) :: (store (s128))
+    RET_ReallyLR
+
+...