Thumb2: Reduce memory used for fixup dependencies.

Mark fixups for far backward branches as 32-bit from the
beginning and ignore their dependencies since they cannot
be expanded anymore. This reduces the memory used by the
fixup_dependents_ when compiling methods with a lot of
slow paths.

Bug: 28256882

This is a cherry-pick of the fixed revert-revert
    Revert "Revert "Thumb2: Reduce memory used for fixup dependencies.""
    (cherry picked from commit a64f249add4a0a9e93aa6f246045ca79f5916dcd)

Change-Id: I9ca766b73d9f5a1da5dc10ad070145ac88fbcd53

diff --git a/compiler/utils/arm/assembler_thumb2.cc b/compiler/utils/arm/assembler_thumb2.cc
index 2c73fb8..546dd65 100644 (file)
--- a/compiler/utils/arm/assembler_thumb2.cc
+++ b/compiler/utils/arm/assembler_thumb2.cc
@@ -14,6 +14,8 @@
  * limitations under the License.
  */
 
+#include <type_traits>
+
 #include "assembler_thumb2.h"
 
 #include "base/bit_utils.h"
@@ -25,33 +27,49 @@
 namespace art {
 namespace arm {
 
-void Thumb2Assembler::Fixup::PrepareDependents(Thumb2Assembler* assembler) {
-  // For each Fixup, it's easy to find the Fixups that it depends on as they are either
-  // the following or the preceding Fixups until we find the target. However, for fixup
-  // adjustment we need the reverse lookup, i.e. what Fixups depend on a given Fixup.
-  // This function creates a compact representation of this relationship, where we have
-  // all the dependents in a single array and Fixups reference their ranges by start
-  // index and count. (Instead of having a per-fixup vector.)
-
-  // Count the number of dependents of each Fixup.
-  const FixupId end_id = assembler->fixups_.size();
+template <typename Function>
+void Thumb2Assembler::Fixup::ForExpandableDependencies(Thumb2Assembler* assembler, Function fn) {
+  static_assert(
+      std::is_same<typename std::result_of<Function(FixupId, FixupId)>::type, void>::value,
+      "Incorrect signature for argument `fn`: expected (FixupId, FixupId) -> void");
   Fixup* fixups = assembler->fixups_.data();
-  for (FixupId fixup_id = 0u; fixup_id != end_id; ++fixup_id) {
+  for (FixupId fixup_id = 0u, end_id = assembler->fixups_.size(); fixup_id != end_id; ++fixup_id) {
     uint32_t target = fixups[fixup_id].target_;
     if (target > fixups[fixup_id].location_) {
       for (FixupId id = fixup_id + 1u; id != end_id && fixups[id].location_ < target; ++id) {
-        fixups[id].dependents_count_ += 1u;
+        if (fixups[id].CanExpand()) {
+          fn(id, fixup_id);
+        }
       }
     } else {
       for (FixupId id = fixup_id; id != 0u && fixups[id - 1u].location_ >= target; --id) {
-        fixups[id - 1u].dependents_count_ += 1u;
+        if (fixups[id - 1u].CanExpand()) {
+          fn(id - 1u, fixup_id);
+        }
       }
     }
   }
+}
+
+void Thumb2Assembler::Fixup::PrepareDependents(Thumb2Assembler* assembler) {
+  // For each Fixup, it's easy to find the Fixups that it depends on as they are either
+  // the following or the preceding Fixups until we find the target. However, for fixup
+  // adjustment we need the reverse lookup, i.e. what Fixups depend on a given Fixup.
+  // This function creates a compact representation of this relationship, where we have
+  // all the dependents in a single array and Fixups reference their ranges by start
+  // index and count. (Instead of having a per-fixup vector.)
+
+  // Count the number of dependents of each Fixup.
+  Fixup* fixups = assembler->fixups_.data();
+  ForExpandableDependencies(
+      assembler,
+      [fixups](FixupId dependency, FixupId dependent ATTRIBUTE_UNUSED) {
+        fixups[dependency].dependents_count_ += 1u;
+      });
   // Assign index ranges in fixup_dependents_ to individual fixups. Record the end of the
   // range in dependents_start_, we shall later decrement it as we fill in fixup_dependents_.
   uint32_t number_of_dependents = 0u;
-  for (FixupId fixup_id = 0u; fixup_id != end_id; ++fixup_id) {
+  for (FixupId fixup_id = 0u, end_id = assembler->fixups_.size(); fixup_id != end_id; ++fixup_id) {
     number_of_dependents += fixups[fixup_id].dependents_count_;
     fixups[fixup_id].dependents_start_ = number_of_dependents;
   }
@@ -61,20 +79,12 @@ void Thumb2Assembler::Fixup::PrepareDependents(Thumb2Assembler* assembler) {
   // Create and fill in the fixup_dependents_.
   assembler->fixup_dependents_.resize(number_of_dependents);
   FixupId* dependents = assembler->fixup_dependents_.data();
-  for (FixupId fixup_id = 0u; fixup_id != end_id; ++fixup_id) {
-    uint32_t target = fixups[fixup_id].target_;
-    if (target > fixups[fixup_id].location_) {
-      for (FixupId id = fixup_id + 1u; id != end_id && fixups[id].location_ < target; ++id) {
-        fixups[id].dependents_start_ -= 1u;
-        dependents[fixups[id].dependents_start_] = fixup_id;
-      }
-    } else {
-      for (FixupId id = fixup_id; id != 0u && fixups[id - 1u].location_ >= target; --id) {
-        fixups[id - 1u].dependents_start_ -= 1u;
-        dependents[fixups[id - 1u].dependents_start_] = fixup_id;
-      }
-    }
-  }
+  ForExpandableDependencies(
+      assembler,
+      [fixups, dependents](FixupId dependency, FixupId dependent) {
+        fixups[dependency].dependents_start_ -= 1u;
+        dependents[fixups[dependency].dependents_start_] = dependent;
+      });
 }
 
 void Thumb2Assembler::BindLabel(Label* label, uint32_t bound_pc) {
@@ -115,6 +125,7 @@ void Thumb2Assembler::AdjustFixupIfNeeded(Fixup* fixup, uint32_t* current_code_s
                                           std::deque<FixupId>* fixups_to_recalculate) {
   uint32_t adjustment = fixup->AdjustSizeIfNeeded(*current_code_size);
   if (adjustment != 0u) {
+    DCHECK(fixup->CanExpand());
     *current_code_size += adjustment;
     for (FixupId dependent_id : fixup->Dependents(*this)) {
       Fixup* dependent = GetFixup(dependent_id);
@@ -2546,9 +2557,19 @@ void Thumb2Assembler::EmitBranch(Condition cond, Label* label, bool link, bool x
       }
     } else {
       branch_type = Fixup::kUnconditional;             // B.
+      // The T2 encoding offset is `SignExtend(imm11:'0', 32)` and there is a PC adjustment of 4.
+      static constexpr size_t kMaxT2BackwardDistance = (1u << 11) - 4u;
+      if (!use32bit && label->IsBound() && pc - label->Position() > kMaxT2BackwardDistance) {
+        use32bit = true;
+      }
     }
   } else {
     branch_type = Fixup::kConditional;                 // B<cond>.
+    // The T1 encoding offset is `SignExtend(imm8:'0', 32)` and there is a PC adjustment of 4.
+    static constexpr size_t kMaxT1BackwardDistance = (1u << 8) - 4u;
+    if (!use32bit && label->IsBound() && pc - label->Position() > kMaxT1BackwardDistance) {
+      use32bit = true;
+    }
   }
 
   Fixup::Size size = use32bit ? Fixup::kBranch32Bit : Fixup::kBranch16Bit;

diff --git a/compiler/utils/arm/assembler_thumb2.h b/compiler/utils/arm/assembler_thumb2.h
index 111a6b0..ce310a4 100644 (file)
--- a/compiler/utils/arm/assembler_thumb2.h
+++ b/compiler/utils/arm/assembler_thumb2.h
@@ -538,6 +538,20 @@ class Thumb2Assembler FINAL : public ArmAssembler {
       return GetType() >= kLoadLiteralNarrow;
     }
 
+    // Returns whether the Fixup can expand from the original size.
+    bool CanExpand() const {
+      switch (GetOriginalSize()) {
+        case kBranch32Bit:
+        case kCbxz48Bit:
+        case kLiteralFar:
+        case kLiteralAddrFar:
+        case kLongOrFPLiteralFar:
+          return false;
+        default:
+          return true;
+      }
+    }
+
     Size GetOriginalSize() const {
       return original_size_;
     }
@@ -611,6 +625,7 @@ class Thumb2Assembler FINAL : public ArmAssembler {
           dependents_count_(0u),
           dependents_start_(0u) {
     }
+
     static size_t SizeInBytes(Size size);
 
     // The size of padding added before the literal pool.
@@ -623,6 +638,9 @@ class Thumb2Assembler FINAL : public ArmAssembler {
 
     int32_t LoadWideOrFpEncoding(Register rbase, int32_t offset) const;
 
+    template <typename Function>
+    static void ForExpandableDependencies(Thumb2Assembler* assembler, Function fn);
+
     static constexpr uint32_t kUnresolved = 0xffffffff;     // Value for target_ for unresolved.
 
     const Register rn_;   // Rn for cbnz/cbz, Rt for literal loads.

diff --git a/compiler/utils/arm/assembler_thumb2_test.cc b/compiler/utils/arm/assembler_thumb2_test.cc
index 650b089..b5cafcb 100644 (file)
--- a/compiler/utils/arm/assembler_thumb2_test.cc
+++ b/compiler/utils/arm/assembler_thumb2_test.cc
@@ -372,6 +372,31 @@ TEST_F(AssemblerThumb2Test, StoreWordPairToNonThumbOffset) {
   DriverStr(expected, "StoreWordPairToNonThumbOffset");
 }
 
+TEST_F(AssemblerThumb2Test, DistantBackBranch) {
+  Label start, end;
+  __ Bind(&start);
+  constexpr size_t kLdrR0R0Count1 = 256;
+  for (size_t i = 0; i != kLdrR0R0Count1; ++i) {
+    __ ldr(arm::R0, arm::Address(arm::R0));
+  }
+  __ b(&end, arm::EQ);
+  __ b(&start, arm::LT);
+  constexpr size_t kLdrR0R0Count2 = 256;
+  for (size_t i = 0; i != kLdrR0R0Count2; ++i) {
+    __ ldr(arm::R0, arm::Address(arm::R0));
+  }
+  __ Bind(&end);
+
+  std::string expected =
+      "0:\n" +
+      RepeatInsn(kLdrR0R0Count1, "ldr r0, [r0]\n") +
+      "beq 1f\n"
+      "blt 0b\n" +
+      RepeatInsn(kLdrR0R0Count2, "ldr r0, [r0]\n") +
+      "1:\n";
+  DriverStr(expected, "DistantBackBranch");
+}
+
 TEST_F(AssemblerThumb2Test, TwoCbzMaxOffset) {
   Label label0, label1, label2;
   __ cbz(arm::R0, &label1);