Revert "Revert "ARM64 asm for region space array allocation""
Also added missing large object check. No regression from the check
N6P CC EAAC time at 1313 for 10 samples vs 1314 before reverts.
Bug: 30162165
Bug: 12687968
Test: test-art-target with CC + heap poisoning
This reverts commit 6ae7f3a4541e70f04243a6fe469aa3bd51e16d79.
Change-Id: Ie28f652f619898d7d37eeebf3f31a88af8fac949
diff --git a/runtime/arch/arm64/quick_entrypoints_arm64.S b/runtime/arch/arm64/quick_entrypoints_arm64.S
index 415bb71..439f8d4 100644
--- a/runtime/arch/arm64/quick_entrypoints_arm64.S
+++ b/runtime/arch/arm64/quick_entrypoints_arm64.S
@@ -1788,7 +1788,20 @@
ONE_ARG_DOWNCALL art_quick_resolve_string, artResolveStringFromCode, RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
// Generate the allocation entrypoints for each allocator.
-GENERATE_ALLOC_ENTRYPOINTS_FOR_EACH_ALLOCATOR
+GENERATE_ALLOC_ENTRYPOINTS_FOR_NON_REGION_TLAB_ALLOCATORS
+// Comment out allocators that have arm64 specific asm.
+// GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT(_region_tlab, RegionTLAB) implemented in asm
+// GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT_RESOLVED(_region_tlab, RegionTLAB)
+// GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT_INITIALIZED(_region_tlab, RegionTLAB)
+GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT_WITH_ACCESS_CHECK(_region_tlab, RegionTLAB)
+// GENERATE_ALLOC_ENTRYPOINTS_ALLOC_ARRAY(_region_tlab, RegionTLAB) implemented in asm
+// GENERATE_ALLOC_ENTRYPOINTS_ALLOC_ARRAY_RESOLVED(_region_tlab, RegionTLAB)
+GENERATE_ALLOC_ENTRYPOINTS_ALLOC_ARRAY_WITH_ACCESS_CHECK(_region_tlab, RegionTLAB)
+GENERATE_ALLOC_ENTRYPOINTS_CHECK_AND_ALLOC_ARRAY(_region_tlab, RegionTLAB)
+GENERATE_ALLOC_ENTRYPOINTS_CHECK_AND_ALLOC_ARRAY_WITH_ACCESS_CHECK(_region_tlab, RegionTLAB)
+GENERATE_ALLOC_ENTRYPOINTS_ALLOC_STRING_FROM_BYTES(_region_tlab, RegionTLAB)
+GENERATE_ALLOC_ENTRYPOINTS_ALLOC_STRING_FROM_CHARS(_region_tlab, RegionTLAB)
+GENERATE_ALLOC_ENTRYPOINTS_ALLOC_STRING_FROM_STRING(_region_tlab, RegionTLAB)
// A hand-written override for GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT(_rosalloc, RosAlloc).
ENTRY art_quick_alloc_object_rosalloc
@@ -1895,6 +1908,71 @@
RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
END art_quick_alloc_object_rosalloc
+
+// The common fast path code for art_quick_alloc_array_region_tlab.
+.macro ALLOC_ARRAY_TLAB_FAST_PATH slowPathLabel, xClass, wClass, xCount, wCount, xTemp0, wTemp0, xTemp1, wTemp1, xTemp2, wTemp2
+ // Check null class
+ cbz \wClass, \slowPathLabel
+ ALLOC_ARRAY_TLAB_FAST_PATH_RESOLVED \slowPathLabel, \xClass, \wClass, \xCount, \wCount, \xTemp0, \wTemp0, \xTemp1, \wTemp1, \xTemp2, \wTemp2
+.endm
+
+// The common fast path code for art_quick_alloc_array_region_tlab.
+.macro ALLOC_ARRAY_TLAB_FAST_PATH_RESOLVED slowPathLabel, xClass, wClass, xCount, wCount, xTemp0, wTemp0, xTemp1, wTemp1, xTemp2, wTemp2
+ // Array classes are never finalizable or uninitialized, no need to check.
+ ldr \wTemp0, [\xClass, #MIRROR_CLASS_COMPONENT_TYPE_OFFSET] // Load component type
+ UNPOISON_HEAP_REF \wTemp0
+ ldr \wTemp0, [\xTemp0, #MIRROR_CLASS_OBJECT_PRIMITIVE_TYPE_OFFSET]
+ lsr \xTemp0, \xTemp0, #PRIMITIVE_TYPE_SIZE_SHIFT_SHIFT // Component size shift is in high 16
+ // bits.
+ // xCount is holding a 32 bit value,
+ // it can not overflow.
+ lsl \xTemp1, \xCount, \xTemp0 // Calculate data size
+ // Add array data offset and alignment.
+ add \xTemp1, \xTemp1, #(MIRROR_INT_ARRAY_DATA_OFFSET + OBJECT_ALIGNMENT_MASK)
+#if MIRROR_LONG_ARRAY_DATA_OFFSET != MIRROR_INT_ARRAY_DATA_OFFSET + 4
+#error Long array data offset must be 4 greater than int array data offset.
+#endif
+
+ add \xTemp0, \xTemp0, #1 // Add 4 to the length only if the
+ // component size shift is 3
+ // (for 64 bit alignment).
+ and \xTemp0, \xTemp0, #4
+ add \xTemp1, \xTemp1, \xTemp0
+ and \xTemp1, \xTemp1, #OBJECT_ALIGNMENT_MASK_TOGGLED // Round up the object size by the
+ // object alignment. (addr + 7) & ~7.
+ // Add by 7 is done above.
+
+ cmp \xTemp1, #MIN_LARGE_OBJECT_THRESHOLD // Possibly a large object, go slow
+ bhs \slowPathLabel // path.
+
+ ldr \xTemp0, [xSELF, #THREAD_LOCAL_POS_OFFSET] // Check tlab for space, note that
+ // we use (end - begin) to handle
+ // negative size arrays. It is
+ // assumed that a negative size will
+ // always be greater unsigned than
+ // region size.
+ ldr \xTemp2, [xSELF, #THREAD_LOCAL_END_OFFSET]
+ sub \xTemp2, \xTemp2, \xTemp0
+ cmp \xTemp1, \xTemp2
+ bhi \slowPathLabel
+
+ // "Point of no slow path". Won't go to the slow path from here on. OK to clobber x0 and x1.
+ // Move old thread_local_pos to x0
+ // for the return value.
+ mov x0, \xTemp0
+ add \xTemp0, \xTemp0, \xTemp1
+ str \xTemp0, [xSELF, #THREAD_LOCAL_POS_OFFSET] // Store new thread_local_pos.
+ ldr \xTemp0, [xSELF, #THREAD_LOCAL_OBJECTS_OFFSET] // Increment thread_local_objects.
+ add \xTemp0, \xTemp0, #1
+ str \xTemp0, [xSELF, #THREAD_LOCAL_OBJECTS_OFFSET]
+ POISON_HEAP_REF \wClass
+ str \wClass, [x0, #MIRROR_OBJECT_CLASS_OFFSET] // Store the class pointer.
+ str \wCount, [x0, #MIRROR_ARRAY_LENGTH_OFFSET] // Store the array length.
+ // Fence.
+ dmb ishst
+ ret
+.endm
+
// The common fast path code for art_quick_alloc_object_tlab and art_quick_alloc_object_region_tlab.
//
// x0: type_idx/return value, x1: ArtMethod*, x2: Class*, xSELF(x19): Thread::Current
@@ -1902,8 +1980,11 @@
// Need to preserve x0 and x1 to the slow path.
.macro ALLOC_OBJECT_TLAB_FAST_PATH slowPathLabel
cbz x2, \slowPathLabel // Check null class
- // Check class status.
- ldr w3, [x2, #MIRROR_CLASS_STATUS_OFFSET]
+ ALLOC_OBJECT_TLAB_FAST_PATH_RESOLVED \slowPathLabel
+.endm
+
+.macro ALLOC_OBJECT_TLAB_FAST_PATH_RESOLVED slowPathLabel
+ ldr w3, [x2, #MIRROR_CLASS_STATUS_OFFSET] // Check class status.
cmp x3, #MIRROR_CLASS_STATUS_INITIALIZED
bne \slowPathLabel
// Add a fake dependence from the
@@ -1916,6 +1997,10 @@
// a load-acquire for the status).
eor x3, x3, x3
add x2, x2, x3
+ ALLOC_OBJECT_TLAB_FAST_PATH_INITIALIZED \slowPathLabel
+.endm
+
+.macro ALLOC_OBJECT_TLAB_FAST_PATH_INITIALIZED slowPathLabel
// Check access flags has
// kAccClassIsFinalizable.
ldr w3, [x2, #MIRROR_CLASS_ACCESS_FLAGS_OFFSET]
@@ -1977,32 +2062,37 @@
RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
END art_quick_alloc_object_tlab
-// A hand-written override for GENERATE_ALLOC_ENTRYPOINTS_ALLOC_OBJECT(_region_tlab, RegionTLAB)
-ENTRY art_quick_alloc_object_region_tlab
+// The common code for art_quick_alloc_object_*region_tlab
+.macro GENERATE_ALLOC_OBJECT_REGION_TLAB name, entrypoint, fast_path, is_resolved
+ENTRY \name
// Fast path region tlab allocation.
- // x0: type_idx/return value, x1: ArtMethod*, xSELF(x19): Thread::Current
+ // x0: type_idx/resolved class/return value, x1: ArtMethod*, xSELF(x19): Thread::Current
+ // If is_resolved is 1 then x0 is the resolved type, otherwise it is the index.
// x2-x7: free.
#if !defined(USE_READ_BARRIER)
mvn x0, xzr // Read barrier must be enabled here.
ret // Return -1.
#endif
+.if \is_resolved
+ mov x2, x0 // class is actually stored in x0 already
+.else
ldr x2, [x1, #ART_METHOD_DEX_CACHE_TYPES_OFFSET_64] // Load dex cache resolved types array
// Load the class (x2)
ldr w2, [x2, x0, lsl #COMPRESSED_REFERENCE_SIZE_SHIFT]
-
+.endif
// Most common case: GC is not marking.
ldr w3, [xSELF, #THREAD_IS_GC_MARKING_OFFSET]
- cbnz x3, .Lart_quick_alloc_object_region_tlab_marking
-.Lart_quick_alloc_object_region_tlab_do_allocation:
- ALLOC_OBJECT_TLAB_FAST_PATH .Lart_quick_alloc_object_region_tlab_slow_path
-.Lart_quick_alloc_object_region_tlab_marking:
+ cbnz x3, .Lmarking\name
+.Ldo_allocation\name:
+ \fast_path .Lslow_path\name
+.Lmarking\name:
// GC is marking, check the lock word of the class for the mark bit.
// If the class is null, go slow path. The check is required to read the lock word.
- cbz w2, .Lart_quick_alloc_object_region_tlab_slow_path
+ cbz w2, .Lslow_path\name
// Class is not null, check mark bit in lock word.
ldr w3, [x2, #MIRROR_OBJECT_LOCK_WORD_OFFSET]
// If the bit is not zero, do the allocation.
- tbnz w3, #LOCK_WORD_MARK_BIT_SHIFT, .Lart_quick_alloc_object_region_tlab_do_allocation
+ tbnz w3, #LOCK_WORD_MARK_BIT_SHIFT, .Ldo_allocation\name
// The read barrier slow path. Mark
// the class.
stp x0, x1, [sp, #-32]! // Save registers (x0, x1, lr).
@@ -2013,14 +2103,79 @@
ldp x0, x1, [sp, #0] // Restore registers.
ldr xLR, [sp, #16]
add sp, sp, #32
- b .Lart_quick_alloc_object_region_tlab_do_allocation
-.Lart_quick_alloc_object_region_tlab_slow_path:
+ b .Ldo_allocation\name
+.Lslow_path\name:
SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // Save callee saves in case of GC.
mov x2, xSELF // Pass Thread::Current.
- bl artAllocObjectFromCodeRegionTLAB // (uint32_t type_idx, Method* method, Thread*)
+ bl \entrypoint // (uint32_t type_idx, Method* method, Thread*)
RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME
RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
-END art_quick_alloc_object_region_tlab
+END \name
+.endm
+
+GENERATE_ALLOC_OBJECT_REGION_TLAB art_quick_alloc_object_region_tlab, artAllocObjectFromCodeRegionTLAB, ALLOC_OBJECT_TLAB_FAST_PATH, 0
+GENERATE_ALLOC_OBJECT_REGION_TLAB art_quick_alloc_object_resolved_region_tlab, artAllocObjectFromCodeResolvedRegionTLAB, ALLOC_OBJECT_TLAB_FAST_PATH_RESOLVED, 1
+GENERATE_ALLOC_OBJECT_REGION_TLAB art_quick_alloc_object_initialized_region_tlab, artAllocObjectFromCodeInitializedRegionTLAB, ALLOC_OBJECT_TLAB_FAST_PATH_INITIALIZED, 1
+
+// The common code for art_quick_alloc_array_*region_tlab
+.macro GENERATE_ALLOC_ARRAY_REGION_TLAB name, entrypoint, fast_path, is_resolved
+ENTRY \name
+ // Fast path array allocation for region tlab allocation.
+ // x0: uint32_t type_idx
+ // x1: int32_t component_count
+ // x2: ArtMethod* method
+ // x3-x7: free.
+#if !defined(USE_READ_BARRIER)
+ mvn x0, xzr // Read barrier must be enabled here.
+ ret // Return -1.
+#endif
+.if \is_resolved
+ mov x3, x0
+ // If already resolved, class is stored in x0
+.else
+ ldr x3, [x2, #ART_METHOD_DEX_CACHE_TYPES_OFFSET_64] // Load dex cache resolved types array
+ // Load the class (x2)
+ ldr w3, [x3, x0, lsl #COMPRESSED_REFERENCE_SIZE_SHIFT]
+.endif
+ // Most common case: GC is not marking.
+ ldr w4, [xSELF, #THREAD_IS_GC_MARKING_OFFSET]
+ cbnz x4, .Lmarking\name
+.Ldo_allocation\name:
+ \fast_path .Lslow_path\name, x3, w3, x1, w1, x4, w4, x5, w5, x6, w6
+.Lmarking\name:
+ // GC is marking, check the lock word of the class for the mark bit.
+ // If the class is null, go slow path. The check is required to read the lock word.
+ cbz w3, .Lslow_path\name
+ // Class is not null, check mark bit in lock word.
+ ldr w4, [x3, #MIRROR_OBJECT_LOCK_WORD_OFFSET]
+ // If the bit is not zero, do the allocation.
+ tbnz w4, #LOCK_WORD_MARK_BIT_SHIFT, .Ldo_allocation\name
+ // The read barrier slow path. Mark
+ // the class.
+ stp x0, x1, [sp, #-32]! // Save registers (x0, x1, x2, lr).
+ stp x2, xLR, [sp, #16]
+ mov x0, x3 // Pass the class as the first param.
+ bl artReadBarrierMark
+ mov x3, x0 // Get the (marked) class back.
+ ldp x2, xLR, [sp, #16]
+ ldp x0, x1, [sp], #32 // Restore registers.
+ b .Ldo_allocation\name
+.Lslow_path\name:
+ // x0: uint32_t type_idx / mirror::Class* klass (if resolved)
+ // x1: int32_t component_count
+ // x2: ArtMethod* method
+ // x3: Thread* self
+ SETUP_REFS_ONLY_CALLEE_SAVE_FRAME // save callee saves in case of GC
+ mov x3, xSELF // pass Thread::Current
+ bl \entrypoint
+ RESTORE_REFS_ONLY_CALLEE_SAVE_FRAME
+ RETURN_IF_RESULT_IS_NON_ZERO_OR_DELIVER
+END \name
+.endm
+
+GENERATE_ALLOC_ARRAY_REGION_TLAB art_quick_alloc_array_region_tlab, artAllocArrayFromCodeRegionTLAB, ALLOC_ARRAY_TLAB_FAST_PATH, 0
+// TODO: art_quick_alloc_array_resolved_region_tlab seems to not get called. Investigate compiler.
+GENERATE_ALLOC_ARRAY_REGION_TLAB art_quick_alloc_array_resolved_region_tlab, artAllocArrayFromCodeResolvedRegionTLAB, ALLOC_ARRAY_TLAB_FAST_PATH_RESOLVED, 1
/*
* Called by managed code when the thread has been asked to suspend.