Skip to content

V
V8
Commit d21acd91 authored by alexandre.rames@arm.com's avatar alexandre.rames@arm.com
Browse files
Options

ARM64: Fix and improve atomic operations. 

* The 'compare and swap' operations should enforce memory ordering even when
  the exchange does not occur.

* The exclusive monitor does not need to be cleared by CLREX if a LDRX was
  not followed by a matching STREX.

* Use LDAR and STLR where possible.

* Use the 'I' and 'J' constraints to hint for constants valid for immediate
  values.

R=jfb@chromium.org, rmcilroy@chromium.org, ulan@chromium.org

Review URL: https://codereview.chromium.org/220793002

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@20446 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
parent 0f52462e
Hide whitespace changes
Inline Side-by-side
Showing with 56 additions and 97 deletions
src/atomicops_internals_arm64_gcc.h
View file @ d21acd91
...@@ -34,12 +34,16 @@ namespace v8 { ...@@ -34,12 +34,16 @@ namespace v8 {
namespace internal { namespace internal {
inline void MemoryBarrier() { inline void MemoryBarrier() {
__asm__ __volatile__ ( // NOLINT __asm__ __volatile__ ("dmb ish" ::: "memory"); // NOLINT
"dmb ish \n\t" // Data memory barrier.
::: "memory"
); // NOLINT
} }
// NoBarrier versions of the operation include "memory" in the clobber list.
// This is not required for direct usage of the NoBarrier versions of the
// operations. However this is required for correctness when they are used as
// part of the Acquire or Release versions, to ensure that nothing from outside
// the call is reordered between the operation and the memory barrier. This does
// not change the code generated, so has no or minimal impact on the
// NoBarrier operations.
inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr, inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr,
Atomic32 old_value, Atomic32 old_value,
...@@ -55,13 +59,12 @@ inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr, ...@@ -55,13 +59,12 @@ inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr,
"stxr %w[temp], %w[new_value], %[ptr] \n\t" // Try to store the new value. "stxr %w[temp], %w[new_value], %[ptr] \n\t" // Try to store the new value.
"cbnz %w[temp], 0b \n\t" // Retry if it did not work. "cbnz %w[temp], 0b \n\t" // Retry if it did not work.
"1: \n\t" "1: \n\t"
"clrex \n\t" // In case we didn't swap.
: [prev]"=&r" (prev), : [prev]"=&r" (prev),
[temp]"=&r" (temp), [temp]"=&r" (temp),
[ptr]"+Q" (*ptr) [ptr]"+Q" (*ptr)
: [old_value]"r" (old_value), : [old_value]"IJr" (old_value),
[new_value]"r" (new_value) [new_value]"r" (new_value)
: "memory", "cc" : "cc", "memory"
); // NOLINT ); // NOLINT
return prev; return prev;
...@@ -101,7 +104,7 @@ inline Atomic32 NoBarrier_AtomicIncrement(volatile Atomic32* ptr, ...@@ -101,7 +104,7 @@ inline Atomic32 NoBarrier_AtomicIncrement(volatile Atomic32* ptr,
: [result]"=&r" (result), : [result]"=&r" (result),
[temp]"=&r" (temp), [temp]"=&r" (temp),
[ptr]"+Q" (*ptr) [ptr]"+Q" (*ptr)
: [increment]"r" (increment) : [increment]"IJr" (increment)
: "memory" : "memory"
); // NOLINT ); // NOLINT
...@@ -110,8 +113,10 @@ inline Atomic32 NoBarrier_AtomicIncrement(volatile Atomic32* ptr, ...@@ -110,8 +113,10 @@ inline Atomic32 NoBarrier_AtomicIncrement(volatile Atomic32* ptr,
inline Atomic32 Barrier_AtomicIncrement(volatile Atomic32* ptr, inline Atomic32 Barrier_AtomicIncrement(volatile Atomic32* ptr,
Atomic32 increment) { Atomic32 increment) {
Atomic32 result;
MemoryBarrier(); MemoryBarrier();
Atomic32 result = NoBarrier_AtomicIncrement(ptr, increment); result = NoBarrier_AtomicIncrement(ptr, increment);
MemoryBarrier(); MemoryBarrier();
return result; return result;
...@@ -121,27 +126,9 @@ inline Atomic32 Acquire_CompareAndSwap(volatile Atomic32* ptr, ...@@ -121,27 +126,9 @@ inline Atomic32 Acquire_CompareAndSwap(volatile Atomic32* ptr,
Atomic32 old_value, Atomic32 old_value,
Atomic32 new_value) { Atomic32 new_value) {
Atomic32 prev; Atomic32 prev;
int32_t temp;
__asm__ __volatile__ ( // NOLINT prev = NoBarrier_CompareAndSwap(ptr, old_value, new_value);
"0: \n\t" MemoryBarrier();
"ldxr %w[prev], %[ptr] \n\t" // Load the previous value.
"cmp %w[prev], %w[old_value] \n\t"
"bne 1f \n\t"
"stxr %w[temp], %w[new_value], %[ptr] \n\t" // Try to store the new value.
"cbnz %w[temp], 0b \n\t" // Retry if it did not work.
"dmb ish \n\t" // Data memory barrier.
"1: \n\t"
// If the compare failed the 'dmb' is unnecessary, but we still need a
// 'clrex'.
"clrex \n\t"
: [prev]"=&r" (prev),
[temp]"=&r" (temp),
[ptr]"+Q" (*ptr)
: [old_value]"r" (old_value),
[new_value]"r" (new_value)
: "memory", "cc"
); // NOLINT
return prev; return prev;
} }
...@@ -150,27 +137,9 @@ inline Atomic32 Release_CompareAndSwap(volatile Atomic32* ptr, ...@@ -150,27 +137,9 @@ inline Atomic32 Release_CompareAndSwap(volatile Atomic32* ptr,
Atomic32 old_value, Atomic32 old_value,
Atomic32 new_value) { Atomic32 new_value) {
Atomic32 prev; Atomic32 prev;
int32_t temp;
MemoryBarrier(); MemoryBarrier();
prev = NoBarrier_CompareAndSwap(ptr, old_value, new_value);
__asm__ __volatile__ ( // NOLINT
"0: \n\t"
"ldxr %w[prev], %[ptr] \n\t" // Load the previous value.
"cmp %w[prev], %w[old_value] \n\t"
"bne 1f \n\t"
"stxr %w[temp], %w[new_value], %[ptr] \n\t" // Try to store the new value.
"cbnz %w[temp], 0b \n\t" // Retry if it did not work.
"1: \n\t"
// If the compare failed the we still need a 'clrex'.
"clrex \n\t"
: [prev]"=&r" (prev),
[temp]"=&r" (temp),
[ptr]"+Q" (*ptr)
: [old_value]"r" (old_value),
[new_value]"r" (new_value)
: "memory", "cc"
); // NOLINT
return prev; return prev;
} }
...@@ -185,8 +154,12 @@ inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value) { ...@@ -185,8 +154,12 @@ inline void Acquire_Store(volatile Atomic32* ptr, Atomic32 value) {
} }
inline void Release_Store(volatile Atomic32* ptr, Atomic32 value) { inline void Release_Store(volatile Atomic32* ptr, Atomic32 value) {
MemoryBarrier(); __asm__ __volatile__ ( // NOLINT
*ptr = value; "stlr %w[value], %[ptr] \n\t"
: [ptr]"=Q" (*ptr)
: [value]"r" (value)
: "memory"
); // NOLINT
} }
inline Atomic32 NoBarrier_Load(volatile const Atomic32* ptr) { inline Atomic32 NoBarrier_Load(volatile const Atomic32* ptr) {
...@@ -194,8 +167,15 @@ inline Atomic32 NoBarrier_Load(volatile const Atomic32* ptr) { ...@@ -194,8 +167,15 @@ inline Atomic32 NoBarrier_Load(volatile const Atomic32* ptr) {
} }
inline Atomic32 Acquire_Load(volatile const Atomic32* ptr) { inline Atomic32 Acquire_Load(volatile const Atomic32* ptr) {
Atomic32 value = *ptr; Atomic32 value;
MemoryBarrier();
__asm__ __volatile__ ( // NOLINT
"ldar %w[value], %[ptr] \n\t"
: [value]"=r" (value)
: [ptr]"Q" (*ptr)
: "memory"
); // NOLINT
return value; return value;
} }
...@@ -221,13 +201,12 @@ inline Atomic64 NoBarrier_CompareAndSwap(volatile Atomic64* ptr, ...@@ -221,13 +201,12 @@ inline Atomic64 NoBarrier_CompareAndSwap(volatile Atomic64* ptr,
"stxr %w[temp], %[new_value], %[ptr] \n\t" "stxr %w[temp], %[new_value], %[ptr] \n\t"
"cbnz %w[temp], 0b \n\t" "cbnz %w[temp], 0b \n\t"
"1: \n\t" "1: \n\t"
"clrex \n\t"
: [prev]"=&r" (prev), : [prev]"=&r" (prev),
[temp]"=&r" (temp), [temp]"=&r" (temp),
[ptr]"+Q" (*ptr) [ptr]"+Q" (*ptr)
: [old_value]"r" (old_value), : [old_value]"IJr" (old_value),
[new_value]"r" (new_value) [new_value]"r" (new_value)
: "memory", "cc" : "cc", "memory"
); // NOLINT ); // NOLINT
return prev; return prev;
...@@ -267,7 +246,7 @@ inline Atomic64 NoBarrier_AtomicIncrement(volatile Atomic64* ptr, ...@@ -267,7 +246,7 @@ inline Atomic64 NoBarrier_AtomicIncrement(volatile Atomic64* ptr,
: [result]"=&r" (result), : [result]"=&r" (result),
[temp]"=&r" (temp), [temp]"=&r" (temp),
[ptr]"+Q" (*ptr) [ptr]"+Q" (*ptr)
: [increment]"r" (increment) : [increment]"IJr" (increment)
: "memory" : "memory"
); // NOLINT ); // NOLINT
...@@ -276,8 +255,10 @@ inline Atomic64 NoBarrier_AtomicIncrement(volatile Atomic64* ptr, ...@@ -276,8 +255,10 @@ inline Atomic64 NoBarrier_AtomicIncrement(volatile Atomic64* ptr,
inline Atomic64 Barrier_AtomicIncrement(volatile Atomic64* ptr, inline Atomic64 Barrier_AtomicIncrement(volatile Atomic64* ptr,
Atomic64 increment) { Atomic64 increment) {
Atomic64 result;
MemoryBarrier(); MemoryBarrier();
Atomic64 result = NoBarrier_AtomicIncrement(ptr, increment); result = NoBarrier_AtomicIncrement(ptr, increment);
MemoryBarrier(); MemoryBarrier();
return result; return result;
...@@ -287,25 +268,9 @@ inline Atomic64 Acquire_CompareAndSwap(volatile Atomic64* ptr, ...@@ -287,25 +268,9 @@ inline Atomic64 Acquire_CompareAndSwap(volatile Atomic64* ptr,
Atomic64 old_value, Atomic64 old_value,
Atomic64 new_value) { Atomic64 new_value) {
Atomic64 prev; Atomic64 prev;
int32_t temp;
__asm__ __volatile__ ( // NOLINT prev = NoBarrier_CompareAndSwap(ptr, old_value, new_value);
"0: \n\t" MemoryBarrier();
"ldxr %[prev], %[ptr] \n\t"
"cmp %[prev], %[old_value] \n\t"
"bne 1f \n\t"
"stxr %w[temp], %[new_value], %[ptr] \n\t"
"cbnz %w[temp], 0b \n\t"
"dmb ish \n\t"
"1: \n\t"
"clrex \n\t"
: [prev]"=&r" (prev),
[temp]"=&r" (temp),
[ptr]"+Q" (*ptr)
: [old_value]"r" (old_value),
[new_value]"r" (new_value)
: "memory", "cc"
); // NOLINT
return prev; return prev;
} }
...@@ -314,26 +279,9 @@ inline Atomic64 Release_CompareAndSwap(volatile Atomic64* ptr, ...@@ -314,26 +279,9 @@ inline Atomic64 Release_CompareAndSwap(volatile Atomic64* ptr,
Atomic64 old_value, Atomic64 old_value,
Atomic64 new_value) { Atomic64 new_value) {
Atomic64 prev; Atomic64 prev;
int32_t temp;
MemoryBarrier(); MemoryBarrier();
prev = NoBarrier_CompareAndSwap(ptr, old_value, new_value);
__asm__ __volatile__ ( // NOLINT
"0: \n\t"
"ldxr %[prev], %[ptr] \n\t"
"cmp %[prev], %[old_value] \n\t"
"bne 1f \n\t"
"stxr %w[temp], %[new_value], %[ptr] \n\t"
"cbnz %w[temp], 0b \n\t"
"1: \n\t"
"clrex \n\t"
: [prev]"=&r" (prev),
[temp]"=&r" (temp),
[ptr]"+Q" (*ptr)
: [old_value]"r" (old_value),
[new_value]"r" (new_value)
: "memory", "cc"
); // NOLINT
return prev; return prev;
} }
...@@ -348,8 +296,12 @@ inline void Acquire_Store(volatile Atomic64* ptr, Atomic64 value) { ...@@ -348,8 +296,12 @@ inline void Acquire_Store(volatile Atomic64* ptr, Atomic64 value) {
} }
inline void Release_Store(volatile Atomic64* ptr, Atomic64 value) { inline void Release_Store(volatile Atomic64* ptr, Atomic64 value) {
MemoryBarrier(); __asm__ __volatile__ ( // NOLINT
*ptr = value; "stlr %x[value], %[ptr] \n\t"
: [ptr]"=Q" (*ptr)
: [value]"r" (value)
: "memory"
); // NOLINT
} }
inline Atomic64 NoBarrier_Load(volatile const Atomic64* ptr) { inline Atomic64 NoBarrier_Load(volatile const Atomic64* ptr) {
...@@ -357,8 +309,15 @@ inline Atomic64 NoBarrier_Load(volatile const Atomic64* ptr) { ...@@ -357,8 +309,15 @@ inline Atomic64 NoBarrier_Load(volatile const Atomic64* ptr) {
} }
inline Atomic64 Acquire_Load(volatile const Atomic64* ptr) { inline Atomic64 Acquire_Load(volatile const Atomic64* ptr) {
Atomic64 value = *ptr; Atomic32 value;
MemoryBarrier();
__asm__ __volatile__ ( // NOLINT
"ldar %x[value], %[ptr] \n\t"
: [value]"=r" (value)
: [ptr]"Q" (*ptr)
: "memory"
); // NOLINT
return value; return value;
} }
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment