2011年9月24日土曜日

Clojure 1.3 のパフォーマンスに驚いた話

Clojureの今によると、 Clojure 1.3 からプリミティブ型による型注釈が可能になったようです。プリミティブ 型による型注釈は、最適化において最も重要な情報の一つですから、当然なが らその結果として、プリミティブな演算を伴う処理の高速化が期待できるわけ です。スライドでは、以下の fib 関数を例にとって、型注釈を与えた場合と 与えなかった場合とで、どれだけ実行速度が違うかを示しています。

;; 型注釈なし
(defn fib [n]
  (if (<= n 1)
      1
      (+ (fib (dec n)) (fib (- n 2)))))
(time (fib 38)) ;=> Elapsed time: 4100.272 msecs

;; 型注釈あり
(defn fib ^long [^long n]
  (if (<= n 1)
      1
      (+ (fib (dec n)) (fib (- n 2)))))
(time (fib 38)) ;=> Elapsed time: 674.706 msecs

この手の情報を安易に鵜呑みにするのはいけないので、実際に手元の環境で確 認してみました。

$ uname -a
Linux thinkpad 2.6.38-11-generic #50-Ubuntu SMP Mon Sep 12 21:17:25 UTC 2011 x86_64 x86_64 x86_64 GNU/Linux
$ java -version
java version "1.6.0_22"
OpenJDK Runtime Environment (IcedTea6 1.10.2) (6b22-1.10.2-0ubuntu1~11.04.1)
OpenJDK 64-Bit Server VM (build 20.0-b11, mixed mode)
$ java -jar ~/opt/clojure-1.3.0-RC0/clojure-1.3.0-RC0.jar
Clojure 1.3.0-RC0
user=> (defn fib [n] (if (<= n 1) 1 (+ (fib (dec n)) (fib (- n 2)))))
#'user/fib
user=> (dotimes [_ 3] (time (fib 38)))
"Elapsed time: 3927.97485 msecs"
"Elapsed time: 3722.368325 msecs"
"Elapsed time: 3828.55781 msecs"
nil
user=> ^D
$ java -jar ~/opt/clojure-1.3.0-RC0/clojure-1.3.0-RC0.jar
Clojure 1.3.0-RC0
user=> (defn fib ^long [^long n] (if (<= n 1) 1 (+ (fib (dec n)) (fib (- n 2)))))
#'user/fib
user=> (dotimes [_ 3] (time (fib 38)))
"Elapsed time: 803.010916 msecs"
"Elapsed time: 793.236553 msecs"
"Elapsed time: 791.748801 msecs"
nil
user=> 

確かに早くなっていますね。ところで我らが SBCL はどれぐらい速いのでしょ うか。

$ sbcl
This is SBCL 1.0.51, an implementation of ANSI Common Lisp.
More information about SBCL is available at <http://www.sbcl.org/>.

SBCL is free software, provided as is, with absolutely no warranty.
It is mostly in the public domain; some portions are provided under
BSD-style licenses.  See the CREDITS and COPYING files in the
distribution for more information.
* 
(defun fib (n)
  (declare (optimize (speed 3) (safety 0))
           (fixnum n))
  (if (<= n 1)
      1
      (the fixnum (+ (fib (1- n)) (fib (- n 2))))))

FIB
* (dotimes '3 (time (fib 38)))

Evaluation took:
  1.211 seconds of real time
  1.210000 seconds of total run time (1.200000 user, 0.010000 system)
  99.92% CPU
  1,932,215,688 processor cycles
  0 bytes consed

Evaluation took:
  1.205 seconds of real time
  1.210000 seconds of total run time (1.210000 user, 0.000000 system)
  100.41% CPU
  1,922,833,256 processor cycles
  0 bytes consed

Evaluation took:
  1.211 seconds of real time
  1.210000 seconds of total run time (1.210000 user, 0.000000 system)
  99.92% CPU
  1,933,261,128 processor cycles
  0 bytes consed

NIL

なぜか Clojure に負ける SBCL 。悔しいので詳しく調べてみます。とりあえず DISASSEMBLE してみましょう。

* (disassemble 'fib)
; disassembly for FIB
; 030680F2:       4883FA08         CMP RDX, 8                 ; no-arg-parsing entry point
;      0F6:       7F0A             JNLE L1
;      0F8:       B908000000       MOV ECX, 8
;      0FD: L0:   488BE5           MOV RSP, RBP
;      100:       5D               POP RBP
;      101:       C3               RET
;      102: L1:   488955F0         MOV [RBP-16], RDX
;      106:       488BCA           MOV RCX, RDX
;      109:       4883E908         SUB RCX, 8
;      10D:       488BDD           MOV RBX, RBP
;      110:       488D4424F0       LEA RAX, [RSP-16]
;      115:       4883EC40         SUB RSP, 64
;      119:       488BD1           MOV RDX, RCX
;      11C:       488918           MOV [RAX], RBX
;      11F:       488BE8           MOV RBP, RAX
;      122:       E8C8FFFFFF       CALL #x10030680EF
;      127:       488B55F0         MOV RDX, [RBP-16]
;      12B:       48894DF8         MOV [RBP-8], RCX
;      12F:       4883EA10         SUB RDX, 16
;      133:       488BCD           MOV RCX, RBP
;      136:       488D4424F0       LEA RAX, [RSP-16]
;      13B:       4883EC40         SUB RSP, 64
;      13F:       488908           MOV [RAX], RCX
;      142:       488BE8           MOV RBP, RAX
;      145:       E8A5FFFFFF       CALL #x10030680EF
;      14A:       488B55F8         MOV RDX, [RBP-8]
;      14E:       4801CA           ADD RDX, RCX
;      151:       488BCA           MOV RCX, RDX
;      154:       EBA7             JMP L0
NIL

至って単純ですし、これ以上に速くなるのが信じられません。参考として C で 書いたものも計測してみましょう。

$ cat fib.c
int fib(int n) {
    if (n <= 1)
        return 1;
    else
        return fib(n - 1) + fib(n - 2);
}

int main(int argc, char *argv[]) {
    fib(38);
    return 0;
}
$ gcc -O2 fib.c
$ time ./a.out
./a.out  1.01s user 0.00s system 98% cpu 1.026 total

Clojure のほうが速いですが、そんなことはありえないから -O3 で再挑戦し ます。

$ gcc -O3 fib.c
$ time ./a.out
./a.out  0.38s user 0.00s system 97% cpu 0.390 total

当然の結果です。長くなるので詳細は省きますが、両者の実行可能ファイルを objdump して比べてみると、 -O3 でコンパイルしたほうは明らかにコード サイズ量が大きくなっており、インライン化された形跡が見られます。たぶん これと同じことが Clojure vs. SBCL にも起きているのでしょう。実際に確か めてみます。

ところで Clojure には disassemble 関数とか、 javap のような decompile tool はないのでしょうか。いくら探しても見つからないので、皆困っ てないのかもしれませんが、 Clojure ほどの抽象化されたプログラミング言語 で、その手の解析ツールがないのは、結構致命的だと思ったりします。少なく とも僕は好きになれません。また、コンパイラに関しても、 Python で言うと ころのバイトコードを、クラスファイルに取り繕うだけのおもちゃみたいなコ ンパイラで、期待外れでがっかりしました。まあこれから改善されるのでしょ う。

以上のような状況ですから、 Clojure のコードを容易には解析できません。た だし、今回に関して言えば、 fib 関数がインライン化されているかどうかが 分かれば良いわけですから、それほど苦労はしませんでした。

まず、予想を確かめるために -XX:-Inline で実験します。ちなみにこのオプ ションはインライン化を無効にするものです。グローバルに無効になるので、 正確な解析にはなりませんが、ちょっとした目安にはなるでしょう。

$ java -XX:-Inline -jar ~/opt/clojure-1.3.0-RC0/clojure-1.3.0-RC0.jar
Clojure 1.3.0-RC0
user=> (defn fib ^long [^long n] (if (<= n 1) 1 (+ (fib (dec n)) (fib (- n 2)))))
#'user/fib
user=> (dotimes [_ 3] (time (fib 38)))
"Elapsed time: 3107.450674 msecs"
"Elapsed time: 3086.921525 msecs"
"Elapsed time: 3073.992303 msecs"
nil

すごく遅くなってますね。やはりインライン化は効いているようです。それで は実際にインライン化が効いているかを確かめましょう。そのためには -XX:+PrintInlining-XX:+PrintAssembly を利用します。ただし、両者 ともデバッグビルドされた HotSpot JDK でしか利用できません。 OpenJDK 6 あるいは OpenJDK 7 のソースコードをダ ウンロードして、以下のようにビルドしてください。ちなみに僕は OpenJDK 7 をインストールしました。新しい物好きです。

$ unzip openjdk-7-*.zip
$ cd openjdk
$ make sanity
$ make debug_build
...
$ build/linux-amd64-debug/bin/java -version
openjdk version "1.7.0-internal-debug"
OpenJDK Runtime Environment (build 1.7.0-internal-debug-tomo_2011_09_24_13_54-b00)
OpenJDK 64-Bit Server VM (build 21.0-b17-jvmg, mixed mode)

まず -XX:-Inline-XX:+PrintAssemblyfib 関数のアセンブリを確 認してみます。

$ ~/opt/openjdk/build/linux-amd64-debug/bin/java -XX:-Inline -XX:+PrintAssembly -jar ~/opt/clojure-1.3.0-RC0/clojure-1.3.0-RC0.jar
user=> (defn fib ^long [^long n] (if (<= n 1) 1 (+ (fib (dec n)) (fib (- n 2)))))
#'user/fib
user=> (fib 38)
;; fib 関数のアセンブリ
{method}
 - klass: {other class}
 - this oop:          0x00000000bebd81e8
 - method holder:     'user$fib'
 - constants:         0x00000000bebd7aa0 constant pool [104] for 'user$fib' cache=0x00000000bebd8838
 - access:            0x81000011  public final
 - name:              'invokePrim'
 - signature:         '(J)J'
 - max stack:         7
 - max locals:        3
 - size of params:    3
 - method size:       17
 - vtable index:      -2
 - i2i entry:         0x00007f87a323e060
 - adapter:           0x0000000002040c90
 - compiled entry     0x00007f87a32eeeb1
 - code size:         54
 - code start:        0x00000000bebd8190
 - code end (excl):   0x00000000bebd81c6
 - method data:       0x00000000bebd95b8
 - checked ex length: 0
 - linenumber start:  0x00000000bebd81c6
 - localvar length:   2
 - localvar start:    0x00000000bebd81ce
#
#  long/half ( user$fib:NotNull:exact *, long, half )
#
#r018 rsi:rsi   : parm 0: user$fib:NotNull:exact *
#r016 rdx:rdx   : parm 1: long
# -- Old rsp -- Framesize: 48 --
#r089 rsp+44: pad2, in_preserve
#r088 rsp+40: pad2, in_preserve
#r087 rsp+36: pad2, in_preserve
#r086 rsp+32: pad2, in_preserve
#r085 rsp+28: pad2, in_preserve
#r084 rsp+24: return address
#r083 rsp+20: Fixed slot 1
#r082 rsp+16: Fixed slot 0
#r093 rsp+12: spill
#r092 rsp+ 8: spill
#r091 rsp+ 4: spill
#r090 rsp+ 0: spill
#
000   N215: #   B1 <- BLOCK HEAD IS JUNK   Freq: 1
000     movl    rscratch1, [j_rarg0 + oopDesc::klass_offset_in_bytes()] # compressed klass
        cmpq    rax, rscratch1   # Inline cache check
        jne     SharedRuntime::_ic_miss_stub
        nop     # nops to align entry point

000
010   B1: #     B12 B2 <- BLOCK HEAD IS JUNK   Freq: 1
010     # stack bang
        pushq   rbp
        subq    rsp, #32        # Create frame
01c     movq    [rsp + #0], RDX # spill
020     cmpq    RDX, #1
024     jle     B12  P=0.500340 C=7351.000000
024
02a   B2: #     B22 B3 <- B1  Freq: 0.49966
02a     movq    R10, java/lang/Class:exact *    # ptr
034     movl    R11, [R10 + #124 (8-bit)]       # compressed ptr ! Field user$fib.const__3
038     movl    RBP, [R11 + #28 (8-bit)]        # compressed ptr ! Field  Volatileclojure/lang/Var.root
03c     MEMBAR-acquire ! (empty encoding)
03c     movl    R10, [RBP + #8 (8-bit)] # compressed klass ptr
040     NullCheck RBP
040
040   B3: #     B14 B4 <- B2  Freq: 0.499659
040     cmpl    R10, narrowoop: precise klass user$fib: 0x00007f879c19ac68:Constant:exact *     # compressed ptr
047     jne,u  B14  P=0.000000 C=-1.000000
047
04d   B4: #     B19 B5 <- B3  Freq: 0.499659
04d     decode_heap_oop_not_null RBP,RBP
04d     # checkcastPP of RBP
04d     movq    RSI, RDX        # spill
050     nop     # 3 bytes pad for loops and calls
053     call,static  clojure.lang.Numbers::dec
        # user$fib::invokePrim @ bci:21  L[0]=_ L[1]=rsp + #0 L[2]=_ STK[0]=RBP
        # OopMap{rbp=Oop off=88}
058
058   B5: #     B20 B6 <- B4  Freq: 0.499649
        # Block is sole successor of call
058     movq    RDX, RAX        # spill
05b     movq    RSI, RBP        # spill
05e     nop     # 1 bytes pad for loops and calls
05f     call,static  user$fib::invokePrim
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=rsp + #0 L[2]=_
        # OopMap{off=100}
064
064   B6: #     B23 B7 <- B5  Freq: 0.499639
        # Block is sole successor of call
064     movq    [rsp + #8], RAX # spill
069     movq    R10, java/lang/Class:exact *    # ptr
073     movl    R11, [R10 + #124 (8-bit)]       # compressed ptr ! Field user$fib.const__3
077     movl    RBP, [R11 + #28 (8-bit)]        # compressed ptr ! Field  Volatileclojure/lang/Var.root
07b     MEMBAR-acquire ! (empty encoding)
07b     movl    R10, [RBP + #8 (8-bit)] # compressed klass ptr
07f     NullCheck RBP
07f
07f   B7: #     B15 B8 <- B6  Freq: 0.499639
07f     cmpl    R10, narrowoop: precise klass user$fib: 0x00007f879c19ac68:Constant:exact *     # compressed ptr
086     jne,us  B15  P=0.000000 C=-1.000000
086
088   B8: #     B18 B9 <- B7  Freq: 0.499639
088     decode_heap_oop_not_null RBP,RBP
088     # checkcastPP of RBP
088     movl    RDX, #2 # long (unsigned 32-bit)
08d     movq    RSI, [rsp + #0] # spill
091     nop     # 2 bytes pad for loops and calls
093     call,static  clojure.lang.Numbers::minus
        # user$fib::invokePrim @ bci:42  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #8 STK[1]=_ STK[2]=RBP
        # OopMap{rbp=Oop off=152}
098
098   B9: #     B17 B10 <- B8  Freq: 0.499629
        # Block is sole successor of call
098     movq    RSI, RBP        # spill
09b     movq    RDX, RAX        # spill
09e     nop     # 1 bytes pad for loops and calls
09f     call,static  user$fib::invokePrim
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #8 STK[1]=_
        # OopMap{off=164}
0a4
0a4   B10: #    B16 B11 <- B9  Freq: 0.499619
        # Block is sole successor of call
0a4     movq    RSI, [rsp + #8] # spill
0a9     movq    RDX, RAX        # spill
0ac     nop     # 3 bytes pad for loops and calls
0af     call,static  clojure.lang.Numbers::add
        # user$fib::invokePrim @ bci:50  L[0]=_ L[1]=_ L[2]=_
        # OopMap{off=180}
0b4
0b4   B11: #    B13 <- B10  Freq: 0.499609
        # Block is sole successor of call
0b4     jmp,s   B13
0b4
0b6   B12: #    B13 <- B1  Freq: 0.50034
0b6     movl    RAX, #1 # long (unsigned 32-bit)
0b6
0bb   B13: #    N215 <- B11 B12  Freq: 0.999949
0bb     addq    rsp, 32 # Destroy frame
        popq   rbp
        testl  rax, [rip + #offset_to_poll_page]        # Safepoint: poll for GC

0c6     ret
0c6
0c7   B14: #    N215 <- B3  Freq: 1e-35
0c7     movl    RSI, #-34       # int
0cc     nop     # 3 bytes pad for loops and calls
0cf     call,static  wrapper for: uncommon_trap(reason='class_check' action='maybe_recompile')
        # user$fib::invokePrim @ bci:17  L[0]=_ L[1]=rsp + #0 L[2]=_ STK[0]=RBP
        # OopMap{rbp=NarrowOop off=212}
0d4     int3    # ShouldNotReachHere
0d4
0d9   B15: #    N215 <- B7  Freq: 1e-35
0d9     movl    RSI, #-34       # int
0de     nop     # 1 bytes pad for loops and calls
0df     call,static  wrapper for: uncommon_trap(reason='class_check' action='maybe_recompile')
        # user$fib::invokePrim @ bci:35  L[0]=_ L[1]=rsp + #0 L[2]=_ STK[0]=rsp + #8 STK[1]=_ STK[2]=RBP
        # OopMap{rbp=NarrowOop off=228}
0e4     int3    # ShouldNotReachHere
0e4
0e9   B16: #    B21 <- B10  Freq: 4.99619e-06
0e9     # exception oop is in rax; no code emitted
0e9     movq    RSI, RAX        # spill
0ec     jmp,s   B21
0ec
0ee   B17: #    B21 <- B9  Freq: 4.99629e-06
0ee     # exception oop is in rax; no code emitted
0ee     movq    RSI, RAX        # spill
0f1     jmp,s   B21
0f1
0f3   B18: #    B21 <- B8  Freq: 4.99639e-06
0f3     # exception oop is in rax; no code emitted
0f3     movq    RSI, RAX        # spill
0f6     jmp,s   B21
0f6
0f8   B19: #    B21 <- B4  Freq: 4.99659e-06
0f8     # exception oop is in rax; no code emitted
0f8     movq    RSI, RAX        # spill
0fb     jmp,s   B21
0fb
0fd   B20: #    B21 <- B5  Freq: 4.99649e-06
0fd     # exception oop is in rax; no code emitted
0fd     movq    RSI, RAX        # spill
0fd
100   B21: #    N215 <- B19 B20 B18 B17 B16  Freq: 2.4982e-05
100     addq    rsp, 32 # Destroy frame
        popq   rbp

105     jmp     rethrow_stub
105
10a   B22: #    N215 <- B2  Freq: 5.06295e-07
10a     movl    RSI, #-12       # int
10f     movq    RBP, RDX        # spill
112     nop     # 1 bytes pad for loops and calls
113     call,static  wrapper for: uncommon_trap(reason='null_check' action='make_not_entrant')
        # user$fib::invokePrim @ bci:17  L[0]=_ L[1]=RBP L[2]=_ STK[0]=#NULL
        # OopMap{off=280}
118     int3    # ShouldNotReachHere
118
11d   B23: #    N215 <- B6  Freq: 5.06274e-07
11d     movl    RSI, #-12       # int
122     movq    RBP, [rsp + #0] # spill
126     nop     # 1 bytes pad for loops and calls
127     call,static  wrapper for: uncommon_trap(reason='null_check' action='make_not_entrant')
        # user$fib::invokePrim @ bci:35  L[0]=_ L[1]=RBP L[2]=_ STK[0]=rsp + #8 STK[1]=_ STK[2]=#NULL
        # OopMap{off=300}
12c     int3    # ShouldNotReachHere
12c

次に -XX:+Inline した場合で fib 関数のアセンブリを確認してみます。

$ ~/opt/openjdk/build/linux-amd64-debug/bin/java -XX:+Inline -XX:+PrintAssembly -jar ~/opt/clojure-1.3.0-RC0/clojure-1.3.0-RC0.jar
user=> (defn fib ^long [^long n] (if (<= n 1) 1 (+ (fib (dec n)) (fib (- n 2)))))
#'user/fib
user=> (fib 38)
;; fib 関数のアセンブリ
{method}
 - klass: {other class}
 - this oop:          0x00000000bebced68
 - method holder:     'user$fib'
 - constants:         0x00000000bebce620 constant pool [104] for 'user$fib' cache=0x00000000bebcf3b8
 - access:            0x81000011  public final
 - name:              'invokePrim'
 - signature:         '(J)J'
 - max stack:         7
 - max locals:        3
 - size of params:    3
 - method size:       17
 - vtable index:      -2
 - i2i entry:         0x00007fd949404060
 - adapter:           0x0000000000ba7c90
 - compiled entry     0x00007fd9494b4eb1
 - code size:         54
 - code start:        0x00000000bebced10
 - code end (excl):   0x00000000bebced46
 - method data:       0x00000000bebd0158
 - checked ex length: 0
 - linenumber start:  0x00000000bebced46
 - localvar length:   2
 - localvar start:    0x00000000bebced4e
#
#  long/half ( user$fib:NotNull:exact *, long, half )
#
#r018 rsi:rsi   : parm 0: user$fib:NotNull:exact *
#r016 rdx:rdx   : parm 1: long
# -- Old rsp -- Framesize: 64 --
#r089 rsp+60: pad2, in_preserve
#r088 rsp+56: pad2, in_preserve
#r087 rsp+52: pad2, in_preserve
#r086 rsp+48: pad2, in_preserve
#r085 rsp+44: pad2, in_preserve
#r084 rsp+40: return address
#r083 rsp+36: Fixed slot 1
#r082 rsp+32: Fixed slot 0
#r097 rsp+28: spill
#r096 rsp+24: spill
#r095 rsp+20: spill
#r094 rsp+16: spill
#r093 rsp+12: spill
#r092 rsp+ 8: spill
#r091 rsp+ 4: spill
#r090 rsp+ 0: spill
#
000   N805: #   B1 <- BLOCK HEAD IS JUNK   Freq: 1
000     movl    rscratch1, [j_rarg0 + oopDesc::klass_offset_in_bytes()] # compressed klass
        cmpq    rax, rscratch1   # Inline cache check
        jne     SharedRuntime::_ic_miss_stub
        nop     # nops to align entry point

000
010   B1: #     B39 B2 <- BLOCK HEAD IS JUNK   Freq: 1
010     # stack bang
        pushq   rbp
        subq    rsp, #48        # Create frame
01c     movq    R8, RDX # spill
01f     cmpq    RDX, #1
023     jle     B39  P=0.500344 C=7273.000000
023
029   B2: #     B68 B3 <- B1  Freq: 0.499656
029     movq    R10, java/lang/Class:exact *    # ptr
033     movl    R11, [R10 + #124 (8-bit)]       # compressed ptr ! Field user$fib.const__3
037     movl    RBP, [R11 + #28 (8-bit)]        # compressed ptr ! Field  Volatileclojure/lang/Var.root
03b     MEMBAR-acquire ! (empty encoding)
03b     movl    R10, [RBP + #8 (8-bit)] # compressed klass ptr
03f     NullCheck RBP
03f
03f   B3: #     B62 B4 <- B2  Freq: 0.499656
03f     cmpl    R10, narrowoop: precise klass user$fib: 0x00007fd940329268:Constant:exact *     # compressed ptr
046     jne,u  B62  P=0.000000 C=-1.000000
046
04c   B4: #     B41 B5 <- B3  Freq: 0.499656
04c     movq    R10, #-9223372036854775808      # long
056     cmpq    RDX, R10
059     je     B41  P=0.000000 C=7240.000000
059
05f   B5: #     B6 <- B4  Freq: 0.499656
05f     movq    R10, RDX        # spill
062     decq    R10     # long
065
065   B6: #     B18 B7 <- B42 B5  Freq: 0.499656
065     movq    R11, java/lang/Class:exact *    # ptr
06f     movl    R11, [R11 + #124 (8-bit)]       # compressed ptr ! Field user$fib.const__3
073     cmpq    R10, #1
077     jle     B18  P=0.500344 C=7273.000000
077
07d   B7: #     B70 B8 <- B6  Freq: 0.249656
07d     movl    RBP, [R11 + #28 (8-bit)]        # compressed ptr ! Field  Volatileclojure/lang/Var.root
081     MEMBAR-acquire ! (empty encoding)
081     movl    R11, [RBP + #8 (8-bit)] # compressed klass ptr
085     NullCheck RBP
085
085   B8: #     B63 B9 <- B7  Freq: 0.249656
085     cmpl    R11, narrowoop: precise klass user$fib: 0x00007fd940329268:Constant:exact *     # compressed ptr
08c     jne,u  B63  P=0.000000 C=-1.000000
08c
092   B9: #     B46 B10 <- B8  Freq: 0.249656
092     decode_heap_oop_not_null RSI,RBP
095     # checkcastPP of RSI
095     movq    R11, #-9223372036854775808      # long
09f     cmpq    R10, R11
0a2     je     B46  P=0.000000 C=7240.000000
0a2
0a8   B10: #    B11 <- B9  Freq: 0.249656
0a8     movq    RDX, R10        # spill
0ab     decq    RDX     # long
0ae
0ae   B11: #    B83 B12 <- B47 B10  Freq: 0.249656
0ae     movq    [rsp + #8], R10 # spill
0b3     movq    [rsp + #0], R8  # spill
0b7     call,static  user$fib::invokePrim
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=rsp + #8 L[2]=_
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=rsp + #0 L[2]=_
        # OopMap{off=188}
0bc
0bc   B12: #    B71 B13 <- B11  Freq: 0.249651
        # Block is sole successor of call
0bc     movq    [rsp + #16], RAX        # spill
0c1     movq    R10, java/lang/Class:exact *    # ptr
0cb     movl    R11, [R10 + #124 (8-bit)]       # compressed ptr ! Field user$fib.const__3
0cf     movl    RBP, [R11 + #28 (8-bit)]        # compressed ptr ! Field  Volatileclojure/lang/Var.root
0d3     MEMBAR-acquire ! (empty encoding)
0d3     movl    R10, [RBP + #8 (8-bit)] # compressed klass ptr
0d7     NullCheck RBP
0d7
0d7   B13: #    B64 B14 <- B12  Freq: 0.249651
0d7     cmpl    R10, narrowoop: precise klass user$fib: 0x00007fd940329268:Constant:exact *     # compressed ptr
0de     jne,u  B64  P=0.000000 C=-1.000000
0de
0e4   B14: #    B48 B15 <- B13  Freq: 0.249651
0e4     decode_heap_oop_not_null RSI,RBP
0e7     # checkcastPP of RSI
0e7     movq    RDX, [rsp + #8] # spill
0ec     addq    RDX, #-2        # long
0f0     movq    R10, [rsp + #8] # spill
0f5     xorq    R10, RDX        # long
0f8     testq   R10, R10
0fb     jl     B48  P=0.000000 C=8268.000000
0fb
101   B15: #    B84 B16 <- B50 B48 B14  Freq: 0.249651
101     nop     # 2 bytes pad for loops and calls
103     call,static  user$fib::invokePrim
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #16 STK[1]=_
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=rsp + #0 L[2]=_
        # OopMap{off=264}
108
108   B16: #    B53 B17 <- B15  Freq: 0.249646
        # Block is sole successor of call
108     movq    R9, [rsp + #16] # spill
10d     addq    R9, RAX # long
110     movq    R11, [rsp + #16]        # spill
115     xorq    R11, R9 # long
118     testq   R11, R11
11b     jl     B53  P=0.000000 C=8899.000000
11b
121   B17: #    B19 <- B55 B53 B16  Freq: 0.249646
121     movq    R10, java/lang/Class:exact *    # ptr
12b     movl    R11, [R10 + #124 (8-bit)]       # compressed ptr ! Field user$fib.const__3
12f     movq    R8, [rsp + #0]  # spill
133     jmp,s   B19
133
135   B18: #    B19 <- B6  Freq: 0.25
135     movl    R9, #1  # long (unsigned 32-bit)
135
13b   B19: #    B69 B20 <- B17 B18  Freq: 0.499645
13b     movl    R10, [R11 + #28 (8-bit)]        # compressed ptr ! Field  Volatileclojure/lang/Var.root
13f     MEMBAR-acquire ! (empty encoding)
13f     movl    RCX, [R10 + #8 (8-bit)] # compressed klass ptr
143     NullCheck R10
143
143   B20: #    B65 B21 <- B19  Freq: 0.499645
143     cmpl    RCX, narrowoop: precise klass user$fib: 0x00007fd940329268:Constant:exact *     # compressed ptr
149     jne,u  B65  P=0.000000 C=-1.000000
149
14f   B21: #    B43 B22 <- B20  Freq: 0.499645
14f     movq    RCX, R8 # spill
152     addq    RCX, #-2        # long
156     xorq    R8, RCX # long
159     testq   R8, R8
15c     jl     B43  P=0.000000 C=8268.000000
15c
162   B22: #    B34 B23 <- B45 B43 B21  Freq: 0.499645
162     cmpq    RCX, #1
166     jle     B34  P=0.500344 C=7273.000000
166
16c   B23: #    B72 B24 <- B22  Freq: 0.249651
16c     movq    R10, java/lang/Class:exact *    # ptr
176     movl    R11, [R10 + #124 (8-bit)]       # compressed ptr ! Field user$fib.const__3
17a     movl    RBP, [R11 + #28 (8-bit)]        # compressed ptr ! Field  Volatileclojure/lang/Var.root
17e     MEMBAR-acquire ! (empty encoding)
17e     movl    R10, [RBP + #8 (8-bit)] # compressed klass ptr
182     NullCheck RBP
182
182   B24: #    B66 B25 <- B23  Freq: 0.24965
182     cmpl    R10, narrowoop: precise klass user$fib: 0x00007fd940329268:Constant:exact *     # compressed ptr
189     jne,u  B66  P=0.000000 C=-1.000000
189
18f   B25: #    B51 B26 <- B24  Freq: 0.24965
18f     decode_heap_oop_not_null RSI,RBP
192     # checkcastPP of RSI
192     movq    R10, #-9223372036854775808      # long
19c     cmpq    RCX, R10
19f     je     B51  P=0.000000 C=7240.000000
19f
1a5   B26: #    B27 <- B25  Freq: 0.24965
1a5     movq    RDX, RCX        # spill
1a8     decq    RDX     # long
1ab
1ab   B27: #    B85 B28 <- B52 B26  Freq: 0.24965
1ab     movq    [rsp + #8], RCX # spill
1b0     movq    [rsp + #0], R9  # spill
1b4     nop     # 3 bytes pad for loops and calls
1b7     call,static  user$fib::invokePrim
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=rsp + #8 L[2]=_
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #0 STK[1]=_
        # OopMap{off=444}
1bc
1bc   B28: #    B73 B29 <- B27  Freq: 0.249645
        # Block is sole successor of call
1bc     movq    [rsp + #16], RAX        # spill
1c1     movq    R10, java/lang/Class:exact *    # ptr
1cb     movl    R10, [R10 + #124 (8-bit)]       # compressed ptr ! Field user$fib.const__3
1cf     movl    RBP, [R10 + #28 (8-bit)]        # compressed ptr ! Field  Volatileclojure/lang/Var.root
1d3     MEMBAR-acquire ! (empty encoding)
1d3     movl    R10, [RBP + #8 (8-bit)] # compressed klass ptr
1d7     NullCheck RBP
1d7
1d7   B29: #    B67 B30 <- B28  Freq: 0.249645
1d7     cmpl    R10, narrowoop: precise klass user$fib: 0x00007fd940329268:Constant:exact *     # compressed ptr
1de     jne,u  B67  P=0.000000 C=-1.000000
1de
1e4   B30: #    B56 B31 <- B29  Freq: 0.249645
1e4     decode_heap_oop_not_null RBP,RBP
1e4     # checkcastPP of RBP
1e4     movq    RDX, [rsp + #8] # spill
1e9     addq    RDX, #-2        # long
1ed     movq    R10, [rsp + #8] # spill
1f2     xorq    R10, RDX        # long
1f5     testq   R10, R10
1f8     jl     B56  P=0.000000 C=8268.000000
1f8
1fe   B31: #    B86 B32 <- B58 B56 B30  Freq: 0.249645
1fe     movq    RSI, RBP        # spill
201     nop     # 2 bytes pad for loops and calls
203     call,static  user$fib::invokePrim
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #16 STK[1]=_
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #0 STK[1]=_
        # OopMap{off=520}
208
208   B32: #    B59 B33 <- B31  Freq: 0.24964
        # Block is sole successor of call
208     movq    R11, [rsp + #16]        # spill
20d     addq    R11, RAX        # long
210     movq    R10, [rsp + #16]        # spill
215     xorq    R10, R11        # long
218     testq   R10, R10
21b     jl     B59  P=0.000000 C=8899.000000
21b
221   B33: #    B35 <- B61 B59 B32  Freq: 0.24964
221     movq    R9, [rsp + #0]  # spill
225     jmp,s   B35
225
227   B34: #    B35 <- B22  Freq: 0.249994
227     movl    R11, #1 # long (unsigned 32-bit)
227
22d   B35: #    B40 B36 <- B33 B34  Freq: 0.499634
22d     movq    RAX, R9 # spill
230     addq    RAX, R11        # long
233     xorq    R9, RAX # long
236     testq   R9, R9
239     jge,s   B40  P=1.000000 C=8899.000000
239
23b   B36: #    B40 B37 <- B35  Freq: 2.38244e-07
23b     xorq    R11, RAX        # long
23e     testq   R11, R11
241     jge,s   B40  P=0.500000 C=-1.000000
241
243   B37: #    B80 B38 <- B36  Freq: 1.19122e-07
243     call,static  clojure.lang.Numbers::throwIntOverflow
        # clojure.lang.Numbers::add @ bci:23  L[0]=_ L[1]=_ L[2]=_ L[3]=_ L[4]=_ L[5]=_
        # user$fib::invokePrim @ bci:50  L[0]=_ L[1]=_ L[2]=_
        # OopMap{off=584}
248
248   B38: #    B40 <- B37  Freq: 1.1912e-07
        # Block is sole successor of call
248     movslq  RAX, RAX        # i2l
24b     jmp,s   B40
24b
24d   B39: #    B40 <- B1  Freq: 0.500344
24d     movl    RAX, #1 # long (unsigned 32-bit)
24d
252   B40: #    N805 <- B38 B36 B35 B39  Freq: 0.999978
252     addq    rsp, 48 # Destroy frame
        popq   rbp
        testl  rax, [rip + #offset_to_poll_page]        # Safepoint: poll for GC

25d     ret
25d
25e   B41: #    B82 B42 <- B4  Freq: 2.38254e-07
25e     movq    [rsp + #0], RDX # spill
262     decode_heap_oop_not_null RBP,RBP
262     # checkcastPP of RBP
262     nop     # 1 bytes pad for loops and calls
263     call,static  clojure.lang.Numbers::throwIntOverflow
        # clojure.lang.Numbers::dec @ bci:8  L[0]=_ L[1]=_
        # user$fib::invokePrim @ bci:21  L[0]=_ L[1]=rsp + #0 L[2]=_ STK[0]=RBP
        # OopMap{rbp=Oop off=616}
268
268   B42: #    B6 <- B41  Freq: 2.3825e-07
        # Block is sole successor of call
268     movslq  R10, RAX        # i2l
26b     movq    R8, [rsp + #0]  # spill
26f     jmp     B6
26f
274   B43: #    B22 B44 <- B21  Freq: 2.38249e-07
274     movq    R11, RCX        # spill
277     xorq    R11, #-3        # long
27b     testq   R11, R11
27e     jge     B22  P=0.500000 C=-1.000000
27e
284   B44: #    B81 B45 <- B43  Freq: 1.19125e-07
284     movq    [rsp + #0], R9  # spill
288     decode_heap_oop_not_null RBP,R10
28b     # checkcastPP of RBP
28b     call,static  clojure.lang.Numbers::throwIntOverflow
        # clojure.lang.Numbers::minus @ bci:27  L[0]=_ L[1]=_ L[2]=_ L[3]=_ L[4]=_ L[5]=_
        # user$fib::invokePrim @ bci:42  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #0 STK[1]=_ STK[2]=RBP
        # OopMap{rbp=Oop off=656}
290
290   B45: #    B22 <- B44  Freq: 1.19122e-07
        # Block is sole successor of call
290     movslq  RCX, RAX        # i2l
293     movq    R9, [rsp + #0]  # spill
297     jmp     B22
297
29c   B46: #    B79 B47 <- B9  Freq: 1.19045e-07
29c     movq    [rsp + #8], RSI # spill
2a1     movq    [rsp + #0], R10 # spill
2a5     movq    RBP, R8 # spill
2a8     nop     # 3 bytes pad for loops and calls
2ab     call,static  clojure.lang.Numbers::throwIntOverflow
        # clojure.lang.Numbers::dec @ bci:8  L[0]=_ L[1]=_
        # user$fib::invokePrim @ bci:21  L[0]=_ L[1]=rsp + #0 L[2]=_ STK[0]=rsp + #8
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=RBP L[2]=_
        # OopMap{[8]=Oop off=688}
2b0
2b0   B47: #    B11 <- B46  Freq: 1.19043e-07
        # Block is sole successor of call
2b0     movslq  RDX, RAX        # i2l
2b3     movq    R8, RBP # spill
2b6     movq    R10, [rsp + #0] # spill
2ba     movq    RSI, [rsp + #8] # spill
2bf     jmp     B11
2bf
2c4   B48: #    B15 B49 <- B14  Freq: 1.19043e-07
2c4     movq    R10, RDX        # spill
2c7     xorq    R10, #-3        # long
2cb     testq   R10, R10
2ce     jge     B15  P=0.500000 C=-1.000000
2ce
2d4   B49: #    B77 B50 <- B48  Freq: 5.95213e-08
2d4     movq    RBP, RSI        # spill
2d7     call,static  clojure.lang.Numbers::throwIntOverflow
        # clojure.lang.Numbers::minus @ bci:27  L[0]=_ L[1]=_ L[2]=_ L[3]=_ L[4]=_ L[5]=_
        # user$fib::invokePrim @ bci:42  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #16 STK[1]=_ STK[2]=RBP
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=rsp + #0 L[2]=_
        # OopMap{rbp=Oop off=732}
2dc
2dc   B50: #    B15 <- B49  Freq: 5.95202e-08
        # Block is sole successor of call
2dc     movslq  RDX, RAX        # i2l
2df     movq    RSI, RBP        # spill
2e2     jmp     B15
2e2
2e7   B51: #    B78 B52 <- B25  Freq: 1.19043e-07
2e7     movq    [rsp + #8], RSI # spill
2ec     movq    [rsp + #0], RCX # spill
2f0     movq    RBP, R9 # spill
2f3     call,static  clojure.lang.Numbers::throwIntOverflow
        # clojure.lang.Numbers::dec @ bci:8  L[0]=_ L[1]=_
        # user$fib::invokePrim @ bci:21  L[0]=_ L[1]=rsp + #0 L[2]=_ STK[0]=rsp + #8
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=RBP STK[1]=_
        # OopMap{[8]=Oop off=760}
2f8
2f8   B52: #    B27 <- B51  Freq: 1.1904e-07
        # Block is sole successor of call
2f8     movslq  RDX, RAX        # i2l
2fb     movq    R9, RBP # spill
2fe     movq    RCX, [rsp + #0] # spill
302     movq    RSI, [rsp + #8] # spill
307     jmp     B27
307
30c   B53: #    B17 B54 <- B16  Freq: 1.1904e-07
30c     xorq    RAX, R9 # long
30f     testq   RAX, RAX
312     jge     B17  P=0.500000 C=-1.000000
312
318   B54: #    B76 B55 <- B53  Freq: 5.95202e-08
318     nop     # 3 bytes pad for loops and calls
31b     call,static  clojure.lang.Numbers::throwIntOverflow
        # clojure.lang.Numbers::add @ bci:23  L[0]=_ L[1]=_ L[2]=_ L[3]=_ L[4]=_ L[5]=_
        # user$fib::invokePrim @ bci:50  L[0]=_ L[1]=_ L[2]=_
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=rsp + #0 L[2]=_
        # OopMap{off=800}
320
320   B55: #    B17 <- B54  Freq: 5.9519e-08
        # Block is sole successor of call
320     movslq  R9, RAX # i2l
323     jmp     B17
323
328   B56: #    B31 B57 <- B30  Freq: 1.1904e-07
328     movq    R10, RDX        # spill
32b     xorq    R10, #-3        # long
32f     testq   R10, R10
332     jge     B31  P=0.500000 C=-1.000000
332
338   B57: #    B75 B58 <- B56  Freq: 5.952e-08
338     nop     # 3 bytes pad for loops and calls
33b     call,static  clojure.lang.Numbers::throwIntOverflow
        # clojure.lang.Numbers::minus @ bci:27  L[0]=_ L[1]=_ L[2]=_ L[3]=_ L[4]=_ L[5]=_
        # user$fib::invokePrim @ bci:42  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #16 STK[1]=_ STK[2]=RBP
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #0 STK[1]=_
        # OopMap{rbp=Oop off=832}
340
340   B58: #    B31 <- B57  Freq: 5.95188e-08
        # Block is sole successor of call
340     movslq  RDX, RAX        # i2l
343     jmp     B31
343
348   B59: #    B33 B60 <- B32  Freq: 1.19038e-07
348     movq    R8, RAX # spill
34b     xorq    R8, R11 # long
34e     testq   R8, R8
351     jge     B33  P=0.500000 C=-1.000000
351
357   B60: #    B74 B61 <- B59  Freq: 5.95188e-08
357     call,static  clojure.lang.Numbers::throwIntOverflow
        # clojure.lang.Numbers::add @ bci:23  L[0]=_ L[1]=_ L[2]=_ L[3]=_ L[4]=_ L[5]=_
        # user$fib::invokePrim @ bci:50  L[0]=_ L[1]=_ L[2]=_
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #0 STK[1]=_
        # OopMap{off=860}
35c
35c   B61: #    B33 <- B60  Freq: 5.95176e-08
        # Block is sole successor of call
35c     movslq  R11, RAX        # i2l
35f     jmp     B33
35f
364   B62: #    N805 <- B3  Freq: 1e-35
364     movl    RSI, #-34       # int
369     movq    [rsp + #0], RDX # spill
36d     nop     # 2 bytes pad for loops and calls
36f     call,static  wrapper for: uncommon_trap(reason='class_check' action='maybe_recompile')
        # user$fib::invokePrim @ bci:17  L[0]=_ L[1]=rsp + #0 L[2]=_ STK[0]=RBP
        # OopMap{rbp=NarrowOop off=884}
374     int3    # ShouldNotReachHere
374
379   B63: #    N805 <- B8  Freq: 1e-35
379     movl    RSI, #-34       # int
37e     movq    [rsp + #0], R8  # spill
382     movq    [rsp + #8], R10 # spill
387     call,static  wrapper for: uncommon_trap(reason='class_check' action='maybe_recompile')
        # user$fib::invokePrim @ bci:17  L[0]=_ L[1]=rsp + #8 L[2]=_ STK[0]=RBP
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=rsp + #0 L[2]=_
        # OopMap{rbp=NarrowOop off=908}
38c     int3    # ShouldNotReachHere
38c
391   B64: #    N805 <- B13  Freq: 1e-35
391     movl    RSI, #-34       # int
396     nop     # 1 bytes pad for loops and calls
397     call,static  wrapper for: uncommon_trap(reason='class_check' action='maybe_recompile')
        # user$fib::invokePrim @ bci:35  L[0]=_ L[1]=rsp + #8 L[2]=_ STK[0]=rsp + #16 STK[1]=_ STK[2]=RBP
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=rsp + #0 L[2]=_
        # OopMap{rbp=NarrowOop off=924}
39c     int3    # ShouldNotReachHere
39c
3a1   B65: #    N805 <- B20  Freq: 1e-35
3a1     movl    RSI, #-34       # int
3a6     movq    RBP, R8 # spill
3a9     movq    [rsp + #0], R9  # spill
3ad     movl    [rsp + #8], R10 # spill
3b2     nop     # 1 bytes pad for loops and calls
3b3     call,static  wrapper for: uncommon_trap(reason='class_check' action='maybe_recompile')
        # user$fib::invokePrim @ bci:35  L[0]=_ L[1]=RBP L[2]=_ STK[0]=rsp + #0 STK[1]=_ STK[2]=rsp + #8
        # OopMap{[8]=NarrowOop off=952}
3b8     int3    # ShouldNotReachHere
3b8
3bd   B66: #    N805 <- B24  Freq: 1e-35
3bd     movl    RSI, #-34       # int
3c2     movq    [rsp + #0], R9  # spill
3c6     movq    [rsp + #8], RCX # spill
3cb     call,static  wrapper for: uncommon_trap(reason='class_check' action='maybe_recompile')
        # user$fib::invokePrim @ bci:17  L[0]=_ L[1]=rsp + #8 L[2]=_ STK[0]=RBP
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #0 STK[1]=_
        # OopMap{rbp=NarrowOop off=976}
3d0     int3    # ShouldNotReachHere
3d0
3d5   B67: #    N805 <- B29  Freq: 1e-35
3d5     movl    RSI, #-34       # int
3da     nop     # 1 bytes pad for loops and calls
3db     call,static  wrapper for: uncommon_trap(reason='class_check' action='maybe_recompile')
        # user$fib::invokePrim @ bci:35  L[0]=_ L[1]=rsp + #8 L[2]=_ STK[0]=rsp + #16 STK[1]=_ STK[2]=RBP
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #0 STK[1]=_
        # OopMap{rbp=NarrowOop off=992}
3e0     int3    # ShouldNotReachHere
3e0
3e5   B68: #    N805 <- B2  Freq: 5.06291e-07
3e5     movl    RSI, #-12       # int
3ea     movq    RBP, RDX        # spill
3ed     nop     # 2 bytes pad for loops and calls
3ef     call,static  wrapper for: uncommon_trap(reason='null_check' action='make_not_entrant')
        # user$fib::invokePrim @ bci:17  L[0]=_ L[1]=RBP L[2]=_ STK[0]=#NULL
        # OopMap{off=1012}
3f4     int3    # ShouldNotReachHere
3f4
3f9   B69: #    N805 <- B19  Freq: 5.0628e-07
3f9     movl    RSI, #-12       # int
3fe     movq    RBP, R8 # spill
401     movq    [rsp + #0], R9  # spill
405     nop     # 2 bytes pad for loops and calls
407     call,static  wrapper for: uncommon_trap(reason='null_check' action='make_not_entrant')
        # user$fib::invokePrim @ bci:35  L[0]=_ L[1]=RBP L[2]=_ STK[0]=rsp + #0 STK[1]=_ STK[2]=#NULL
        # OopMap{off=1036}
40c     int3    # ShouldNotReachHere
40c
411   B70: #    N805 <- B7  Freq: 2.52971e-07
411     movl    RSI, #-12       # int
416     movq    RBP, R8 # spill
419     movq    [rsp + #0], R10 # spill
41d     nop     # 2 bytes pad for loops and calls
41f     call,static  wrapper for: uncommon_trap(reason='null_check' action='make_not_entrant')
        # user$fib::invokePrim @ bci:17  L[0]=_ L[1]=rsp + #0 L[2]=_ STK[0]=#NULL
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=RBP L[2]=_
        # OopMap{off=1060}
424     int3    # ShouldNotReachHere
424
429   B71: #    N805 <- B12  Freq: 2.52966e-07
429     movl    RSI, #-12       # int
42e     movq    RBP, [rsp + #0] # spill
432     nop     # 1 bytes pad for loops and calls
433     call,static  wrapper for: uncommon_trap(reason='null_check' action='make_not_entrant')
        # user$fib::invokePrim @ bci:35  L[0]=_ L[1]=rsp + #8 L[2]=_ STK[0]=rsp + #16 STK[1]=_ STK[2]=#NULL
        # user$fib::invokePrim @ bci:24  L[0]=_ L[1]=RBP L[2]=_
        # OopMap{off=1080}
438     int3    # ShouldNotReachHere
438
43d   B72: #    N805 <- B23  Freq: 2.52966e-07
43d     movl    RSI, #-12       # int
442     movq    RBP, R9 # spill
445     movq    [rsp + #0], RCX # spill
449     nop     # 2 bytes pad for loops and calls
44b     call,static  wrapper for: uncommon_trap(reason='null_check' action='make_not_entrant')
        # user$fib::invokePrim @ bci:17  L[0]=_ L[1]=rsp + #0 L[2]=_ STK[0]=#NULL
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=RBP STK[1]=_
        # OopMap{off=1104}
450     int3    # ShouldNotReachHere
450
455   B73: #    N805 <- B28  Freq: 2.5296e-07
455     movl    RSI, #-12       # int
45a     nop     # 1 bytes pad for loops and calls
45b     call,static  wrapper for: uncommon_trap(reason='null_check' action='make_not_entrant')
        # user$fib::invokePrim @ bci:35  L[0]=_ L[1]=rsp + #8 L[2]=_ STK[0]=rsp + #16 STK[1]=_ STK[2]=#NULL
        # user$fib::invokePrim @ bci:45  L[0]=_ L[1]=_ L[2]=_ STK[0]=rsp + #0 STK[1]=_
        # OopMap{off=1120}
460     int3    # ShouldNotReachHere
460
465   B74: #    B87 <- B60  Freq: 5.95188e-13
465     # exception oop is in rax; no code emitted
465     movq    RSI, RAX        # spill
468     jmp,s   B87
468
46a   B75: #    B87 <- B57  Freq: 5.952e-13
46a     # exception oop is in rax; no code emitted
46a     movq    RSI, RAX        # spill
46d     jmp,s   B87
46d
46f   B76: #    B87 <- B54  Freq: 5.95202e-13
46f     # exception oop is in rax; no code emitted
46f     movq    RSI, RAX        # spill
472     jmp,s   B87
472
474   B77: #    B87 <- B49  Freq: 5.95213e-13
474     # exception oop is in rax; no code emitted
474     movq    RSI, RAX        # spill
477     jmp,s   B87
477
479   B78: #    B87 <- B51  Freq: 1.19043e-12
479     # exception oop is in rax; no code emitted
479     movq    RSI, RAX        # spill
47c     jmp,s   B87
47c
47e   B79: #    B87 <- B46  Freq: 1.19045e-12
47e     # exception oop is in rax; no code emitted
47e     movq    RSI, RAX        # spill
481     jmp,s   B87
481
483   B80: #    B87 <- B37  Freq: 1.19122e-12
483     # exception oop is in rax; no code emitted
483     movq    RSI, RAX        # spill
486     jmp,s   B87
486
488   B81: #    B87 <- B44  Freq: 1.19125e-12
488     # exception oop is in rax; no code emitted
488     movq    RSI, RAX        # spill
48b     jmp,s   B87
48b
48d   B82: #    B87 <- B41  Freq: 2.38254e-12
48d     # exception oop is in rax; no code emitted
48d     movq    RSI, RAX        # spill
490     jmp,s   B87
490
492   B83: #    B87 <- B11  Freq: 2.49656e-06
492     # exception oop is in rax; no code emitted
492     movq    RSI, RAX        # spill
495     jmp,s   B87
495
497   B84: #    B87 <- B15  Freq: 2.49651e-06
497     # exception oop is in rax; no code emitted
497     movq    RSI, RAX        # spill
49a     jmp,s   B87
49a
49c   B85: #    B87 <- B27  Freq: 2.4965e-06
49c     # exception oop is in rax; no code emitted
49c     movq    RSI, RAX        # spill
49f     jmp,s   B87
49f
4a1   B86: #    B87 <- B31  Freq: 2.49645e-06
4a1     # exception oop is in rax; no code emitted
4a1     movq    RSI, RAX        # spill
4a1
4a4   B87: #    N805 <- B82 B79 B83 B77 B84 B76 B81 B78 B85 B75 B86 B74 B80  Freq: 9.98603e-06
4a4     addq    rsp, 48 # Destroy frame
        popq   rbp

4a9     jmp     rethrow_stub
4a9

読み方がよく分かりませんが、明らかにコードサイズが大きくなってますね。

以上から fib 関数はインライン化されていることが分かりました。こんなに インライン化されては、さすがの SBCL でも敵いません。

結論: Clojure コンパイラがすごいんじゃなくて HotSpot がすごい。 まあで もプリミティブ型をちゃんとコンパイルするようになったのは偉いか。

ところで SBCL が再帰関数のインライン化を行わないのは、何か理由があるの でしょうか。僕が知る限り、インライン化にこれといった障壁はないと思うの ですが。

2 件のコメント:

  1. :gen-classつけたりでコンパイルする場合バイトコードのclassファイルができるので、その場合はjavapできたと思います。

    返信削除
  2. このコメントは投稿者によって削除されました。

    返信削除