In Java, can & be faster than &&?

SusanW picture SusanW · Sep 20, 2016 · Viewed 8.5k times · Source

In this code:

if (value >= x && value <= y) {

when value >= x and value <= y are as likely true as false with no particular pattern, would using the & operator be faster than using &&?

Specifically, I am thinking about how && lazily evaluates the right-hand-side expression (ie only if the LHS is true), which implies a conditional, whereas in Java & in this context guarantees strict evaluation of both (boolean) sub-expressions. The value result is the same either way.

But whilst a >= or <= operator will use a simple comparison instruction, the && must involve a branch, and that branch is susceptible to branch prediction failure - as per this Very Famous Question: Why is it faster to process a sorted array than an unsorted array?

So, forcing the expression to have no lazy components will surely be more deterministic and not be vulnerable to prediction failure. Right?

Notes:

  • obviously the answer to my question would be No if the code looked like this: if(value >= x && verySlowFunction()). I am focusing on "sufficiently simple" RHS expressions.
  • there's a conditional branch in there anyway (the if statement). I can't quite prove to myself that that is irrelevant, and that alternative formulations might be better examples, like boolean b = value >= x && value <= y;
  • this all falls into the world of horrendous micro-optimizations. Yeah, I know :-) ... interesting though?

Update Just to explain why I'm interested: I've been staring at the systems that Martin Thompson has been writing about on his Mechanical Sympathy blog, after he came and did a talk about Aeron. One of the key messages is that our hardware has all this magical stuff in it, and we software developers tragically fail to take advantage of it. Don't worry, I'm not about to go s/&&/\&/ on all my code :-) ... but there are a number of questions on this site on improving branch prediction by removing branches, and it occurred to me that the conditional boolean operators are at the core of test conditions.

Of course, @StephenC makes the fantastic point that bending your code into weird shapes can make it less easy for JITs to spot common optimizations - if not now, then in the future. And that the Very Famous Question mentioned above is special because it pushes the prediction complexity far beyond practical optimization.

I'm pretty much aware that in most (or almost all) situations, && is the clearest, simplest, fastest, best thing to do - although I'm very grateful to the people who have posted answers demonstrating this! I'm really interested to see if there are actually any cases in anyone's experience where the answer to "Can & be faster?" might be Yes...

Update 2: (Addressing advice that the question is overly broad. I don't want to make major changes to this question because it might compromise some of the answers below, which are of exceptional quality!) Perhaps an example in the wild is called for; this is from the Guava LongMath class (thanks hugely to @maaartinus for finding this):

public static boolean isPowerOfTwo(long x) {
    return x > 0 & (x & (x - 1)) == 0;
}

See that first &? And if you check the link, the next method is called lessThanBranchFree(...), which hints that we are in branch-avoidance territory - and Guava is really widely used: every cycle saved causes sea-levels to drop visibly. So let's put the question this way: is this use of & (where && would be more normal) a real optimization?

Answer

walen picture walen · Sep 20, 2016

Ok, so you want to know how it behaves at the lower level... Let's have a look at the bytecode then!

EDIT : added the generated assembly code for AMD64, at the end. Have a look for some interesting notes.
EDIT 2 (re: OP's "Update 2"): added asm code for Guava's isPowerOfTwo method as well.

Java source

I wrote these two quick methods:

public boolean AndSC(int x, int value, int y) {
    return value >= x && value <= y;
}

public boolean AndNonSC(int x, int value, int y) {
    return value >= x & value <= y;
}

As you can see, they are exactly the same, save for the type of AND operator.

Java bytecode

And this is the generated bytecode:

  public AndSC(III)Z
   L0
    LINENUMBER 8 L0
    ILOAD 2
    ILOAD 1
    IF_ICMPLT L1
    ILOAD 2
    ILOAD 3
    IF_ICMPGT L1
   L2
    LINENUMBER 9 L2
    ICONST_1
    IRETURN
   L1
    LINENUMBER 11 L1
   FRAME SAME
    ICONST_0
    IRETURN
   L3
    LOCALVARIABLE this Ltest/lsoto/AndTest; L0 L3 0
    LOCALVARIABLE x I L0 L3 1
    LOCALVARIABLE value I L0 L3 2
    LOCALVARIABLE y I L0 L3 3
    MAXSTACK = 2
    MAXLOCALS = 4

  // access flags 0x1
  public AndNonSC(III)Z
   L0
    LINENUMBER 15 L0
    ILOAD 2
    ILOAD 1
    IF_ICMPLT L1
    ICONST_1
    GOTO L2
   L1
   FRAME SAME
    ICONST_0
   L2
   FRAME SAME1 I
    ILOAD 2
    ILOAD 3
    IF_ICMPGT L3
    ICONST_1
    GOTO L4
   L3
   FRAME SAME1 I
    ICONST_0
   L4
   FRAME FULL [test/lsoto/AndTest I I I] [I I]
    IAND
    IFEQ L5
   L6
    LINENUMBER 16 L6
    ICONST_1
    IRETURN
   L5
    LINENUMBER 18 L5
   FRAME SAME
    ICONST_0
    IRETURN
   L7
    LOCALVARIABLE this Ltest/lsoto/AndTest; L0 L7 0
    LOCALVARIABLE x I L0 L7 1
    LOCALVARIABLE value I L0 L7 2
    LOCALVARIABLE y I L0 L7 3
    MAXSTACK = 3
    MAXLOCALS = 4

The AndSC (&&) method generates two conditional jumps, as expected:

  1. It loads value and x onto the stack, and jumps to L1 if value is lower. Else it keeps running the next lines.
  2. It loads value and y onto the stack, and jumps to L1 also, if value is greater. Else it keeps running the next lines.
  3. Which happen to be a return true in case none of the two jumps were made.
  4. And then we have the lines marked as L1 which are a return false.

The AndNonSC (&) method, however, generates three conditional jumps!

  1. It loads value and x onto the stack and jumps to L1 if value is lower. Because now it needs to save the result to compare it with the other part of the AND, so it has to execute either "save true" or "save false", it can't do both with the same instruction.
  2. It loads value and y onto the stack and jumps to L1 if value is greater. Once again it needs to save true or false and that's two different lines depending on the comparison result.
  3. Now that both comparisons are done, the code actually executes the AND operation -- and if both are true, it jumps (for a third time) to return true; or else it continues execution onto the next line to return false.

(Preliminary) Conclusion

Though I'm not that very much experienced with Java bytecode and I may have overlooked something, it seems to me that & will actually perform worse than && in every case: it generates more instructions to execute, including more conditional jumps to predict and possibly fail at.

A rewriting of the code to replace comparisons with arithmetical operations, as someone else proposed, might be a way to make & a better option, but at the cost of making the code much less clear.
IMHO it is not worth the hassle for 99% of the scenarios (it may be very well worth it for the 1% loops that need to be extremely optimized, though).

EDIT: AMD64 assembly

As noted in the comments, the same Java bytecode can lead to different machine code in different systems, so while the Java bytecode might give us a hint about which AND version performs better, getting the actual ASM as generated by the compiler is the only way to really find out.
I printed the AMD64 ASM instructions for both methods; below are the relevant lines (stripped entry points etc.).

NOTE: all methods compiled with java 1.8.0_91 unless otherwise stated.

Method AndSC with default options

  # {method} {0x0000000016da0810} 'AndSC' '(III)Z' in 'AndTest'
  ...
  0x0000000002923e3e: cmp    %r8d,%r9d
  0x0000000002923e41: movabs $0x16da0a08,%rax   ;   {metadata(method data for {method} {0x0000000016da0810} 'AndSC' '(III)Z' in 'AndTest')}
  0x0000000002923e4b: movabs $0x108,%rsi
  0x0000000002923e55: jl     0x0000000002923e65
  0x0000000002923e5b: movabs $0x118,%rsi
  0x0000000002923e65: mov    (%rax,%rsi,1),%rbx
  0x0000000002923e69: lea    0x1(%rbx),%rbx
  0x0000000002923e6d: mov    %rbx,(%rax,%rsi,1)
  0x0000000002923e71: jl     0x0000000002923eb0  ;*if_icmplt
                                                ; - AndTest::AndSC@2 (line 22)

  0x0000000002923e77: cmp    %edi,%r9d
  0x0000000002923e7a: movabs $0x16da0a08,%rax   ;   {metadata(method data for {method} {0x0000000016da0810} 'AndSC' '(III)Z' in 'AndTest')}
  0x0000000002923e84: movabs $0x128,%rsi
  0x0000000002923e8e: jg     0x0000000002923e9e
  0x0000000002923e94: movabs $0x138,%rsi
  0x0000000002923e9e: mov    (%rax,%rsi,1),%rdi
  0x0000000002923ea2: lea    0x1(%rdi),%rdi
  0x0000000002923ea6: mov    %rdi,(%rax,%rsi,1)
  0x0000000002923eaa: jle    0x0000000002923ec1  ;*if_icmpgt
                                                ; - AndTest::AndSC@7 (line 22)

  0x0000000002923eb0: mov    $0x0,%eax
  0x0000000002923eb5: add    $0x30,%rsp
  0x0000000002923eb9: pop    %rbp
  0x0000000002923eba: test   %eax,-0x1c73dc0(%rip)        # 0x0000000000cb0100
                                                ;   {poll_return}
  0x0000000002923ec0: retq                      ;*ireturn
                                                ; - AndTest::AndSC@13 (line 25)

  0x0000000002923ec1: mov    $0x1,%eax
  0x0000000002923ec6: add    $0x30,%rsp
  0x0000000002923eca: pop    %rbp
  0x0000000002923ecb: test   %eax,-0x1c73dd1(%rip)        # 0x0000000000cb0100
                                                ;   {poll_return}
  0x0000000002923ed1: retq   

Method AndSC with -XX:PrintAssemblyOptions=intel option

  # {method} {0x00000000170a0810} 'AndSC' '(III)Z' in 'AndTest'
  ...
  0x0000000002c26e2c: cmp    r9d,r8d
  0x0000000002c26e2f: jl     0x0000000002c26e36  ;*if_icmplt
  0x0000000002c26e31: cmp    r9d,edi
  0x0000000002c26e34: jle    0x0000000002c26e44  ;*iconst_0
  0x0000000002c26e36: xor    eax,eax            ;*synchronization entry
  0x0000000002c26e38: add    rsp,0x10
  0x0000000002c26e3c: pop    rbp
  0x0000000002c26e3d: test   DWORD PTR [rip+0xffffffffffce91bd],eax        # 0x0000000002910000
  0x0000000002c26e43: ret    
  0x0000000002c26e44: mov    eax,0x1
  0x0000000002c26e49: jmp    0x0000000002c26e38

Method AndNonSC with default options

  # {method} {0x0000000016da0908} 'AndNonSC' '(III)Z' in 'AndTest'
  ...
  0x0000000002923a78: cmp    %r8d,%r9d
  0x0000000002923a7b: mov    $0x0,%eax
  0x0000000002923a80: jl     0x0000000002923a8b
  0x0000000002923a86: mov    $0x1,%eax
  0x0000000002923a8b: cmp    %edi,%r9d
  0x0000000002923a8e: mov    $0x0,%esi
  0x0000000002923a93: jg     0x0000000002923a9e
  0x0000000002923a99: mov    $0x1,%esi
  0x0000000002923a9e: and    %rsi,%rax
  0x0000000002923aa1: cmp    $0x0,%eax
  0x0000000002923aa4: je     0x0000000002923abb  ;*ifeq
                                                ; - AndTest::AndNonSC@21 (line 29)

  0x0000000002923aaa: mov    $0x1,%eax
  0x0000000002923aaf: add    $0x30,%rsp
  0x0000000002923ab3: pop    %rbp
  0x0000000002923ab4: test   %eax,-0x1c739ba(%rip)        # 0x0000000000cb0100
                                                ;   {poll_return}
  0x0000000002923aba: retq                      ;*ireturn
                                                ; - AndTest::AndNonSC@25 (line 30)

  0x0000000002923abb: mov    $0x0,%eax
  0x0000000002923ac0: add    $0x30,%rsp
  0x0000000002923ac4: pop    %rbp
  0x0000000002923ac5: test   %eax,-0x1c739cb(%rip)        # 0x0000000000cb0100
                                                ;   {poll_return}
  0x0000000002923acb: retq   

Method AndNonSC with -XX:PrintAssemblyOptions=intel option

  # {method} {0x00000000170a0908} 'AndNonSC' '(III)Z' in 'AndTest'
  ...
  0x0000000002c270b5: cmp    r9d,r8d
  0x0000000002c270b8: jl     0x0000000002c270df  ;*if_icmplt
  0x0000000002c270ba: mov    r8d,0x1            ;*iload_2
  0x0000000002c270c0: cmp    r9d,edi
  0x0000000002c270c3: cmovg  r11d,r10d
  0x0000000002c270c7: and    r8d,r11d
  0x0000000002c270ca: test   r8d,r8d
  0x0000000002c270cd: setne  al
  0x0000000002c270d0: movzx  eax,al
  0x0000000002c270d3: add    rsp,0x10
  0x0000000002c270d7: pop    rbp
  0x0000000002c270d8: test   DWORD PTR [rip+0xffffffffffce8f22],eax        # 0x0000000002910000
  0x0000000002c270de: ret    
  0x0000000002c270df: xor    r8d,r8d
  0x0000000002c270e2: jmp    0x0000000002c270c0
  • First of all, the generated ASM code differs depending on whether we choose the default AT&T syntax or the Intel syntax.
  • With AT&T syntax:
    • The ASM code is actually longer for the AndSC method, with every bytecode IF_ICMP* translated to two assembly jump instructions, for a total of 4 conditional jumps.
    • Meanwhile, for the AndNonSC method the compiler generates a more straight-forward code, where each bytecode IF_ICMP* is translated to only one assembly jump instruction, keeping the original count of 3 conditional jumps.
  • With Intel syntax:
    • The ASM code for AndSC is shorter, with just 2 conditional jumps (not counting the non-conditional jmp at the end). Actually it's just two CMP, two JL/E and a XOR/MOV depending on the result.
    • The ASM code for AndNonSC is now longer than the AndSC one! However, it has just 1 conditional jump (for the first comparison), using the registers to directly compare the first result with the second, without any more jumps.

Conclusion after ASM code analysis

  • At AMD64 machine-language level, the & operator seems to generate ASM code with fewer conditional jumps, which might be better for high prediction-failure rates (random values for example).
  • On the other hand, the && operator seems to generate ASM code with fewer instructions (with the -XX:PrintAssemblyOptions=intel option anyway), which might be better for really long loops with prediction-friendly inputs, where the fewer number of CPU cycles for each comparison can make a difference in the long run.

As I stated in some of the comments, this is going to vary greatly between systems, so if we're talking about branch-prediction optimization, the only real answer would be: it depends on your JVM implementation, your compiler, your CPU and your input data.


Addendum: Guava's isPowerOfTwo method

Here, Guava's developers have come up with a neat way of calculating if a given number is a power of 2:

public static boolean isPowerOfTwo(long x) {
    return x > 0 & (x & (x - 1)) == 0;
}

Quoting OP:

is this use of & (where && would be more normal) a real optimization?

To find out if it is, I added two similar methods to my test class:

public boolean isPowerOfTwoAND(long x) {
    return x > 0 & (x & (x - 1)) == 0;
}

public boolean isPowerOfTwoANDAND(long x) {
    return x > 0 && (x & (x - 1)) == 0;
}

Intel's ASM code for Guava's version

  # {method} {0x0000000017580af0} 'isPowerOfTwoAND' '(J)Z' in 'AndTest'
  # this:     rdx:rdx   = 'AndTest'
  # parm0:    r8:r8     = long
  ...
  0x0000000003103bbe: movabs rax,0x0
  0x0000000003103bc8: cmp    rax,r8
  0x0000000003103bcb: movabs rax,0x175811f0     ;   {metadata(method data for {method} {0x0000000017580af0} 'isPowerOfTwoAND' '(J)Z' in 'AndTest')}
  0x0000000003103bd5: movabs rsi,0x108
  0x0000000003103bdf: jge    0x0000000003103bef
  0x0000000003103be5: movabs rsi,0x118
  0x0000000003103bef: mov    rdi,QWORD PTR [rax+rsi*1]
  0x0000000003103bf3: lea    rdi,[rdi+0x1]
  0x0000000003103bf7: mov    QWORD PTR [rax+rsi*1],rdi
  0x0000000003103bfb: jge    0x0000000003103c1b  ;*lcmp
  0x0000000003103c01: movabs rax,0x175811f0     ;   {metadata(method data for {method} {0x0000000017580af0} 'isPowerOfTwoAND' '(J)Z' in 'AndTest')}
  0x0000000003103c0b: inc    DWORD PTR [rax+0x128]
  0x0000000003103c11: mov    eax,0x1
  0x0000000003103c16: jmp    0x0000000003103c20  ;*goto
  0x0000000003103c1b: mov    eax,0x0            ;*lload_1
  0x0000000003103c20: mov    rsi,r8
  0x0000000003103c23: movabs r10,0x1
  0x0000000003103c2d: sub    rsi,r10
  0x0000000003103c30: and    rsi,r8
  0x0000000003103c33: movabs rdi,0x0
  0x0000000003103c3d: cmp    rsi,rdi
  0x0000000003103c40: movabs rsi,0x175811f0     ;   {metadata(method data for {method} {0x0000000017580af0} 'isPowerOfTwoAND' '(J)Z' in 'AndTest')}
  0x0000000003103c4a: movabs rdi,0x140
  0x0000000003103c54: jne    0x0000000003103c64
  0x0000000003103c5a: movabs rdi,0x150
  0x0000000003103c64: mov    rbx,QWORD PTR [rsi+rdi*1]
  0x0000000003103c68: lea    rbx,[rbx+0x1]
  0x0000000003103c6c: mov    QWORD PTR [rsi+rdi*1],rbx
  0x0000000003103c70: jne    0x0000000003103c90  ;*lcmp
  0x0000000003103c76: movabs rsi,0x175811f0     ;   {metadata(method data for {method} {0x0000000017580af0} 'isPowerOfTwoAND' '(J)Z' in 'AndTest')}
  0x0000000003103c80: inc    DWORD PTR [rsi+0x160]
  0x0000000003103c86: mov    esi,0x1
  0x0000000003103c8b: jmp    0x0000000003103c95  ;*goto
  0x0000000003103c90: mov    esi,0x0            ;*iand
  0x0000000003103c95: and    rsi,rax
  0x0000000003103c98: and    esi,0x1
  0x0000000003103c9b: mov    rax,rsi
  0x0000000003103c9e: add    rsp,0x50
  0x0000000003103ca2: pop    rbp
  0x0000000003103ca3: test   DWORD PTR [rip+0xfffffffffe44c457],eax        # 0x0000000001550100
  0x0000000003103ca9: ret    

Intel's asm code for && version

  # {method} {0x0000000017580bd0} 'isPowerOfTwoANDAND' '(J)Z' in 'AndTest'
  # this:     rdx:rdx   = 'AndTest'
  # parm0:    r8:r8     = long
  ...
  0x0000000003103438: movabs rax,0x0
  0x0000000003103442: cmp    rax,r8
  0x0000000003103445: jge    0x0000000003103471  ;*lcmp
  0x000000000310344b: mov    rax,r8
  0x000000000310344e: movabs r10,0x1
  0x0000000003103458: sub    rax,r10
  0x000000000310345b: and    rax,r8
  0x000000000310345e: movabs rsi,0x0
  0x0000000003103468: cmp    rax,rsi
  0x000000000310346b: je     0x000000000310347b  ;*lcmp
  0x0000000003103471: mov    eax,0x0
  0x0000000003103476: jmp    0x0000000003103480  ;*ireturn
  0x000000000310347b: mov    eax,0x1            ;*goto
  0x0000000003103480: and    eax,0x1
  0x0000000003103483: add    rsp,0x40
  0x0000000003103487: pop    rbp
  0x0000000003103488: test   DWORD PTR [rip+0xfffffffffe44cc72],eax        # 0x0000000001550100
  0x000000000310348e: ret    

In this specific example, the JIT compiler generates far less assembly code for the && version than for Guava's & version (and, after yesterday's results, I was honestly surprised by this).
Compared to Guava's, the && version translates to 25% less bytecode for JIT to compile, 50% less assembly instructions, and only two conditional jumps (the & version has four of them).

So everything points to Guava's & method being less efficient than the more "natural" && version.

... Or is it?

As noted before, I'm running the above examples with Java 8:

C:\....>java -version
java version "1.8.0_91"
Java(TM) SE Runtime Environment (build 1.8.0_91-b14)
Java HotSpot(TM) 64-Bit Server VM (build 25.91-b14, mixed mode)

But what if I switch to Java 7?

C:\....>c:\jdk1.7.0_79\bin\java -version
java version "1.7.0_79"
Java(TM) SE Runtime Environment (build 1.7.0_79-b15)
Java HotSpot(TM) 64-Bit Server VM (build 24.79-b02, mixed mode)
C:\....>c:\jdk1.7.0_79\bin\java -XX:+UnlockDiagnosticVMOptions -XX:CompileCommand=print,*AndTest.isPowerOfTwoAND -XX:PrintAssemblyOptions=intel AndTestMain
  .....
  0x0000000002512bac: xor    r10d,r10d
  0x0000000002512baf: mov    r11d,0x1
  0x0000000002512bb5: test   r8,r8
  0x0000000002512bb8: jle    0x0000000002512bde  ;*ifle
  0x0000000002512bba: mov    eax,0x1            ;*lload_1
  0x0000000002512bbf: mov    r9,r8
  0x0000000002512bc2: dec    r9
  0x0000000002512bc5: and    r9,r8
  0x0000000002512bc8: test   r9,r9
  0x0000000002512bcb: cmovne r11d,r10d
  0x0000000002512bcf: and    eax,r11d           ;*iand
  0x0000000002512bd2: add    rsp,0x10
  0x0000000002512bd6: pop    rbp
  0x0000000002512bd7: test   DWORD PTR [rip+0xffffffffffc0d423],eax        # 0x0000000002120000
  0x0000000002512bdd: ret    
  0x0000000002512bde: xor    eax,eax
  0x0000000002512be0: jmp    0x0000000002512bbf
  .....

Surprise! The assembly code generated for the & method by the JIT compiler in Java 7, has only one conditional jump now, and is way shorter! Whereas the && method (you'll have to trust me on this one, I don't want to clutter the ending!) remains about the same, with its two conditional jumps and a couple less instructions, tops.
Looks like Guava's engineers knew what they were doing, after all! (if they were trying to optimize Java 7 execution time, that is ;-)

So back to OP's latest question:

is this use of & (where && would be more normal) a real optimization?

And IMHO the answer is the same, even for this (very!) specific scenario: it depends on your JVM implementation, your compiler, your CPU and your input data.