I'm reading Agner Fog's "Optimizing software in C++" (specific to x86 processors for Intel, AMD and VIA) and it states on page 34
Boolean variables are stored as 8-bit integers with the value 0 for false and 1 for true. Boolean variables are overdetermined in the sense that all operators that have Boolean variables as input check if the inputs have any other value than 0 or 1, but operators that have Booleans as output can produce no other value than 0 or 1. This makes operations with Boolean variables as input less efficient than necessary.
Is this still true today and on what compilers? Can you please give an example? The author states
The Boolean operations can be made much more efficient if it is known with certainty that the operands have no other values than 0 and 1. The reason why the compiler doesn't make such an assumption is that the variables might have other values if they are uninitialized or come from unknown sources.
Does this mean that if I take a function pointer bool(*)()
for example and call it, then operations on it produce inefficient code? Or is it the case when I access a boolean by dereferencing a pointer or reading from a reference and then operate on it?
The purpose here is clarity of intent. The return type is part of a functions interface and a bool tells you more about what to expect from the function than an int does. Even BOOL is more expressive than int , even though it's the same type it at least shows your intent.
A boolean type normally follows the smallest unit of addressable memory of the target machine (i.e. usually the 8bits byte). Access to memory is always in "chunks" (multiple of words, this is for efficiency at the hardware level, bus transactions): a boolean bit cannot be addressed "alone" in most CPU systems.
Boolean variables are stored as 16-bit (2-byte) numbers, but they can only be True or False. Boolean variables display as either: True or False (when Print is used), or. #TRUE# or #FALSE# (when Write # is used).
Boolean values A variable of the primitive data type boolean can have two values: true and false (Boolean literals).
TL:DR: current compilers still have bool
missed-optimizations when doing stuff like(a&&b) ? x : y
. But the reason why is not that they don't assume 0/1, they just suck at this.
Many uses of bool
are for locals, or inline functions, so booleanizing to a 0
/ 1
can optimize away and branch (or cmov or whatever) on the original condition. Only worry about optimizing bool
inputs / outputs when it does have to get passed/returned across something that doesn't inline, or really stored in memory.
Possible optimization guideline: combine bool
s from external sources (function args / memory) with bitwise operators, like a&b
. MSVC and ICC do better with this. IDK if it's ever worse for local bool
s. Beware that a&b
is only equivalent to a&&b
for bool
, not integer types. 2 && 1
is true, but 2 & 1
is 0 which is false. Bitwise OR doesn't have this problem.
IDK if this guideline will ever hurt for locals that were set from a comparison within the function (or in something that inlined). E.g. it might lead the compiler to actually make integer booleans instead of just using comparison results directly when possible. Also note that it doesn't seem to help with current gcc and clang.
Yes, C++ implementations on x86 store bool
in a byte that's always 0 or 1 (at least across function-call boundaries where the compiler has to respect the ABI / calling convention which requires this.)
Compilers do sometimes take advantage of this, e.g. for bool
->int
conversion even gcc 4.4 simply zero-extends to 32-bit (movzx eax, dil
). Clang and MSVC do this, too. C and C++ rules require this conversion to produce 0 or 1, so this behaviour is only safe if it's always safe to assume that a bool
function arg or global variable has a 0 or 1 value.
Even old compilers typically did take advantage of it for bool
->int
, but not in other cases. Thus, Agner is wrong about the reason when he says:
The reason why the compiler doesn't make such an assumption is that the variables might have other values if they are uninitialized or come from unknown sources.
MSVC CL19 does make code that assumes bool
function args are 0 or 1, so the Windows x86-64 ABI must guarantee this.
In the x86-64 System V ABI (used by everything other than Windows), the changelog for revision 0.98 says "Specify that _Bool
(aka bool
) is booleanized at the caller." I think even before that change, compilers were assuming it, but this just documents what compilers were already relying on. The current language in the x86-64 SysV ABI is:
3.1.2 Data Representation
Booleans, when stored in a memory object, are stored as single byte objects the value of which is always 0 (false) or 1 (true). When stored in integer registers (except for passing as arguments), all 8 bytes of the register are significant; any nonzero value is considered true.
The second sentence is nonsense: the ABI has no business telling compilers how to store things in registers inside a function, only at boundaries between different compilation units (memory / function args and return values). I reported this ABI defect a while ago on the github page where it's maintained.
3.2.3 Parameter passing:
When a value of type
_Bool
is returned or passed in a register or on the stack, bit 0 contains the truth value and bits 1 to 7 shall be zero16.(footnote 16): Other bits are left unspecified, hence the consumer side of those values can rely on it being 0 or 1 when truncated to 8 bit.
The language in the i386 System V ABI is the same, IIRC.
Any compiler that assumes 0/1 for one thing (e.g. conversion to int
) but fails to take advantage of it in other cases has a missed optimization. Unfortunately such missed-optimizations still exist, although they are rarer than when Agner wrote that paragraph about compilers always re-booleanizing.
(Source + asm on the Godbolt compiler explorer for gcc4.6 / 4.7, and clang/MSVC. See also Matt Godbolt's CppCon2017 talk What Has My Compiler Done for Me Lately? Unbolting the Compiler's Lid)
bool logical_or(bool a, bool b) { return a||b; } # gcc4.6.4 -O3 for the x86-64 System V ABI test dil, dil # test a against itself (for non-zero) mov eax, 1 cmove eax, esi # return a ? 1 : b; ret
So even gcc4.6 didn't re-booleanize b
, but it did miss the optimization that gcc4.7 makes: (and clang and later compilers as shown in other answers):
# gcc4.7 -O3 to present: looks ideal to me. mov eax, esi or eax, edi ret
(Clang's or dil, sil
/ mov eax, edi
is silly: it's guaranteed to cause a partial-register stall on Nehalem or earlier Intel when reading edi
after writing dil
, and it has worse code size from needing a REX prefix to use the low-8 part of edi. A better choice might be or dil,sil
/ movzx eax, dil
if you want to avoid reading any 32-bit registers in case your caller left some arg-passing registers with "dirty" partial registers.)
MSVC emits this code that checks a
then b
separately, completely failing to take advantage of anything, and even using xor al,al
instead of xor eax,eax
. So it has a false dependency on the old value of eax
on most CPUs (including Haswell/Skylake, which don't rename low-8 partial regs separately from the whole register, only AH/BH/...). This is just dumb. The only reason to ever use xor al,al
is when you explicitly want to preserve the upper bytes.
logical_or PROC ; x86-64 MSVC CL19 test cl, cl ; Windows ABI passes args in ecx, edx jne SHORT $LN3@logical_or test dl, dl jne SHORT $LN3@logical_or xor al, al ; missed peephole: xor eax,eax is strictly better ret 0 $LN3@logical_or: mov al, 1 ret 0 logical_or ENDP
ICC18 also doesn't take advantage of the known 0/1 nature of the inputs, it just uses an or
instruction to set flags according to the bitwise OR of the two inputs, and setcc
to produce a 0/1.
logical_or(bool, bool): # ICC18 xor eax, eax #4.42 movzx edi, dil #4.33 movzx esi, sil #4.33 or edi, esi #4.42 setne al #4.42 ret #4.42
ICC emits the same code even for bool bitwise_or(bool a, bool b) { return a|b; }
. It promotes to int
(with movzx
), and uses or
to set flags according to the bitwise OR. This is dumb compared to or dil,sil
/ setne al
.
For bitwise_or
, MSVC does just use an or
instruction (after movzx
on each input), but anyway doesn't re-booleanize.
Only ICC/MSVC were making dumb code with the simple function above, but this function still gives gcc and clang trouble:
int select(bool a, bool b, int x, int y) { return (a&&b) ? x : y; }
Source+asm on the Godbolt compiler explorer (Same source, different compilers selected vs. last time).
Looks simple enough; you'd hope that a smart compiler would do it branchlessly with one test
/cmov
. x86's test
instruction sets flags according to a bitwise AND. It's an AND instruction that doesn't actually write the destination. (Just like cmp
is a sub
that doesn't write the destination).
# hand-written implementation that no compilers come close to making select: mov eax, edx # retval = x test edi, esi # ZF = ((a & b) == 0) cmovz eax, ecx # conditional move: return y if ZF is set ret
But even the daily builds of gcc and clang on the Godbolt compiler explorer make much more complicated code, checking each boolean separately. They know how to optimize bool ab = a&&b;
if you return ab
, but even writing it that way (with a separate boolean variable to hold the result) doesn't manage to hand-hold them into making code that doesn't suck.
Note that test same,same
is exactly equivalent to cmp reg, 0
, and is smaller, so it's what compilers use.
Clang's version is strictly worse than my hand-written version. (Note that it requires that the caller zero-extended the bool
args to 32-bit, like it does for narrow integer types as an unofficial part of the ABI which it and gcc implement but only clang depends on).
select: # clang 6.0 trunk 317877 nightly build on Godbolt test esi, esi cmove edx, ecx # x = b ? y : x test edi, edi cmove edx, ecx # x = a ? y : x mov eax, edx # return x ret
gcc 8.0.0 20171110 nightly makes branchy code for this, similar to what older gcc versions do.
select(bool, bool, int, int): # gcc 8.0.0-pre 20171110 test dil, dil mov eax, edx ; compiling with -mtune=intel or -mtune=haswell would keep test/jcc together for macro-fusion. je .L8 test sil, sil je .L8 rep ret .L8: mov eax, ecx ret
MSVC x86-64 CL19 makes very similar branchy code. It's targeting the Windows calling convention, where integer args are in rcx, rdx, r8, r9.
select PROC test cl, cl ; a je SHORT $LN3@select mov eax, r8d ; retval = x test dl, dl ; b jne SHORT $LN4@select $LN3@select: mov eax, r9d ; retval = y $LN4@select: ret 0 ; 0 means rsp += 0 after popping the return address, not C return 0. ; MSVC doesn't emit the `ret imm16` opcode here, so IDK why they put an explicit 0 as an operand. select ENDP
ICC18 also makes branchy code, but with both mov
instructions after the branches.
select(bool, bool, int, int): test dil, dil #8.13 je ..B4.4 # Prob 50% #8.13 test sil, sil #8.16 jne ..B4.5 # Prob 50% #8.16 ..B4.4: # Preds ..B4.2 ..B4.1 mov edx, ecx #8.13 ..B4.5: # Preds ..B4.2 ..B4.4 mov eax, edx #8.13 ret #8.13
Trying to help the compiler by using
int select2(bool a, bool b, int x, int y) { bool ab = a&&b; return (ab) ? x : y; }
leads MSVC into making hilariously bad code:
;; MSVC CL19 -Ox = full optimization select2 PROC test cl, cl je SHORT $LN3@select2 test dl, dl je SHORT $LN3@select2 mov al, 1 ; ab = 1 test al, al ;; and then test/cmov on an immediate constant!!! cmovne r9d, r8d mov eax, r9d ret 0 $LN3@select2: xor al, al ;; ab = 0 test al, al ;; and then test/cmov on another path with known-constant condition. cmovne r9d, r8d mov eax, r9d ret 0 select2 ENDP
This is only with MSVC (and ICC18 has the same missed optimization of test/cmov on a register that was just set to a constant).
gcc and clang as usual don't make code as bad as MSVC; they make the same asm they do for select()
, which is still not good but at least trying to help them doesn't make it worse like with MSVC.
bool
with bitwise operators helps MSVC and ICCIn my very limited testing, |
and &
seem to work better than ||
and &&
for MSVC and ICC. Look at the compiler output for your own code with your compiler + compile options to see what happens.
int select_bitand(bool a, bool b, int x, int y) { return (a&b) ? x : y; }
Gcc still branches separately on separate test
s of the two inputs, same code as the other versions of select
. clang still does two separate test/cmov
, same asm as for the other source versions.
MSVC comes through and optimizes correctly, beating all the other compilers (at least in the stand-alone definition):
select_bitand PROC ;; MSVC test cl, dl ;; ZF = !(a & b) cmovne r9d, r8d mov eax, r9d ;; could have done the mov to eax in parallel with the test, off the critical path, but close enough. ret 0
ICC18 wastes two movzx
instructions zero-extending the bool
s to int
, but then makes the same code as MSVC
select_bitand: ## ICC18 movzx edi, dil #16.49 movzx esi, sil #16.49 test edi, esi #17.15 cmovne ecx, edx #17.15 mov eax, ecx #17.15 ret #17.15
I think this is not the case.
First of all, this reasoning is completely unacceptable:
The reason why the compiler doesn't make such an assumption is that the variables might have other values if they are uninitialized or come from unknown sources.
Let's check some code (compiled with clang 6, but GCC 7 and MSVC 2017 produces similar code).
Boolean or:
bool fn(bool a, bool b) { return a||b; } 0000000000000000 <fn(bool, bool)>: 0: 40 08 f7 or dil,sil 3: 40 88 f8 mov al,dil 6: c3 ret
As can be seen, no 0/1 check here, simple or
.
Convert bool to int:
int fn(bool a) { return a; } 0000000000000000 <fn(bool)>: 0: 40 0f b6 c7 movzx eax,dil 4: c3 ret
Again, no check, simple move.
Convert char to bool:
bool fn(char a) { return a; } 0000000000000000 <fn(char)>: 0: 40 84 ff test dil,dil 3: 0f 95 c0 setne al 6: c3 ret
Here, char is checked whether it is 0, or not, and bool value set to 0 or 1 accordingly.
So I think it is safe to say that the compiler uses bool in a way so it always contains a 0/1. It never checks its validity.
About efficiency: I think bool is optimal. The only case I can imagine, where this approach is not optimal is char->bool conversion. That operation could be a simple mov, if bool value wouldn't be restricted to 0/1. For all other operations, the current approach is equally good, or better.
EDIT: Peter Cordes mentioned ABI. Here's the relevant text from the System V ABI for AMD64 (the text for i386 is similar):
Booleans, when stored in a memory object, are stored as single byte objects the value of which is always 0 (false) or 1 (true). When stored in integer registers (except for passing as arguments), all 8 bytes of the register are significant; any nonzero value is considered true
So for platforms which follow SysV ABI, we can be sure that a bool
has a 0/1 value.
I searched for ABI document for MSVC, but unfortunately I didn't find anything about bool
.
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With