INC instruction vs ADD 1: Does it matter?
Depending on the CPU implementation of the instructions, a partial register update may cause a stall. According to Agner Fog's optimization guide, page 62,
For historical reasons, the
INC
andDEC
instructions leave the carry flag unchanged, while the other arithmetic flags are written to. This causes a false dependence on the previous value of the flags and costs an extra μop. To avoid these problems, it is recommended that you always useADD
andSUB
instead ofINC
andDEC
. For example,INC EAX
should be replaced byADD EAX,1
.
See also page 83 on "Partial flags stalls" and page 100 on "Partial flags stall".
Update: the Efficiency cores on Alder Lake are Gracemont, and run inc reg
as a single uop, but at only 1/clock, vs. 4/clock for add reg, 1
(https://uops.info/). This may be a false dependency on FLAGS like P4 had; the uops.info tests didn't try adding a dep-breaking instruction. Other than the TL:DR, I haven't updated other parts of this answer.
TL:DR/advice for modern CPUs: Probably use add
; Intel Alder Lake's E-cores are relevant for "generic" tuning and seem to run inc
slowly.
Other than Alder Lake and earlier Silvermont-family, use inc
except with a memory destination; that's fine on mainstream Intel or any AMD. (e.g. like gcc -mtune=core2
, -mtune=haswell
, or -mtune=znver1
). inc mem
costs an extra uop vs. add
on Intel P6 / SnB-family; the load can't micro-fuse.
If you care about Silvermont-family (including KNL in Xeon Phi, and some netbooks, chromebooks, and NAS servers), probably avoid inc
. add 1
only costs 1 extra byte in 64-bit code, or 2 in 32-bit code. But it's not a performance disaster (just locally 1 extra ALU port used, not creating false dependencies or big stalls), so if you don't care much about SMont then don't worry about it.
Writing CF instead of leaving it unmodified can potentially be useful with other surrounding code that might benefit from CF dep-breaking, e.g. shifts. See below.
If you want to inc/dec without touching any flags, lea eax, [rax+1]
runs efficiently and has the same code-size as add eax, 1
. (Usually on fewer possible execution ports than add/inc, though, so add/inc are better when destroying FLAGS is not a problem. https://agner.org/optimize/)
On modern CPUs, add
is never slower than inc
(except for indirect code-size / decode effects), but usually it's not faster either, so you should prefer inc
for code-size reasons. Especially if this choice is repeated many times in the same binary (e.g. if you are a compiler-writer).
inc
saves 1 byte (64-bit mode), or 2 bytes (opcodes 0x40..F inc r32
/dec r32
short form in 32-bit mode, re-purposed as the REX prefix for x86-64). This makes a small percentage difference in total code size. This helps instruction-cache hit rates, iTLB hit rate, and number of pages that have to be loaded from disk.
Advantages of inc
:
- code-size directly
- Not using an immediate can have uop-cache effects on Sandybridge-family, which could offset the better micro-fusion of
add
. (See Agner Fog's table 9.1 in the Sandybridge section of his microarch guide.) Perf counters can easily measure issue-stage uops, but it's harder to measure how things pack into the uop cache and uop-cache read bandwidth effects. - Leaving CF unmodified is an advantage in some cases, on CPUs where you can read CF after
inc
without a stall. (Not on Nehalem and earlier.)
There is one exception among modern CPUs: Silvermont/Goldmont/Knight's Landing decodes inc
/dec
efficiently as 1 uop, but expands to 2 in the allocate/rename (aka issue) stage. The extra uop merges partial flags. inc
throughput is only 1 per clock, vs. 0.5c (or 0.33c Goldmont) for independent add r32, imm8
because of the dep chain created by the flag-merging uops.
Unlike P4, the register result doesn't have a false-dep on flags (see below), so out-of-order execution takes the flag-merging off the latency critical path when nothing uses the flag result. (But the OOO window is much smaller than mainstream CPUs like Haswell or Ryzen.) Running inc
as 2 separate uops is probably a win for Silvermont in most cases; most x86 instructions write all the flags without reading them, breaking these flag dependency chains.
SMont/KNL has a queue between decode and allocate/rename (See Intel's optimization manual, figure 16-2) so expanding to 2 uops during issue can fill bubbles from decode stalls (on instructions like one-operand mul
, or pshufb
, which produce more than 1 uop from the decoder and cause a 3-7 cycle stall for microcode). Or on Silvermont, just an instruction with more than 3 prefixes (including escape bytes and mandatory prefixes), e.g. REX + any SSSE3 or SSE4 instruction. But note that there is a ~28 uop loop buffer, so small loops don't suffer from these decode stalls.
inc
/dec
aren't the only instructions that decode as 1 but issue as 2: push
/pop
, call
/ret
, and lea
with 3 components do this too. So do KNL's AVX512 gather instructions. Source: Intel's optimization manual, 17.1.2 Out-of-Order Engine (KNL). It's only a small throughput penalty (and sometimes not even that if anything else is a bigger bottleneck), so it's generally fine to still use inc
for "generic" tuning.
Intel's optimization manual still recommends add 1
over inc
in general, to avoid risks of partial-flag stalls. But since Intel's compiler doesn't do that by default, it's not too likely that future CPUs will make inc
slow in all cases, like P4 did.
Clang 5.0 and Intel's ICC 17 (on Godbolt) do use inc
when optimizing for speed (-O3
), not just for size. -mtune=pentium4
makes them avoid inc
/dec
, but the default -mtune=generic
doesn't put much weight on P4.
ICC17 -xMIC-AVX512
(equivalent to gcc's -march=knl
) does avoid inc
, which is probably a good bet in general for Silvermont / KNL. But it's not usually a performance disaster to use inc
, so it's probably still appropriate for "generic" tuning to use inc
/dec
in most code, especially when the flag result isn't part of the critical path.
Other than Silvermont, this is mostly-stale optimization advice left over from Pentium4. On modern CPUs, there's only a problem if you actually read a flag that wasn't written by the last insn that wrote any flags. e.g. in BigInteger adc
loops. (And in that case, you need to preserve CF so using add
would break your code.)
add
writes all the condition-flag bits in the EFLAGS register. Register-renaming makes write-only easy for out-of-order execution: see write-after-write and write-after-read hazards. add eax, 1
and add ecx, 1
can execute in parallel because they are fully independent of each other. (Even Pentium4 renames the condition flag bits separate from the rest of EFLAGS, since even add
leaves the interrupts-enabled and many other bits unmodified.)
On P4, inc
and dec
depend on the previous value of the all the flags, so they can't execute in parallel with each other or preceding flag-setting instructions. (e.g. add eax, [mem]
/ inc ecx
makes the inc
wait until after the add
, even if the add's load misses in cache.) This is called a false dependency. Partial-flag writes work by reading the old value of the flags, updating the bits other than CF, then writing the full flags.
All other out-of-order x86 CPUs (including AMD's), rename different parts of flags separately, so internally they do a write-only update to all the flags except CF. (source: Agner Fog's microarchitecture guide). Only a few instructions, like adc
or cmc
, truly read and then write flags. But also shl r, cl
(see below).
Cases where add dest, 1
is preferable to inc dest
, at least for Intel P6/SnB uarch families:
- Memory-destination:
add [rdi], 1
can micro-fuse the store and the load+add on Intel Core2 and SnB-family, so it's 2 fused-domain uops / 4 unfused-domain uops.inc [rdi]
can only micro-fuse the store, so it's 3F / 4U.
According to Agner Fog's tables, AMD and Silvermont run memory-destinc
andadd
the same, as a single macro-op / uop.
But beware of uop-cache effects with add [label], 1
which needs a 32-bit address and an 8-bit immediate for the same uop.
- Before a variable-count shift/rotate to break the dependency on flags and avoid partial-flag merging:
shl reg, cl
has an input dependency on the flags, because of unfortunate CISC history: it has to leave them unmodified if the shift count is 0.
On Intel SnB-family, variable-count shifts are 3 uops (up from 1 on Core2/Nehalem). AFAICT, two of the uops read/write flags, and an independent uop reads reg
and cl
, and writes reg
. It's a weird case of having better latency (1c + inevitable resource conflicts) than throughput (1.5c), and only being able to achieve max throughput if mixed with instructions that break dependencies on flags. (I posted more about this on Agner Fog's forum). Use BMI2 shlx
when possible; it's 1 uop and the count can be in any register.
Anyway, inc
(writing flags but leaving CF
unmodified) before variable-count shl
leaves it with a false dependency on whatever wrote CF last, and on SnB/IvB can require an extra uop to merge flags.
Core2/Nehalem manage to avoid even the false dep on flags: Merom runs a loop of 6 independent shl reg,cl
instructions at nearly two shifts per clock, same performance with cl=0 or cl=13. Anything better than 1 per clock proves there's no input-dependency on flags.
I tried loops with shl edx, 2
and shl edx, 0
(immediate-count shifts), but didn't see a speed difference between dec
and sub
on Core2, HSW, or SKL. I don't know about AMD.
Update: The nice shift performance on Intel P6-family comes at the cost of a large performance pothole which you need to avoid: when an instruction depends on the flag-result of a shift instruction: The front end stalls until the instruction is retired. (Source: Intel's optimization manual, (Section 3.5.2.6: Partial Flag Register Stalls)). So shr eax, 2
/ jnz
is pretty catastrophic for performance on Intel pre-Sandybridge, I guess! Use shr eax, 2
/ test eax,eax
/ jnz
if you care about Nehalem and earlier. Intel's examples makes it clear this applies to immediate-count shifts, not just count=cl
.
In processors based on Intel Core microarchitecture [this means Core 2 and later], shift immediate by 1 is handled by special hardware such that it does not experience partial flag stall.
Intel actually means the special opcode with no immediate, which shifts by an implicit 1
. I think there is a performance difference between the two ways of encoding shr eax,1
, with the short encoding (using the original 8086 opcode D1 /5
) producing a write-only (partial) flag result, but the longer encoding (C1 /5, imm8
with an immediate 1
) not having its immediate checked for 0 until execution time, but without tracking the flag output in the out-of-order machinery.
Since looping over bits is common, but looping over every 2nd bit (or any other stride) is very uncommon, this seems like a reasonable design choice. This explains why compilers like to test
the result of a shift instead of directly using flag results from shr
.
Update: for variable count shifts on SnB-family, Intel's optimization manual says:
3.5.1.6 Variable Bit Count Rotation and Shift
In Intel microarchitecture code name Sandy Bridge, The “ROL/ROR/SHL/SHR reg, cl” instruction has three micro-ops. When the flag result is not needed, one of these micro-ops may be discarded, providing better performance in many common usages. When these instructions update partial flag results that are subsequently used, the full three micro-ops flow must go through the execution and retirement pipeline, experiencing slower performance. In Intel microarchitecture code name Ivy Bridge, executing the full three micro-ops flow to use the updated partial flag result has additional delay.
Consider the looped sequence below:
loop: shl eax, cl add ebx, eax dec edx ; DEC does not update carry, causing SHL to execute slower three micro-ops flow jnz loop
The DEC instruction does not modify the carry flag. Consequently, the SHL EAX, CL instruction needs to execute the three micro-ops flow in subsequent iterations. The SUB instruction will update all flags. So replacing
DEC
withSUB
will allowSHL EAX, CL
to execute the two micro-ops flow.
Terminology
Partial-flag stalls happen when flags are read, if they happen at all. P4 never has partial-flag stalls, because they never need to be merged. It has false dependencies instead.
Several answers / comments mix up the terminology. They describe a false dependency, but then call it a partial-flag stall. It's a slowdown which happens because of writing only some of the flags, but the term "partial-flag stall" is what happens on pre-SnB Intel hardware when partial-flag writes have to be merged. Intel SnB-family CPUs insert an extra uop to merge flags without stalling. Nehalem and earlier stall for ~7 cycles. I'm not sure how big the penalty is on AMD CPUs.
(Note that partial-register penalties are not always the same as partial-flags, see below).
### Partial flag stall on Intel P6-family CPUs:
bigint_loop:
adc eax, [array_end + rcx*4] # partial-flag stall when adc reads CF
inc rcx # rcx counts up from negative values towards zero
# test rcx,rcx # eliminate partial-flag stalls by writing all flags, or better use add rcx,1
jnz
# this loop doesn't do anything useful; it's not normally useful to loop the carry-out back to the carry-in for the same accumulator.
# Note that `test` will change the input to the next adc, and so would replacing inc with add 1
In other cases, e.g. a partial flag write followed by a full flag write, or a read of only flags written by inc
, is fine. On SnB-family CPUs, inc/dec
can even macro-fuse with a jcc
, the same as add/sub
.
After P4, Intel mostly gave up on trying to get people to re-compile with -mtune=pentium4
or modify hand-written asm as much to avoid serious bottlenecks. (Tuning for a specific microarchitecture will always be a thing, but P4 was unusual in deprecating so many things that used to be fast on previous CPUs, and thus were common in existing binaries.) P4 wanted people to use a RISC-like subset of the x86, and also had branch-prediction hints as prefixes for JCC instructions. (It also had other serious problems, like the trace cache that just wasn't good enough, and weak decoders that meant bad performance on trace-cache misses. Not to mention the whole philosophy of clocking very high ran into the power-density wall.)
When Intel abandoned P4 (NetBurst uarch), they returned to P6-family designs (Pentium-M / Core2 / Nehalem) which inherited their partial-flag / partial-reg handling from earlier P6-family CPUs (PPro to PIII) which pre-dated the netburst mis-step. (Not everything about P4 was inherently bad, and some of the ideas re-appeared in Sandybridge, but overall NetBurst is widely considered a mistake.) Some very-CISC instructions are still slower than the multi-instruction alternatives, e.g. enter
, loop
, or bt [mem], reg
(because the value of reg affects which memory address is used), but these were all slow in older CPUs so compilers already avoided them.
Pentium-M even improved hardware support for partial-regs (lower merging penalties). In Sandybridge, Intel kept partial-flag and partial-reg renaming and made it much more efficient when merging is needed (merging uop inserted with no or minimal stall). SnB made major internal changes and is considered a new uarch family, even though it inherits a lot from Nehalem, and some ideas from P4. (But note that SnB's decoded-uop cache is not a trace cache, though, so it's a very different solution to the decoder throughput/power problem that NetBurst's trace cache tried to solve.)
For example, inc al
and inc ah
can run in parallel on P6/SnB-family CPUs, but reading eax
afterwards requires merging.
PPro/PIII stall for 5-6 cycles when reading the full reg. Core2/Nehalem stall for only 2 or 3 cycles while inserting a merging uop for partial regs, but partial flags are still a longer stall.
SnB inserts a merging uop without stalling, like for flags. Intel's optimization guide says that for merging AH/BH/CH/DH into the wider reg, inserting the merging uop takes an entire issue/rename cycle during which no other uops can be allocated. But for low8/low16, the merging uop is "part of the flow", so it apparently doesn't cause additional front-end throughput penalties beyond taking up one of the 4 slots in an issue/rename cycle.
In IvyBridge (or at least Haswell), Intel dropped partial-register renaming for low8 and low16 registers, keeping it only for high8 registers (AH/BH/CH/DH). Reading high8 registers has extra latency. Also, setcc al
has a false dependency on the old value of rax, unlike in Nehalem and earlier (and probably Sandybridge). See this HSW/SKL partial-register performance Q&A for the details.
(I've previously claimed that Haswell could merge AH with no uop, but that's not true and not what Agner Fog's guide says. I skimmed too quickly and unfortunately repeated my wrong understanding in lots of comments and other posts.)
AMD CPUs, and Intel Silvermont, don't rename partial regs (other than flags), so mov al, [mem]
has a false dependency on the old value of eax. (The upside is no partial-reg merging slowdowns when reading the full reg later.)
Normally, the only time add
instead of inc
will make your code faster on AMD or mainstream Intel is when your code actually depends on the doesn't-touch-CF behaviour of inc
. i.e. usually add
only helps when it would break your code, but note the shl
case mentioned above, where the instruction reads flags but usually your code doesn't care about that, so it's a false dependency.
If you do actually want to leave CF unmodified, pre SnB-family CPUs have serious problems with partial-flag stalls, but on SnB-family the overhead of having the CPU merge the partial flags is very low, so it can be best to keep using inc
or dec
as part of a loop condition when targeting those CPU, with some unrolling. (For details, see the BigInteger adc
Q&A I linked earlier). It can be useful to use lea
to do arithmetic without affecting flags at all, if you don't need to branch on the result.
Skylake doesn't have partial-flag merging costs
Update: Skylake doesn't have partial-flag merging uops at all: CF is just a separate register from the rest of FLAGS. Instructions that need both parts (like cmovbe
) read both inputs separately. That makes cmovbe
a 2-uop instruction, but most other cmovcc
instructions 1-uop on Skylake. See What is a Partial Flag Stall?.
adc
only reads CF so it can be single-uop on Skylake with no interaction at all with an inc
or dec
in the same loop.
(TODO: rewrite earlier parts of this answer.)