[llvm-dev] [RFC] Spill2Reg: Selectively replace spills to stack with spills to vector registers

[Vasileios Porpodas via llvm-dev]

Hi everyone,

This is an RFC for a new machine IR pass called Spill2Reg, which selectively replaces spills to the stack with spills to vector registers.
The chain of patches can be found here: https://reviews.llvm.org/D118298 (it includes all patches from D118298 to D118305).

### Overview
The register allocator emits spill/reload instructions to temporarily save register values to memory. These are typically stores to the stack (aka spills) and loads from the stack (aka reloads or fills). These instructions hurt performance not only because they increase the number of instructions executed, but also because they add pressure to the memory resources, which may already be heavily used by the actual memory instructions of the workload.

Spill2Reg aims at reducing this overhead by selectively replacing spills/reloads to/from the stack with spills/reloads to/from registers only when it is profitable. The target registers that will hold the spilled value must be of a different class than those that caused the spill, and in architectures like x86 or ARM we can use vector registers to save values from general purpose registers. Spill2reg can be profitable in a couple of cases: (i) on targets where spills to stack are always slower than spills to registers, (ii) in pathological cases with lots of spill/reload instructions back-to-back, (iii) in memory intensive workloads. It is worth pointing out that Spill2Reg can be profitable even on targets where spills to registers and spills to stack have a similar latency. This is because replacing some of the stack instructions with register instructions can help remove some stalls caused by bottle-necks in the memory resources.

Early evaluation on a Skylake(Server) x86_64 system shows that Spill2Reg can improve performance of both synthetic tests and of real-life workloads.


Why Spill to Registers?
=======================
There are a couple of reasons why it makes sense to spill to registers instead of memory. To summarize: (i) there is usually a lot of free register space even when spilling and (ii) spilling to vector registers can remove back-end stalls. The following sections discuss these points in more detail.


1. Free register space even when spilling
-----------------------------------------
Consider the following code:
```
int D0, D1, D2, ..., D18;
foo() {
   int t0 = D0
   int t1 = D1
   ...
   int t18 = D18
   // Some code
   ... = t0
   ... = t1
       ...
   ... = t18
}
```

Variables t0 to t18 are all live across the middle point (marked with `// Some code`). When compiled for x86_64, this code will assign t0 to t14 to registers, but will spill t15 to t18.
Here is what the assembly looks like:
```
        movl    D0(%rip), %eax
        movl    %eax, -8(%rsp)                  # 4-byte Spill
        movl    D1(%rip), %ecx
        movl    D2(%rip), %edx
        movl    D3(%rip), %esi
        movl    D4(%rip), %edi
        movl    D5(%rip), %r8d
        movl    D6(%rip), %r9d
        movl    D7(%rip), %r10d
        movl    D8(%rip), %r11d
        movl    D9(%rip), %ebx
        movl    D10(%rip), %ebp
        movl    D11(%rip), %r14d
        movl    D12(%rip), %r15d
        movl    D13(%rip), %r12d
        movl    D14(%rip), %r13d
        movl    D15(%rip), %eax
        movl    %eax, -4(%rsp)                  # 4-byte Spill
        movl    D16(%rip), %eax
        movl    %eax, -12(%rsp)                 # 4-byte Spill
        movl    D17(%rip), %eax
        movl    %eax, -16(%rsp)                 # 4-byte Spill
        movl    D18(%rip), %eax
        movl    %eax, -20(%rsp)                 # 4-byte Spill
        # ...  Some code ...
        movl    -8(%rsp), %eax                  # 4-byte Reload
        movl    %eax, U0(%rip)
        movl    %ecx, U1(%rip)
        movl    %edx, U2(%rip)
        movl    %esi, U3(%rip)
        movl    %edi, U4(%rip)
        movl    %r8d, U5(%rip)
        movl    %r9d, U6(%rip)
        movl    %r10d, U7(%rip)
        movl    %r11d, U8(%rip)
        movl    %ebx, U9(%rip)
        movl    %ebp, U10(%rip)
        movl    %r14d, U11(%rip)
        movl    %r15d, U12(%rip)
        movl    %r12d, U13(%rip)
        movl    %r13d, U14(%rip)
        movl    -4(%rsp), %eax                  # 4-byte Reload
        movl    %eax, U15(%rip)
        movl    -12(%rsp), %eax                 # 4-byte Reload
        movl    %eax, U16(%rip)
        movl    -16(%rsp), %eax                 # 4-byte Reload
        movl    %eax, U17(%rip)
        movl    -20(%rsp), %eax                 # 4-byte Reload
        movl    %eax, U18(%rip)
```

Meanwhile there is a lot of free space in the vector register file, yet we are spilling to memory.


2. Memory and register spills use different ports in x86
--------------------------------------------------------
According to [1] spill/reloads to/from the stack use ports 2, 3, 4, 7 on a Skylake Server x86 micro-architecture, while spills/reloads to/from the vector register file use Ports 0, 1, 5. This means that the stack-based and vector-register-based spills/reloads use different back-end resources, and by replacing one with the other can shift the overhead from some type of resources to another.

This is particularly important for Spill2Reg, because it shows that it can improve performance even if spills-to-stack and spills-to-registers have a similar latency. If the CPU is stalling due to over-subscribed memory resources, Spill2Reg can replace some of them with spills-to-registers, which can help remove some of the stalls.


I think this looks familiar
===========================
If spilling to vector registers looks familiar, it is probably because GCC has included support for spilling to registers for quite some time. In x86 this was done with `vmovd` instructions spilling to `xmm` registers.

However, to the best of my knowledge, spilling to vector registers in x86 has been disabled [2] due to stability [3,4,5], correctness [6], and performance [7] issues.
The performance issues highlighted in [7] seem to be related to:
(i) double spilling: the vector register used for spilling get spilled to memory, and
(ii) folded reloads: if a reload can be folded into an instruction, then spilling to a vector register results in an additional instruction: the one extracting the value from the vector register and inserting it into the general purpose register.

I believe that the proposed Spill2Reg design can address all these issues. These points are discussed in the following sections.


Spill2Reg in x86
================
There are several interesting points about spilling to vector registers in x86:

- Writing/Reading to vector registers can be done with a single assembly instruction. So spilling to vector register does not introduce more instructions.
- The vector register space is quite large at 2KB (32 ZMM registers X 64 bytes) in Skylake server, so the chances are that there will be a lot of free vector register space in general purpose workloads.
- If needed we could insert values to any vector lane using `(V)PINSR{B,W,D,Q}` and `(V)PEXTR{B,W,D,Q}` (supported since SSE4.1), or to lane 0 using `MOVD` (supported since MMX).
- We can choose between two types of instructions for moving data to vector registers: `MOVD/Q` and `PINSRD/Q`:
  - `PINSRD/Q` `PEXTRD/Q` allow us to insert/extract a value to/from any lane of the vector register. It is implemented in SSE4_1+.
  - `MOVD/Q` moves a 32/64bit value to the first lane of a vector register. It is implemented in MMX+ instruction sets, so it is widely supported. GCC’s implementation uses `MOVD`.
  - According to Agner Fog's latency/throughput data for Skylake [9], `MOVD` has a lower latency than `PINSRD/Q`, same latency than `MOV` to/from memory, but lower throughput:

```
                  uops   uops      uops
                  fused  unfused   each   latency  throughput
                  domain domain    port
Spill-to-reg
------------
MOVD mm/x r32/64    1     1        p5       2       1
MOVD r32/64 mm/x    1     1        p0       2       1

PINSRD/Q x,r,i      2     2        2p5      3       2
PEXTRB/W/D/Q r,x,i  2     2       p0 p5     3       1

Spill-to-stack
--------------
MOV  m,r            1     2       p237 p4   2       1
MOV r32/64, m       1     1        p23      2       0.5
------------------------------------------------------------
mm  : 64-bit mmx register
x   : 128-bit xmm register
m   : memory operand
m32 : 32-bit memory operand

Source: Agner Fog's Instruction tables: https://www.agner.org/optimize/instruction_tables.pdf
```


Spill2Reg on ARM
----------------
According to ARM's software optimization guide [8] section 4.3 (page 57), it is encouraged to spill to vector registers because these instructions have a lower latency than spills to memory.
Also, given the RISC nature of ARM instructions, there are no folded memory operands that we have to worry about when applying Spill2Reg.


Caveats
=======
- Instruction throughput/latency: Vector insert/extract instructions may have lower throughput or higher latency than memory load/stores on certain targets. If this is the case they should be used only when memory resources are over subscribed.
- Double spilling: We must take extra care not to spill the vector registers that hold the spilled value, because this will only add overhead. This was also observed in GCC's implementation [7]. This is a concern when Spill2Reg is implemented as part of the register allocator's spiller.
- Additional instructions: Stack based spills/reloads can potentially be folded. This is quite common in x86. For example, given a reload `MOV Reg1, [Spill]` and an instruction using `Reg1`, like `ADD Reg2, Reg1`, this can be folded into `Add Reg2, [Spill]`, saving one instruction. Spill2Reg can potentially block this folding, which will add an additional instruction that reloads the data from the vector before the `ADD`. This is a concern when Spill2Reg is implemented as part of the register allocator's spiller, because folding of spills/reloads may be taking place later on, after all spills/reloads have been emitted.
- Frequency throttling: Some targets will lower their turbo frequency when running vector instructions [10]. Throttling is the highest for AVX-512 but decreases as the vector width decreases, and it seems that there is no frequency drop for 128-bit instructions.
- Profitability models: There may be different trade-offs for different targets or different generations of the same target. Each target should implement its own profitability model.


Proposed Design
===============
Spill2Reg can be implemented either as part of the spiller, within the register allocator, or as a standalone pass after register allocation. Each design point has its own benefits and shortcomings. The following sections list some of the most important points for and against each design. The discussion provides some insights into why we believe Spill2Reg is best implemented as a separate pass after the register allocator and after physical registers are introduced.

In the spiller versus as a standalone pass
------------------------------------------

### In the spiller
- Pros:
  - The spiller is a natural location for the Spill2Reg code.
  - All analysis data needed are already computed and readily available.
  - The actual logic is quite simple.

- Cons:
  - The register allocator expects that a Spill/Reload is a store/load. So supporting Spill2Reg requires some refactoring.
  - Spill/Reload removal optimizations within the register allocator need to be updated to handle spills to registers.
  - The register allocation pass is already quite complex, and adding Spill2Reg can increase complexity even further.
  - Folded spills/reloads (see GCC issue [7]): We need to be able to skip Spill2Reg if spills/reloads to the stack can be folded. Folding happens after the spill code is emitted, so implementing this is not trivial.
  - Double-spills must be avoided (see GCC issue [7]): The vector pseudo register that holds the spilled value needs to be colored by the register allocator itself, and we need to guarantee that it won't get spilled.

### As a standalone pass
- Pros:
  - Small pass, easy to design and maintain.
  - Easier pass testing.
  - Straight-forward tracking of performance improvements/regressions, without having to worry if changes in Spill2Reg may affect the decisions of the register allocator.

- Cons:
  - May replicate analysis that is already available within the register allocator.
  - Yet one more compiler pass.


Given the points listed above, and the performance/stability bugs reported in GCC [2,3,4,5,6,7], I believe that it makes sense to implement Spill2Reg as a standalone pass. This results in a simpler design, with simpler testing and easier performance tracking to avoid performance regressions.


As a standalone-pass: Before or after VirtRegRewriter
-----------------------------------------------------
Another design decision is whether we place the pass before or after the `Virtual Register Rewriter` pass.

### Before VirtRegRewriter
- Pros:
  - We are still working with pseudo-registers.
  - Straight forward checking of pseudo register interference using LiveIntervals and LiveRegMatrix.
 
- Cons:
  - The main issue is with testing. Testing Spill2Reg would require running the register allocator pass first, and relying on it to generate spill/reload code, which would make tests very verbose and tricky to write.
  - The pass would need to maintain the state of several data-structures, like LiveIntervals and LiveRegMatrix.
 
### After VirtRegRewriter
- Pros:
  - The main point is ease of testing: We can test the pass in isolation, using small tests containing hand-written spill/reload code. No need to run other passes before it.
  - Fewer data structures to maintain.

- Cons:
  - Spill2Reg needs its own live register tracking implementation since it can no longer rely on LiveIntervals and LiveRegMatrix for finding free physical registers.


Given that the pass design is quite similar in both cases, and that testing is significantly nicer in one of them, the preferred option is after VirtRegRewriter.


Target independent component
----------------------------
The Spill2Reg pass works as follows:
1. It collects all candidate spill/reloads. This filters out folded spills/reloads and unsupported data-types by the target (target dependent legality check).
2. Then it iterates through the collected candidates and checks if it is profitable to spill to the vector register (target dependent cost model).
3. If profitable, it generates the new spills/reloads to/from the vector register file and it removes the original instructions (target dependent spill/reload instruction generation).

Target dependent components
---------------------------
This includes:
- Legality checks that test whether the opcodes and types of spills/reloads can be handled by the target.
- The profitability heuristic, which checks whether we applying Spill2Reg for a specific set of spills/reloads will lead to better performing code on the target.
- The generation of spill/reload instructions to/from a vector register.


Profitability Heuristic in x86
------------------------------
Given a candidate set of spills/reloads, we need to decide whether applying Spill2Reg is more profitable. We are currently implementing an all or nothing approach for the whole set: we will either replace all spills/reloads or none.

According to [9] spills to vector registers have the same latency as spills to memory, but have lower throughput on Skylake. So replacing all spills to memory with spills to the stack can degrade performance. Instead, a better strategy is to spill to registers only when the memory units are over-subscribed, in an attempt to reduce any potential back-end stalls.

Since our goal is to avoid back-end stalls caused by bottle-necks on memory resources, we need some way to measure when these stalls could happen. Ideally we would query a pipeline model (like the one used in instruction schedulers) to determine if spills-to-stack can cause pipeline stalls. For now the implementation is based on a simple instruction count: if the count of memory instructions in the proximity of the spill/reload is above a threshold, then apply Spill2Reg.

Please note that the profitability heuristic is implemented as a target-dependent components, so other targets can implement their own specialized heuristics.


Performance Results
===================
The performance evaluation was done using `ba51d26ec4519f5b31de3acf643264504ea7bc7c` as a base commit on a Skylake Xeon Gold 6154. The code was compiled with `-O3 -march=native`.
Applying Spill2Reg to the code of the motivating example shown in a previous section, replaces some of the spills/reloads with `movd` instructions, leading to about 10% better performance.
Spill2Reg also works on real-life applications: In SPEC CPU 2017 525.x264 Spill2Reg improves performance by about 0.3%.



References
----------
[1] WikiChip Skylake microarchitecture: https://en.wikichip.org/wiki/intel/microarchitectures/skylake_(server)

[2] GCC spill to register: https://github.com/gcc-mirror/gcc/blob/5a431b60d1f221992e5e9f7a5c032df3b5fa35a5/gcc/config/i386/i386.c#L22874

[3] GCC stability bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70902

[4] GCC stability bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71596

[5] GCC stability bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71555

[6] GCC correctness bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71657

[7] GCC performance bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71453

[8] ARM Software Optimization Guide: https://developer.arm.com/documentation/swog309707/a

[9] Agner Fog's Instruction tables: https://www.agner.org/optimize/instruction_tables.pdf

[10] Daniel Lemire's AVX throttling blog: https://lemire.me/blog/2018/09/07/avx-512-when-and-how-to-use-these-new-instructions/


Thanks,
Vasileios

[Philip Reames via llvm-dev]

This is a wonderful writeup of an interesting design problem, and the chosen solution.  I learned several things from reading it.  Thank you taking the time to write this up.

My takeaway after reading was that you've clearly thought through all the issues in some depth.  I might not lean the same direction on each choice, but you've more than convinced me that you've thought about them and made a reasoned choice.  Solid +1 on the design here.

Philip

_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

[Vasileios Porpodas via llvm-dev]

Thanks Philip for the feedback, I am glad you enjoyed reading it.

Vasileios

[Sjoerd Meijer via llvm-dev]

Same here, enjoyed reading this. Just a drive by comment about this:

> Spill2Reg also works on real-life applications: In SPEC CPU 2017 525.x264 Spill2Reg improves performance by about 0.3%.

Our codegen for SPEC is relatively inefficient and there is a lot to gain here (compared to GCC), especially for x264 where we are more than 20% behind. That's for AArch64 by the way, but since we are mostly fixing target independent things, I guess that's also the case for X86. While this 0.3% is nice (every little helps), I think things would look quite different when the inefficiencies are addressed (we are working on it/some), and spill2reg that is fixing up things now may not be able to do so when that is the case. Long story short, we only have a 10% uplift for 1 case, but we need some more performance numbers? Ideally for AArch64 too? 🙂

But thanks again for the write up and the work, really nice.

Cheers.

---

From: llvm-dev <llvm-dev...@lists.llvm.org> on behalf of Vasileios Porpodas via llvm-dev <llvm...@lists.llvm.org>
Sent: 27 January 2022 07:34
To: Philip Reames <list...@philipreames.com>
Cc: llvm...@lists.llvm.org <llvm...@lists.llvm.org>
Subject: Re: [llvm-dev] [RFC] Spill2Reg: Selectively replace spills to stack with spills to vector registers

 

[Quentin Colombet via llvm-dev]

Hi Vasileios,

Spilling to registers is actually already possible with the existing allocator.

See https://lists.llvm.org/pipermail/llvm-dev/2019-December/137700.html

Depending on how you write to your vector registers that may be tricky to pull off, but I thought I share the information to make sure we’re not reinventing something that is already there.

Cheers,

-Quentin

[Momchil Velikov via llvm-dev]

It's interesting how this interacts with lazy save/restore of the FP context across context switches.

~chill

--

Compiler scrub, Arm

[Vasileios Porpodas via llvm-dev]

Thanks for your comments Sjoerd,

So far I have only tried a small synthetic test on AArch64, like the one with the back-to-back spills and reloads, and it seems to improve performance there too. I definitely need to do some more extensive performance evaluation, including AArch64.

I have seen a few SPEC benchmarks improve, not just x264, but this depends on the compiler options used and the exact compiler commit checked out. But yeah the performance improvements are usually around .5%.

Vasileios

[Madhur Amilkanthwar via llvm-dev]

+1. Liked the overall approach and thought process. BTW, in case you didn't notice, AMDGPU backend does have a similar pass in place. 

I have some suggestions/thoughts.

> Please note that the profitability heuristic is implemented as a target-dependent components, so other targets can implement their own specialized heuristics.

Was there any thought about a finer control? e.g. Controlling spilling decisions for a certain region of the code like loop. Attaching metadata to a loop can help control the decisions. It is quite possible that spilling to vector registers is profitable in one loop but not other. Designing robust heuristics is a hard problem. 

[Vasileios Porpodas via llvm-dev]

Hi Quentin,

Thanks for pointing that out. I have not actually tried that exact configuration you are referring to, but what I tried instead was generating spills to vector registers in the spiller, which was fairly straightforward. I would assume that both approaches would have the same issues I mention in the RFC. To summarize, the most important ones that were not straightforward to handle from within the register allocator: 

i. avoiding performance regressions: avoiding spilling to register when the spill can be folded, and also avoiding double spilling.

ii. not trivial changes in the existing code: updating the spill/reload removal optimizations to handle non-memory instruction spills, and

iii. testing is quite a bit trickier than when this is done in a separate pass.

Another point is that the spill2reg pass itself is not very large or complex (~500 lines of code), so I think it makes sense to keep it as a separate pass.

Thanks,

Vasileios

[Vasileios Porpodas via llvm-dev]

Hi Momchil,

Spill2Reg definitely adds to the FP context, but I am not sure how it would interact with lazy save/restore.

Vasileios

[Vasileios Porpodas via llvm-dev]

Hi Madhur,

Thank you for the feedback. I am not aware of the AMDGPU backend pass, which pass are you referring to?

The current goal, on x86 at least, is to spill to vectors only when we get stalls due to the memory units being fully utilized. This is a safe choice because it helps avoid performance regressions. The current heuristic is fairly simple, it just counts nearby instructions and checks against a threshold, but I assume that this can't be very accurate. So I think the next step is to query a pipeline model about stalls in the memory units instead. This should be relatively safe.

The next thing we can improve on is to step away from the all-or-nothing approach. So instead of applying spill2reg only when all spills/reloads are causing stalls, we could apply it even if a fraction of them are doing so. This would definitely require more sophisticated heuristics or manually marked regions, like what you are describing, because it would be much harder to avoid performance regressions. But I think this should also improve Spill2Reg's coverage quite a bit, and possibly lead to more speedups. So yeah, there is a lot of room for improving the profitability decision, and any ideas on this are always welcome.

Thanks,

Vasileios

[Craig Topper via llvm-dev]

On Thu, Jan 27, 2022 at 11:25 AM Vasileios Porpodas via llvm-dev <llvm...@lists.llvm.org> wrote:

Hi Momchil,

Spill2Reg definitely adds to the FP context, but I am not sure how it would interact with lazy save/restore.

Vasileios

The XSAVE instructions allow the OS to skip restoring the XMM registers on context switches if they are in their initial state or haven't changed. Spilling to XMM registers will make the registers dirty in more cases than they might have been otherwise.

That reminds me that this needs to be disabled when the NoImplicitFloat attribute is present. One use for NoImplicitFloat is to prevent uses of FP and vector registers when compiling OS kernels. If the kernel code doesn't change the FP or vector registers, they don't need to be saved/restored during system calls. I hope you already took into account -mno-sse/-no-sse2.

~Craig

 

On Thu, Jan 27, 2022 at 9:34 AM Momchil Velikov <momchil...@gmail.com> wrote:

It's interesting how this interacts with lazy save/restore of the FP context across context switches.

~chill

--

Compiler scrub, Arm

_______________________________________________

[Vasileios Porpodas via llvm-dev]

Thank you for your comments Craig, 

Great point, I will add a check for NoImplicitFloat. 

Regarding `-mno-sse`, the pass is currently enabled only if `X86STI->hasSSE41()` so I guess this is covered.

Vasileios