ziglang/zig
261
6.0k
Fork
You've already forked zig
772

ban (most) runtime slices on spriv #36118

Merged
alichraghi merged 1 commit from sphaerophoria/zig:ban-slices into master 2026年07月11日 19:50:20 +02:00
Contributor
Copy link

spirv slices are represented via an element pointer + len. On access we
get pointers to individual elements with OpPtrAccessChain

Unfortunately this is not a valid operation in most cases. For any array
defined in a function, we are generated with storage class "function".
The SPIRV spec only allows for variable pointers on data in the Workgroup
(addrspace(.shared)) or StorageBuffer storage classes (definition of
variable pointer in SPIRV spec 2.16)

SPIRV kernels with this constraint unmet cause segfaults on my machine
when trying to load them, and fail to check against spirv-val

AFAICT there is no way to represent a slice of data with storage class
"function". We can use OpAccessChain to extract an offset from an array,
however that would require a slice representation resembling {array,
offset, len}. Since arrays are typed according to their type + len,
there would be no way to re-assign a slice from a [4]f32 array to a
slice from a [2]f32 array. The types would just not match up

So the best we can do is just ban construction of slices in the
scenarios we cannot generate valid code for, and give a useful error
message to the caller

A few examples...

var buf: [3]f32 = undefined;
for (buf[0..3]) |v| {} // errors
fnThatAcceptsSlice(&buf); // errors
for (&buf) |v| {} // succeeds
comptime cbuf: [3]f32 = .{1, 2, 3};
for (cbuf[0..3]) |v| {} // errors
for (comptime cbuf[0..3]) |v| {} // succeeds

This is arguably over-restrictive. We theoretically could handle
buf[0..N] correctly, but currently SPIRV codegen will still produce
invalid OpPtrAccessChains, so I've just completely removed the ability
to produce slices

spirv slices are represented via an element pointer + len. On access we get pointers to individual elements with OpPtrAccessChain Unfortunately this is not a valid operation in most cases. For any array defined in a function, we are generated with storage class "function". The SPIRV spec only allows for variable pointers on data in the Workgroup (addrspace(.shared)) or StorageBuffer storage classes (definition of variable pointer in SPIRV spec 2.16) SPIRV kernels with this constraint unmet cause segfaults on my machine when trying to load them, and fail to check against spirv-val AFAICT there is no way to represent a slice of data with storage class "function". We can use OpAccessChain to extract an offset from an array, however that would require a slice representation resembling {array, offset, len}. Since arrays are typed according to their type + len, there would be no way to re-assign a slice from a [4]f32 array to a slice from a [2]f32 array. The types would just not match up So the best we can do is just ban construction of slices in the scenarios we cannot generate valid code for, and give a useful error message to the caller A few examples... var buf: [3]f32 = undefined; for (buf[0..3]) |v| {} // errors fnThatAcceptsSlice(&buf); // errors for (&buf) |v| {} // succeeds comptime cbuf: [3]f32 = .{1, 2, 3}; for (cbuf[0..3]) |v| {} // errors for (comptime cbuf[0..3]) |v| {} // succeeds This is arguably over-restrictive. We theoretically could handle buf[0..N] correctly, but currently SPIRV codegen will still produce invalid OpPtrAccessChains, so I've just completely removed the ability to produce slices
ban (most) runtime slices on spriv
Some checks failed
ci / aarch64-macos-release (pull_request) Successful in 42m58s
ci / x86_64-freebsd-release (pull_request) Successful in 1h0m16s
ci / x86_64-freebsd-debug (pull_request) Successful in 1h3m3s
ci / x86_64-netbsd-release (pull_request) Successful in 1h2m26s
ci / x86_64-netbsd-debug (pull_request) Successful in 1h11m31s
ci / aarch64-macos-debug (pull_request) Successful in 1h47m8s
ci / aarch64-linux-release (pull_request) Successful in 1h51m53s
ci / x86_64-openbsd-release (pull_request) Successful in 1h0m3s
ci / x86_64-openbsd-debug (pull_request) Successful in 1h10m30s
ci / powerpc64le-linux-release (pull_request) Successful in 2h2m57s
ci / aarch64-linux-debug (pull_request) Successful in 2h43m36s
ci / x86_64-windows-debug (pull_request) Successful in 1h26m45s
ci / powerpc64le-linux-debug (pull_request) Successful in 3h56m1s
ci / x86_64-windows-release (pull_request) Successful in 1h6m38s
ci / x86_64-linux-debug (pull_request) Successful in 1h39m15s
ci / s390x-linux-release (pull_request) Successful in 1h32m4s
ci / s390x-linux-debug (pull_request) Successful in 2h47m58s
ci / x86_64-linux-release (pull_request) Successful in 3h42m13s
ci / x86_64-linux-debug-llvm (pull_request) Successful in 4h54m1s
ci / loongarch64-linux-release (pull_request) Successful in 2h34m32s
ci / loongarch64-linux-debug (pull_request) Successful in 3h40m28s
ci / aarch64-freebsd-debug (pull_request) Has been cancelled
ci / aarch64-freebsd-release (pull_request) Has been cancelled
ci / aarch64-netbsd-debug (pull_request) Has been cancelled
ci / aarch64-netbsd-release (pull_request) Has been cancelled
ci / riscv64-linux-debug (pull_request) Has been cancelled
ci / riscv64-linux-release (pull_request) Has been cancelled
3a6fd93493
spirv slices are represented via an element pointer + len. On access we
get pointers to individual elements with OpPtrAccessChain
Unfortunately this is not a valid operation in most cases. For any array
defined in a function, we are generated with storage class "function".
The SPIRV spec only allows for variable pointers on data in the Workgroup
(addrspace(.shared)) or StorageBuffer storage classes (definition of
variable pointer in SPIRV spec 2.16)
SPIRV kernels with this constraint unmet cause segfaults on my machine
when trying to load them, and fail to check against spirv-val
AFAICT there is no way to represent a slice of data with storage class
"function". We can use OpAccessChain to extract an offset from an array,
however that would require a slice representation resembling {array,
offset, len}. Since arrays are typed according to their type + len,
there would be no way to re-assign a slice from a [4]f32 array to a
slice from a [2]f32 array. The types would just not match up
So the best we can do is just ban construction of slices in the
scenarios we cannot generate valid code for, and give a useful error
message to the caller
A few examples...
 var buf: [3]f32 = undefined;
 for (buf[0..3]) |v| {} // errors
 fnThatAcceptsSlice(&buf); // errors
 for (&buf) |v| {} // succeeds
 comptime cbuf: [3]f32 = .{1, 2, 3};
 for (cbuf[0..3]) |v| {} // errors
 for (comptime cbuf[0..3]) |v| {} // succeeds
This is arguably over-restrictive. We theoretically could handle
buf[0..N] correctly, but currently SPIRV codegen will still produce
invalid OpPtrAccessChains, so I've just completely removed the ability
to produce slices
Author
Contributor
Copy link

Once again, feel free to tell me that these PRs aren't welcome. I'm making changes that I find useful, but if they don't line up with the team's direction or priorities I can leave it alone :)

Once again, feel free to tell me that these PRs aren't welcome. I'm making changes that I find useful, but if they don't line up with the team's direction or priorities I can leave it alone :)
alichraghi requested changes 2026年07月10日 11:14:40 +02:00
Dismissed
src/Sema.zig Outdated
@ -29755,0 +29768,4 @@
if(!target.cpu.has(.spirv,.variable_pointers)and!target.cpu.has(.spirv,.variable_pointers_storage_buffer)){
returnsema.failWithOwnedErrorMsg(
block,
trysema.errMsg(src,"cannot construct slices without the .variable_pointers or .variable_pointers_storage_buffer features",.{}),

Tags should be wrapped inside quotes without the dot. .variable_pointers => 'variable_pointers'

Tags should be wrapped inside quotes without the dot. `.variable_pointers => 'variable_pointers'`
alichraghi marked this conversation as resolved
src/Sema.zig Outdated
@ -29755,0 +29779,4 @@
returnsema.failWithOwnedErrorMsg(block,msg:{
constmsg=trysema.errMsg(src,"cannot construct slice from address space {t}",.{address_space});
errdefermsg.destroy(sema.gpa);
trysema.errNote(src,msg,"the SPIRV backend can only slice data in the .shared or .storage_buffer address spaces",.{});
-- try sema.errNote(src, msg, "the SPIRV backend can only slice data in the .shared or .storage_buffer address spaces", .{});
++ try sema.errNote(src, msg, "only 'shared' and 'storage_buffer' address spaces support slicing", .{});
```diff -- try sema.errNote(src, msg, "the SPIRV backend can only slice data in the .shared or .storage_buffer address spaces", .{}); ++ try sema.errNote(src, msg, "only 'shared' and 'storage_buffer' address spaces support slicing", .{}); ```
Author
Contributor
Copy link

I threw the SPRIV backend part on there because I thought that was useful info. Does it make sense to keep that information somehow? Like to me this seems like it is no longer a globally true statement. If you're on a normal target you can slice outside of those spaces

I threw the SPRIV backend part on there because I thought that was useful info. Does it make sense to keep that information somehow? Like to me this seems like it is no longer a globally true statement. If you're on a normal target you can slice outside of those spaces

Makes sense. It indeed makes it less confusing. Don't forget to put the dash in SPIR-V though!

Makes sense. It indeed makes it less confusing. Don't forget to put the dash in `SPIR-V` though!
Author
Contributor
Copy link

Cool, going with "only 'shared' and 'storage_buffer' address spaces support slicing on SPIR-V"

Cool, going with "only 'shared' and 'storage_buffer' address spaces support slicing on SPIR-V"
alichraghi marked this conversation as resolved

Once again, feel free to tell me that these PRs aren't welcome. I'm making changes that I find useful, but if they don't line up with the team's direction or priorities I can leave it alone :)

Nope. Please send more PRs and find more bugs :P

> Once again, feel free to tell me that these PRs aren't welcome. I'm making changes that I find useful, but if they don't line up with the team's direction or priorities I can leave it alone :) Nope. Please send more PRs and find more bugs :P

Btw this needs a test case in test/cases/compile_errors

Btw this needs a test case in `test/cases/compile_errors`
sphaerophoria force-pushed ban-slices from 3a6fd93493
Some checks failed
ci / aarch64-macos-release (pull_request) Successful in 42m58s
ci / x86_64-freebsd-release (pull_request) Successful in 1h0m16s
ci / x86_64-freebsd-debug (pull_request) Successful in 1h3m3s
ci / x86_64-netbsd-release (pull_request) Successful in 1h2m26s
ci / x86_64-netbsd-debug (pull_request) Successful in 1h11m31s
ci / aarch64-macos-debug (pull_request) Successful in 1h47m8s
ci / aarch64-linux-release (pull_request) Successful in 1h51m53s
ci / x86_64-openbsd-release (pull_request) Successful in 1h0m3s
ci / x86_64-openbsd-debug (pull_request) Successful in 1h10m30s
ci / powerpc64le-linux-release (pull_request) Successful in 2h2m57s
ci / aarch64-linux-debug (pull_request) Successful in 2h43m36s
ci / x86_64-windows-debug (pull_request) Successful in 1h26m45s
ci / powerpc64le-linux-debug (pull_request) Successful in 3h56m1s
ci / x86_64-windows-release (pull_request) Successful in 1h6m38s
ci / x86_64-linux-debug (pull_request) Successful in 1h39m15s
ci / s390x-linux-release (pull_request) Successful in 1h32m4s
ci / s390x-linux-debug (pull_request) Successful in 2h47m58s
ci / x86_64-linux-release (pull_request) Successful in 3h42m13s
ci / x86_64-linux-debug-llvm (pull_request) Successful in 4h54m1s
ci / loongarch64-linux-release (pull_request) Successful in 2h34m32s
ci / loongarch64-linux-debug (pull_request) Successful in 3h40m28s
ci / aarch64-freebsd-debug (pull_request) Has been cancelled
ci / aarch64-freebsd-release (pull_request) Has been cancelled
ci / aarch64-netbsd-debug (pull_request) Has been cancelled
ci / aarch64-netbsd-release (pull_request) Has been cancelled
ci / riscv64-linux-debug (pull_request) Has been cancelled
ci / riscv64-linux-release (pull_request) Has been cancelled
to 4630728b26
All checks were successful
ci / x86_64-freebsd-release (pull_request) Successful in 49m0s
ci / x86_64-netbsd-release (pull_request) Successful in 50m41s
ci / x86_64-windows-release (pull_request) Successful in 54m4s
ci / x86_64-freebsd-debug (pull_request) Successful in 56m58s
ci / x86_64-openbsd-release (pull_request) Successful in 59m45s
ci / aarch64-macos-release (pull_request) Successful in 1h2m17s
ci / x86_64-netbsd-debug (pull_request) Successful in 1h9m1s
ci / x86_64-openbsd-debug (pull_request) Successful in 1h9m46s
ci / x86_64-linux-debug (pull_request) Successful in 1h16m48s
ci / aarch64-macos-debug (pull_request) Successful in 1h18m35s
ci / x86_64-windows-debug (pull_request) Successful in 1h20m43s
ci / aarch64-linux-release (pull_request) Successful in 1h37m58s
ci / s390x-linux-release (pull_request) Successful in 1h45m47s
ci / powerpc64le-linux-release (pull_request) Successful in 2h15m56s
ci / s390x-linux-debug (pull_request) Successful in 2h37m0s
ci / aarch64-linux-debug (pull_request) Successful in 3h10m39s
ci / x86_64-linux-debug-llvm (pull_request) Successful in 3h8m30s
ci / x86_64-linux-release (pull_request) Successful in 3h15m6s
ci / powerpc64le-linux-debug (pull_request) Successful in 4h51m50s
ci / loongarch64-linux-release (pull_request) Successful in 2h7m33s
ci / loongarch64-linux-debug (pull_request) Successful in 3h38m27s
ci / aarch64-freebsd-debug (pull_request) Has been skipped
ci / aarch64-freebsd-release (pull_request) Has been skipped
ci / aarch64-netbsd-debug (pull_request) Has been skipped
ci / aarch64-netbsd-release (pull_request) Has been skipped
ci / riscv64-linux-debug (pull_request) Has been skipped
ci / riscv64-linux-release (pull_request) Has been skipped
2026年07月10日 21:21:08 +02:00
Compare
Author
Contributor
Copy link

Updated PR with an error test case and message changes.

While implementing the test I discovered that there already was a ptr arithmetic check on vulkan. I've added OpenGL to the arithmetic check and moved the slice check further up. I'd argue that the new error should take precedence over the ptr arithmetic error (as the caller is not necessarily aware that they are trying to do ptr arithmetic).

I also noticed the test won't run if I specify the selfhosted backend runner. I put it as auto so that it runs, but I would think the correct thing to do is leave it as selfhosted and fix whatever isn't running the spirv + selfhosted tests

Updated PR with an error test case and message changes. While implementing the test I discovered that there already was a ptr arithmetic check on vulkan. I've added OpenGL to the arithmetic check and moved the slice check further up. I'd argue that the new error should take precedence over the ptr arithmetic error (as the caller is not necessarily aware that they are trying to do ptr arithmetic). I also noticed the test won't run if I specify the selfhosted backend runner. I put it as auto so that it runs, but I would think the correct thing to do is leave it as selfhosted and fix whatever isn't running the spirv + selfhosted tests
Sign in to join this conversation.
No reviewers
Labels
Clear labels
abi/f32
abi/ilp32
abi/sf
accepted
This proposal is planned.
arch/21k
arch/6502
arch/aarch64
arch/alpha
arch/amdgcn
arch/arc
arch/arc32
arch/arc64
arch/arm
arch/avr
arch/bfin
arch/bpf
arch/colossus
arch/cris
arch/csky
arch/dlx
arch/epiphany
arch/fr30
arch/frv
arch/hexagon
arch/hppa
arch/hppa64
arch/ia64
arch/kalimba
arch/kvx
arch/lanai
arch/lm32
arch/loongarch32
arch/loongarch64
arch/m32r
arch/m68k
arch/m88k
arch/mcore
arch/microblaze
arch/mips
arch/mips64
arch/mmix
arch/moxie
arch/mrisc32
arch/msp430
arch/nds32
arch/ns32k
arch/nvptx
arch/or1k
arch/powerpc
arch/powerpc64
arch/propeller
arch/riscv32
arch/riscv64
arch/rl78
arch/rx
arch/s390x
arch/sh
arch/sparc
arch/sparc64
arch/spirv
arch/spu
arch/tricore
arch/v850
arch/vax
arch/vc4
arch/ve
arch/wasm
arch/x86
arch/x86_64
arch/xcore
arch/xtensa
autodoc
The web application for interactive documentation and generation of its assets.
backend/c
The C backend outputs C source code.
backend/llvm
The LLVM backend outputs an LLVM bitcode module.
backend/self-hosted
The self-hosted backends produce machine code directly.
binutils
Zig's included binary utilities: zig ar, zig dlltool, zig lib, zig ranlib, zig objcopy, and zig rc.
breaking
Implementing this issue could cause existing code to no longer compile or have different behavior.
build system
The Zig build system - zig build, std.Build, the build runner, and package management.
debug info
An issue related to debug information (e.g. DWARF) produced by the Zig compiler.
docs
An issue with documentation, e.g. the language reference or standard library doc comments.
error message
This issue points out an error message that is unhelpful and should be improved.
frontend
Tokenization, parsing, AstGen, ZonGen, Sema, Legalize, and Liveness.
fuzzing
An issue related to Zig's integrated fuzz testing.
incremental
Reuse of internal compiler state for faster compilation.
lib/c
This issue relates to Zig's libc implementation and/or vendored libcs.
lib/compiler-rt
This issue relates to Zig's compiler-rt library.
lib/cxx
This issue relates to Zig's vendored libc++ and/or libc++abi.
lib/std
This issue relates to Zig's standard library.
lib/tsan
This issue relates to Zig's vendored libtsan.
lib/ubsan-rt
This issue relates to Zig's ubsan-rt library.
lib/unwind
This issue relates to Zig's vendored libunwind.
linking
Zig's integrated object file and incremental linker.
miscompilation
The compiler reports success but produces semantically incorrect code.
os/android
os/contiki
os/dragonfly
os/driverkit
os/emscripten
os/freebsd
os/fuchsia
os/haiku
os/hermit
os/hurd
os/illumos
os/ios
os/linux
os/maccatalyst
os/macos
os/managarm
os/netbsd
os/ohos
os/openbsd
os/plan9
os/redox
os/rtems
os/serenity
os/tvos
os/uefi
os/visionos
os/wasi
os/watchos
os/windows
proposal
This issue suggests language modifications. If it also has the "accepted" label then it is planned.
release notes
This issue or pull request should be mentioned in the release notes.
testing
This issue is related to testing the compiler, standard library, or other parts of Zig.
zig cc
Zig as a drop-in C-family compiler.
zig fmt
The Zig source code formatter.
zig reduce
The Zig source code reduction tool.
bounty
https://ziglang.org/news/announcing-donor-bounties
bug
Observed behavior contradicts documented or intended behavior.
contributor-friendly
This issue is limited in scope and/or knowledge of project internals.
downstream
An issue with a third-party project that uses this project.
enhancement
Solving this issue will likely involve adding new logic or components to the codebase.
infra
An issue related to project infrastructure, e.g. continuous integration.
optimization
A task to improve performance and/or resource usage.
question
No questions on the issue tracker; use a community space instead.
regression
Something that used to work in a previous version stopped working
upstream
An issue with a third-party project that this project uses.
use case
Describes a real use case that is difficult or impossible, but does not propose a solution.
No labels
abi/f32
abi/ilp32
abi/sf
accepted
arch/21k
arch/6502
arch/aarch64
arch/alpha
arch/amdgcn
arch/arc
arch/arc32
arch/arc64
arch/arm
arch/avr
arch/bfin
arch/bpf
arch/colossus
arch/cris
arch/csky
arch/dlx
arch/epiphany
arch/fr30
arch/frv
arch/hexagon
arch/hppa
arch/hppa64
arch/ia64
arch/kalimba
arch/kvx
arch/lanai
arch/lm32
arch/loongarch32
arch/loongarch64
arch/m32r
arch/m68k
arch/m88k
arch/mcore
arch/microblaze
arch/mips
arch/mips64
arch/mmix
arch/moxie
arch/mrisc32
arch/msp430
arch/nds32
arch/ns32k
arch/nvptx
arch/or1k
arch/powerpc
arch/powerpc64
arch/propeller
arch/riscv32
arch/riscv64
arch/rl78
arch/rx
arch/s390x
arch/sh
arch/sparc
arch/sparc64
arch/spirv
arch/spu
arch/tricore
arch/v850
arch/vax
arch/vc4
arch/ve
arch/wasm
arch/x86
arch/x86_64
arch/xcore
arch/xtensa
autodoc
backend/c
backend/llvm
backend/self-hosted
binutils
breaking
build system
debug info
docs
error message
frontend
fuzzing
incremental
lib/c
lib/compiler-rt
lib/cxx
lib/std
lib/tsan
lib/ubsan-rt
lib/unwind
linking
miscompilation
os/android
os/contiki
os/dragonfly
os/driverkit
os/emscripten
os/freebsd
os/fuchsia
os/haiku
os/hermit
os/hurd
os/illumos
os/ios
os/linux
os/maccatalyst
os/macos
os/managarm
os/netbsd
os/ohos
os/openbsd
os/plan9
os/redox
os/rtems
os/serenity
os/tvos
os/uefi
os/visionos
os/wasi
os/watchos
os/windows
proposal
release notes
testing
zig cc
zig fmt
zig reduce
bounty
bug
contributor-friendly
downstream
enhancement
infra
optimization
question
regression
upstream
use case
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
ziglang/zig!36118
Reference in a new issue
ziglang/zig
No description provided.
Delete branch "sphaerophoria/zig:ban-slices"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?