This PR is to thoroughly rework duplicate tracker implementation and
tracking of IR entities and types. These are legacy parts of the project
resulting in an extremely bloated intermediate representation and
computational delays due to inefficient data flow and structure choices.
Main results of the rework:
1) Improved compile-time performance. The reference binary LLVM IR used
to measure speed gains in
https://github.com/llvm/llvm-project/pull/120415 shows ~x5 speed up also
after this PR. The timing before this PR is ~42s and after this PR it's
~7.5s. In total this PR and the previous overhaul of the module analysis
in https://github.com/llvm/llvm-project/pull/120415 results in ~x25
speed improvement.
```
$ time llc -O0 -mtriple=spirv64v1.6-unknown-unknown _group_barrier_phi.bc -o 1 --filetype=obj
real 0m7.545s
user 0m6.685s
sys 0m0.859s
```
2) Less bloated intermediate representation of internal translation
steps. Elimination of `spv_track_constant` intrinsic usage for scalar
constants, rework of `spv_assign_name`, removal of the gMIR `GET_XXX`
pseudo code and a smaller number of generated `ASSIGN_TYPE` pseudo codes
substantially decrease volume of data generated during translation.
3) Simpler code and easier maintenance. The duplicate tracker
implementation is simplified, as well as other features.
4) Numerous fixes of issues and logical flaws in different passes. The
main achievement is rework of the duplicate tracker itself that had
never guaranteed a correct caching of LLVM IR entities, rarely and
randomly returning stale/incorrect records (like, remove an instruction
from gMIR but still refer to it). Other fixes comprise consistent
generation of OpConstantNull, assigning types to newly created
registers, creation of integer/bool types, and other minor fixes.
5) Numerous fixes of LIT tests: mainly CHECK-DAG to properly reflect
SPIR-V spec guarantees, `{{$}}` at the end of constants to avoid
matching of substrings, and XFAILS for `SPV_INTEL_long_composites` test
cases, because the feature is not completed in full yet and doesn't
generate a requested by the extension sequence of instructions.
6) New test cases are added.
This PR is to address legacy issues with module analysis that currently
uses a complicated and not so efficient approach to trace dependencies
between SPIR-V id's via a duplicate tracker data structures and an
explicitly built dependency graph. Even a quick performance check
without any specialized benchmarks points to this part of the
implementation as a biggest bottleneck.
This PR specifically:
* eliminates a need to build a dependency graph as a data structure,
* updates the test suite (mainly, by fixing incorrect CHECK's referring
to a hardcoded order of definitions, contradicting the spec requirement
to allow certain definitions to go "in any order", see
https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_logical_layout_of_a_module),
* improves function pointers implementation so that it now passes
EXPENSIVE_CHECKS (thus removing 3 XFAIL's in the test suite).
As a quick sanity check of whether goals of the PR are achieved, we can
measure time of translation for any big LLVM IR. While testing the PR in
the local development environment, improvements of the x5 order have
been observed.
For example, the SYCL test case "group barrier" that is a ~1Mb binary IR
input shows the following values of the naive performance metric that we
can nevertheless apply here to roughly estimate effects of the PR.
before the PR:
```
$ time llc -O0 -mtriple=spirv64v1.6-unknown-unknown _group_barrier_phi.bc -o 1 --filetype=obj
real 3m33.241s
user 3m14.688s
sys 0m18.530s
```
after the PR
```
$ time llc -O0 -mtriple=spirv64v1.6-unknown-unknown _group_barrier_phi.bc -o 1 --filetype=obj
real 0m42.031s
user 0m38.834s
sys 0m3.193s
```
Next work should probably address Duplicate Tracker further, as it needs
analysis now from the perspective of what parts of it are not necessary
now, after changing the approach to implementation of the module
analysis step.
This PR contains changes in virtual register processing aimed to improve
correctness of emitted MIR between passes from the perspective of
MachineVerifier. This potentially helps to detect previously missed
flaws in code emission and harden the test suite. As a measure of
correctness and usefulness of this PR we may use a mode with expensive
checks set on, and MachineVerifier reports problems in the test suite.
In order to satisfy Machine Verifier requirements to MIR correctness not
only a rework of usage of virtual registers' types and classes is
required, but also corrections into pre-legalizer and instruction
selection logics. Namely, the following changes are introduced:
* scalar virtual registers have proper bit width,
* detect register class by SPIR-V type,
* add a superclass for id virtual register classes,
* fix Tablegen rules used for instruction selection,
* fixes of minor existed issues (missed flag for proper representation
of a null constant for OpenCL vs. HLSL, wrong usage of integer virtual
registers as a synonym of any non-type virtual register).
This PR addresses an issue that may arise when an integer argument size
differs from a machine word size for the target in a call to llvm
intrinsic. The following example demonstrates the issue:
```
@__const.test.arr = private unnamed_addr addrspace(2) constant [3 x i32] [i32 1, i32 2, i32 3]
define spir_func void @test() {
entry:
%arr = alloca [3 x i32], align 4
%dest = bitcast ptr %arr to ptr
call void @llvm.memcpy.p0.p2.i32(ptr align 4 %dest, ptr addrspace(2) align 4 @__const.test.arr, i32 1024, i1 false)
ret void
}
declare void @llvm.memcpy.p0.p2.i32(ptr nocapture writeonly, ptr addrspace(2) nocapture readonly, i32, i1)
```
Depending on the target this code may work or may fail without this PR
due to the fact that IR Translation step introduces additional `zext`
when type of the 3rd argument of `@llvm.memcpy.p0.p2.i32` differs from
machine word.
This PR addresses the issue by adding type deduction for a newly
inserted G_ZEXT generic opcode.
Add support to generate valid SPIR-V for the WaveGetLaneIndex() HLSL
builtin.
To implement this, I had to fix a few small issues in the backend, like
the i8* pointer type being emitted, even if we have the type information
elsewhere.
Signed-off-by: Nathan Gauër <brioche@google.com>
This pull request aims to remove any dependency on OpenCL/SPIR-V type
information in LLVM IR metadata. While, using metadata might simplify
and prettify the resulting SPIR-V output (and restore some of the
information missed in the transformation to opaque pointers), the
overall methodology for resolving kernel parameter types is highly
inefficient.
The high-level strategy is to assign kernel parameter types in this order:
1. Resolving the types using builtin function calls as mangled names
must contain type information or by looking up builtin definition in
SPIRVBuiltins.td. Then:
- Assigning the type temporarily using an intrinsic and later setting
the right SPIR-V type in SPIRVGlobalRegistry after IRTranslation
- Inserting a bitcast
2. Defaulting to LLVM IR types (in case of pointers the generic i8*
type or types from byval/byref attributes)
In case of type incompatibility (e.g. parameter defined initially as
sampler_t and later used as image_t) the error will be found early on
before IRTranslation (in the SPIRVEmitIntrinsics pass).
This patch introduces a new spv_ptrcast intrinsic for tracking expected
pointer types. The change fixes multiple OpenCL CTS regressions due the
switch to opaque pointers (e.g. basic/hiloeo).
