-
Notifications
You must be signed in to change notification settings - Fork 12.2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[CodeGen][OOM] Switch statements without default branching result in long compile times or OOM #78578
Comments
…tchs, r=<try> Replace the default branch with an unreachable branch If it is the last variant Fixes rust-lang#119520. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
…tchs, r=oli-obk Replace the default branch with an unreachable branch If it is the last variant Fixes rust-lang#119520. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
…witchs, r=oli-obk Replace the default branch with an unreachable branch If it is the last variant Fixes rust-lang#119520. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
…tchs, r=oli-obk Replace the default branch with an unreachable branch If it is the last variant Fixes rust-lang#119520. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
…tchs, r=<try> Replace the default branch with an unreachable branch If it is the last variant Fixes rust-lang#119520. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
The current known IR includes two types. I conducted compile-time and runtime performance analysis based on these two IRs, without considering code size in this analysis. First, here are the characteristics of these two IRs:
Time AnalysisFirstly, in
In
Due to the ease of modifying
Based on the data above, the relationship between compile time and the number of branches appears to be The less drastic increase in compilation time for I believe the main issue causing this increased time is the duplication of critical BBs. A critical BB refers to one with multiple predecessors and multiple successors. In Runtime PerformancePer this comment, I assume that In What to Do?Backend isn't my suit, and I have zero knowledge of register allocation/instruction selection algorithms. :) My suggestion is to refrain from duplicating critical BBs. However, I cannot restrict the duplication to only BBs with one predecessor or one successor, as such duplication is clearly beneficial. |
…tchs, r=<try> Replace the default branch with an unreachable branch If it is the last variant Fixes rust-lang#119520. Fixes rust-lang#110097. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
…tchs, r=<try> Replace the default branch with an unreachable branch If it is the last variant Fixes rust-lang#119520. Fixes rust-lang#110097. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
…tchs, r=<try> Replace the default branch with an unreachable branch If it is the last variant Fixes rust-lang#119520. Fixes rust-lang#110097. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
…tchs, r=<try> Replace the default branch with an unreachable branch If it is the last variant Fixes rust-lang#119520. Fixes rust-lang#110097. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
…tchs, r=oli-obk Replace the default branch with an unreachable branch If it is the last variant Fixes rust-lang#119520. Fixes rust-lang#110097. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
…li-obk Replace the default branch with an unreachable branch If it is the last variant Fixes #119520. Fixes #110097. LLVM currently has limited ability to eliminate dead branches in switches, even with the patch of llvm/llvm-project#73446. The main reasons are as follows: - Additional costs are required to calculate the range of values, and there exist many scenarios that cannot be analyzed accurately. - Matching values by bitwise calculation cannot handle odd branches, nor can it handle values like `-1, 0, 1`. See [SimplifyCFG.cpp#L5424](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/Transforms/Utils/SimplifyCFG.cpp#L5424) and https://llvm.godbolt.org/z/qYMqhvMa8 - The current range information is continuous, even if the metadata for the range is submitted. See [ConstantRange.cpp#L1869-L1870](https://github.com/llvm/llvm-project/blob/llvmorg-17.0.6/llvm/lib/IR/ConstantRange.cpp#L1869-L1870). - The metadata of the range may be lost in passes such as SROA. See https://rust.godbolt.org/z/e7f87vKMK. Although we can make improvements, I think it would be more appropriate to put this issue to rustc first. After all, we can easily know the possible values. Note that we've currently found a slow compilation problem in the presence of unreachable branches. See llvm/llvm-project#78578. r? compiler
…tail duplicating blocks (#78582) Fixes #78578. Duplicating a BB which has both multiple predecessors and successors will result in a complex CFG and also may cause huge amount of PHI nodes. See #78578 (comment) for a detailed description of the limit.
I tried the following code, which takes more than 3s to compile using clang 17.0.6. (The code size is also a bit hard to accept.)
oom_manual2.c
If I change the 127 branch to the default branch, the time returns to normal.
The complete command I used is as follows:
main.c
I've found that the main time-consumption is
Early Tail Duplication
.By adding the new
TimeTraceScope
, I can see thatTailDuplicator::tailDuplicateAndUpdate
andTailDuplicator::shouldTailDuplicate
take up most of the time.TimeTraceScope
The original question is as follows (I've abbreviated it a bit)
oom_manual.cc.txt
Originally posted by @dwblaikie in #76669 (comment)
cc @aeubanks @joanahalili @alexfh
The text was updated successfully, but these errors were encountered: