[ARM] Expose supportsTailCalls in ARMTargetTransformInfo #67932
Conversation
| @llvm/pr-subscribers-llvm-transforms @llvm/pr-subscribers-backend-arm ChangesThe ARM target doesn't support tail calls on certain older subarchitectures, however this information is not communicated to the target transform information. This causes coroutine symmetric transfer to fail as it adds musttail attributes which throw an assert during code generation. Exposing the underlying subtarget's Full diff: https://github.com/llvm/llvm-project/pull/67932.diff 1 Files Affected:
diff --git a/llvm/lib/Target/ARM/ARMTargetTransformInfo.h b/llvm/lib/Target/ARM/ARMTargetTransformInfo.h index bb4b321b5300916..95cd1442e689299 100644 --- a/llvm/lib/Target/ARM/ARMTargetTransformInfo.h +++ b/llvm/lib/Target/ARM/ARMTargetTransformInfo.h @@ -333,6 +333,14 @@ class ARMTTIImpl : public BasicTTIImplBase<ARMTTIImpl> { bool hasArmWideBranch(bool Thumb) const; + bool supportsTailCalls() const { + return ST->supportsTailCall(); + } + + bool supportsTailCallFor(const CallBase *CB) const { + return supportsTailCalls(); + } + /// @} }; |
| ✅ With the latest revision this PR passed the C/C++ code formatter. |
mincrmatt12 force-pushed the supports-tail-calls-tti branch from 13b1d74 to b2bced0 Compare 
| This sounds OK. Do you have a testcase for the change in coroutine behaviour? |
The ARM target doesn't support tail calls on certain older architectures, however this information is not communicated to the target transform information. This causes coroutine symmetric transfer to fail as it adds musttail attributes which throw an assert during code generation.
mincrmatt12 force-pushed the supports-tail-calls-tti branch from b2bced0 to 53783c9 Compare 
| Sorry about the large delay -- I've just added a test (that's almost identical to the existing one for similar behaviour on ppc64) to verify the musttail calls are not emitted on ARMv6 thumb. |
davemgreen left a comment
davemgreen left a comment There was a problem hiding this comment.
Thanks. The supportsTailCallFor could be more precise but that is likely not a v6m issue. What you have LGTM in that regard.
| @@ -0,0 +1,60 @@ | |||
| ; Test that ARMv6 correctly masks coroutine infrastructure from generating musttail calls | |||
| ; RUN: opt < %s -passes='cgscc(coro-split),simplifycfg,early-cse' -S \ | |||
| ; RUN: -mtriple=arm-none-eabi -mcpu=cortex-m0 | FileCheck %s | |||
There was a problem hiding this comment.
It might be worth having an extra check line for an arch that does support tail calls, to show the difference between the two
4 participants
The ARM target doesn't support tail calls on certain older subarchitectures, however this information is not communicated to the target transform information. This causes coroutine symmetric transfer to fail as it adds musttail attributes which throw an assert during code generation. Exposing the underlying subtarget's
supportsTailCallmethod to transforms (specifically the coroutine ones) allows coroutines to compile without crashing on these older subarchitectures (like ARMv6 on a Cortex-M0+)See also 0b5ead6 and d2d77e0.