[Arm] Control forced unrolling of small loops

[VladiKrapp-Arm]

• Add flag to control cost threshold for forced unrolling of loops. Existing value preserved as default.
• Document flag in release notes, along with equivalent Aarch64 flag.

[llvmbot]

@llvm/pr-subscribers-backend-arm

Author: Vladi Krapp (VladiKrapp-Arm)

Changes

• Add flag to control cost threshold for forced unrolling of loops. Existing value preserved as default.
• Document flag in release notes, along with equivalent Aarch64 flag.

---

Full diff: https://github.com/llvm/llvm-project/pull/170127.diff

2 Files Affected:

• (modified) llvm/docs/ReleaseNotes.md (+4)
• (modified) llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp (+5-1)

diff --git a/llvm/docs/ReleaseNotes.md b/llvm/docs/ReleaseNotes.md index c6c527d1ae964..1ec700047ced5 100644 --- a/llvm/docs/ReleaseNotes.md +++ b/llvm/docs/ReleaseNotes.md @@ -111,12 +111,16 @@ Changes to the AArch64 Backend * `FEAT_TME` support has been removed, as it has been withdrawn from all future versions of the A-profile architecture. +* Allow forced unrolling of small loops. Threshold controlled by flag `aarch64-force-unroll-threshold` + Changes to the AMDGPU Backend ----------------------------- Changes to the ARM Backend -------------------------- +* Allow forced unrolling of small loops. Threshold controlled by flag `arm-force-unroll-threshold` (existing default preserved) + Changes to the AVR Backend -------------------------- diff --git a/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp b/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp index fdb0ec40cb41f..b5749bf51ce08 100644 --- a/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp +++ b/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp @@ -66,6 +66,10 @@ extern cl::opt<bool> EnableMaskedGatherScatters; extern cl::opt<unsigned> MVEMaxSupportedInterleaveFactor; +static cl::opt<int> ArmForceUnrollThreshold( + "arm-force-unroll-threshold", cl::init(12), cl::Hidden, + cl::desc("Threshold for forced unrolling of small loops in Arm architecture")); + /// Convert a vector load intrinsic into a simple llvm load instruction. /// This is beneficial when the underlying object being addressed comes /// from a constant, since we get constant-folding for free. @@ -2731,7 +2735,7 @@ void ARMTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE, // Force unrolling small loops can be very useful because of the branch // taken cost of the backedge. - if (Cost < 12) + if (Cost < ArmForceUnrollThreshold) UP.Force = true; } 

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[jthackray]

LGTM

[davemgreen]

OK thanks. LGTM