[LoopUnroll][RISCV] Crash when unrolling loop containing vector instructions

[frasercrmck]

When optimizing for RISC-V without the vector extension enabled, the RISCVTargetTransformInfo returns an invalid cost and the loop unroller asserts.

This is happening on both tip (64422cf) and on LLVM 18.1.0 RC4 (461274b).

; RUN: opt -mattr=+f,+d -S --passes=loop-unroll-full -S -disable-output
target datalayout = "e-m:e-p:64:64-i64:64-i128:128-n32:64-S128"
target triple = "riscv64-unknown-unknown-elf"

define void @foo() {
entry:
  br label %for.body

for.body:                                         ; preds = %for.body, %entry
  %indvars.iv1 = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
  %0 = load float, ptr null, align 4
  %splat.splat.i.i.i = shufflevector <2 x float> zeroinitializer, <2 x float> zeroinitializer, <2 x i32> zeroinitializer
  %cmp1.i.i.i = fcmp ogt <2 x float> zeroinitializer, zeroinitializer
  %splat.splat3.i.i.i = shufflevector <2 x i32> zeroinitializer, <2 x i32> zeroinitializer, <2 x i32> zeroinitializer
  %xor3.i.i.i.i.i = select <2 x i1> zeroinitializer, <2 x i32> zeroinitializer, <2 x i32> zeroinitializer
  %1 = load float, ptr null, align 4
  %splat.splat8.i.i.i = shufflevector <2 x float> zeroinitializer, <2 x float> zeroinitializer, <2 x i32> zeroinitializer
  %sub.i.i.i = fsub <2 x float> zeroinitializer, zeroinitializer
  %mul.i.i.i = shl i64 0, 0
  %2 = load float, ptr null, align 4
  %splat.splat.i.i.i.i = shufflevector <2 x float> zeroinitializer, <2 x float> zeroinitializer, <2 x i32> zeroinitializer
  %xor3.i.i.i.v.i.i = select <2 x i1> zeroinitializer, <2 x float> zeroinitializer, <2 x float> zeroinitializer
  %indvars.iv.next = add i64 %indvars.iv1, 1
  %exitcond = icmp ne i64 %indvars.iv1, 8
  br i1 %exitcond, label %for.body, label %exit

exit:                                             ; preds = %for.body
  ret void
}

Should show:

opt: /llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp:680: std::optional<EstimatedUnrollCost> analyzeLoopUnrollCost(const llvm::Loop *, unsigned int, llvm::DominatorTree &, llvm::ScalarEvolution &, const SmallPtrSetImpl<const llvm::Value *> &, const llvm::TargetTransformInfo &, unsigned int, unsigned int): Assertion `UnrolledCost.isValid() && RolledDynamicCost.isValid() && "All instructions must have a valid cost, whether the " "loop is rolled or unrolled."' failed.

This is a regression against LLVM 17, which reports (with the same input but with --debug):

Loop Unroll: F[foo] Loop %for.body
  Loop Size = Invalid
  Not unrolling loop which contains instructions with invalid cost.

I've been away from LLVM for a while so don't know if this is a failure in the RISCV cost model or an overly restrictive loop unroller.

It looks like with LLVM 17, the CodeSize costs were deemed to be Invalid by the UnrollCostEstimator, so the unroller exited early. Now since LLVM 18 they are no longer Invalid so the unroller tries to calculate profitability in analyzeLoopUnrollCost which can't handle Invalid costs and asserts.

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Fraser Cormack (frasercrmck)

When optimizing for RISC-V _without_ the vector extension enabled, the `RISCVTargetTransformInfo` returns an invalid cost and the loop unroller asserts.

This is happening on both tip (64422cf) and on LLVM 18.1.0 RC4 (461274b).

; RUN: opt -mattr=+f,+d -S --passes=loop-unroll-full -S -disable-output
target datalayout = "e-m:e-p:64:64-i64:64-i128:128-n32:64-S128"
target triple = "riscv64-unknown-unknown-elf"

define void @<!-- -->foo() {
entry:
  br label %for.body

for.body:                                         ; preds = %for.body, %entry
  %indvars.iv1 = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
  %0 = load float, ptr null, align 4
  %splat.splat.i.i.i = shufflevector &lt;2 x float&gt; zeroinitializer, &lt;2 x float&gt; zeroinitializer, &lt;2 x i32&gt; zeroinitializer
  %cmp1.i.i.i = fcmp ogt &lt;2 x float&gt; zeroinitializer, zeroinitializer
  %splat.splat3.i.i.i = shufflevector &lt;2 x i32&gt; zeroinitializer, &lt;2 x i32&gt; zeroinitializer, &lt;2 x i32&gt; zeroinitializer
  %xor3.i.i.i.i.i = select &lt;2 x i1&gt; zeroinitializer, &lt;2 x i32&gt; zeroinitializer, &lt;2 x i32&gt; zeroinitializer
  %1 = load float, ptr null, align 4
  %splat.splat8.i.i.i = shufflevector &lt;2 x float&gt; zeroinitializer, &lt;2 x float&gt; zeroinitializer, &lt;2 x i32&gt; zeroinitializer
  %sub.i.i.i = fsub &lt;2 x float&gt; zeroinitializer, zeroinitializer
  %mul.i.i.i = shl i64 0, 0
  %2 = load float, ptr null, align 4
  %splat.splat.i.i.i.i = shufflevector &lt;2 x float&gt; zeroinitializer, &lt;2 x float&gt; zeroinitializer, &lt;2 x i32&gt; zeroinitializer
  %xor3.i.i.i.v.i.i = select &lt;2 x i1&gt; zeroinitializer, &lt;2 x float&gt; zeroinitializer, &lt;2 x float&gt; zeroinitializer
  %indvars.iv.next = add i64 %indvars.iv1, 1
  %exitcond = icmp ne i64 %indvars.iv1, 8
  br i1 %exitcond, label %for.body, label %exit

exit:                                             ; preds = %for.body
  ret void
}

Should show:

opt: /llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp:680: std::optional<EstimatedUnrollCost> analyzeLoopUnrollCost(const llvm::Loop *, unsigned int, llvm::DominatorTree &, llvm::ScalarEvolution &, const SmallPtrSetImpl<const llvm::Value *> &, const llvm::TargetTransformInfo &, unsigned int, unsigned int): Assertion `UnrolledCost.isValid() && RolledDynamicCost.isValid() && "All instructions must have a valid cost, whether the " "loop is rolled or unrolled."' failed.

This is a regression against LLVM 17, which reports (with the same input but with --debug):

Loop Unroll: F[foo] Loop %for.body
  Loop Size = Invalid
  Not unrolling loop which contains instructions with invalid cost.

I've been away from LLVM for a while so don't know if this is a failure in the RISCV cost model or an overly restrictive loop unroller.

It looks like with LLVM 17, the CodeSize costs were deemed to be Invalid by the UnrollCostEstimator, so the unroller exited early. Now since LLVM 18 they are no longer Invalid so the unroller tries to calculate profitability in analyzeLoopUnrollCost which can't handle Invalid costs and asserts.

[topperc]

CC @arcbbb

[arcbbb]

Thanks! I created #83384 to address this.

[frasercrmck]

/cherry-pick fb67dce

[llvmbot]

Failed to create pull request for issue83294 https://github.com/llvm/llvm-project/actions/runs/8133150600

[llvmbot]

Failed to cherry-pick: fb67dce

https://github.com/llvm/llvm-project/actions/runs/8133152130

Please manually backport the fix and push it to your github fork. Once this is done, please create a pull request