[PATCH] R600: Don't crash on unhandled instruction in promote alloca

[Tom Stellard]

On Fri, Jun 27, 2014 at 04:33:21AM +0000, Matt Arsenault wrote:
> http://reviews.llvm.org/D4318

LGTM.

> 
> Files:
>   lib/Target/R600/AMDGPUPromoteAlloca.cpp
>   test/CodeGen/R600/private-memory-broken.ll
>   test/CodeGen/R600/private-memory.ll

> Index: lib/Target/R600/AMDGPUPromoteAlloca.cpp
> ===================================================================
> --- lib/Target/R600/AMDGPUPromoteAlloca.cpp
> +++ lib/Target/R600/AMDGPUPromoteAlloca.cpp
> @@ -129,6 +129,22 @@
>    return GEP->getOperand(2);
>  }
>  
> +// Not an instruction handled below to turn into a vector.
> +//
> +// TODO: Check isTriviallyVectorizable for calls and handle other
> +// instructions.
> +static bool canVectorizeInst(Instruction *Inst) {
> +  switch (Inst->getOpcode()) {
> +  case Instruction::Load:
> +  case Instruction::Store:
> +  case Instruction::BitCast:
> +  case Instruction::AddrSpaceCast:
> +    return true;
> +  default:
> +    return false;
> +  }
> +}
> +
>  static bool tryPromoteAllocaToVector(AllocaInst *Alloca) {
>    Type *AllocaTy = Alloca->getAllocatedType();
>  
> @@ -149,6 +165,9 @@
>    for (User *AllocaUser : Alloca->users()) {
>      GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(AllocaUser);
>      if (!GEP) {
> +      if (!canVectorizeInst(cast<Instruction>(AllocaUser)))
> +        return false;
> +
>        WorkList.push_back(AllocaUser);
>        continue;
>      }
> @@ -164,6 +183,9 @@
>  
>      GEPVectorIdx[GEP] = Index;
>      for (User *GEPUser : AllocaUser->users()) {
> +      if (!canVectorizeInst(cast<Instruction>(GEPUser)))
> +        return false;
> +
>        WorkList.push_back(GEPUser);
>      }
>    }
> @@ -201,12 +223,12 @@
>        break;
>      }
>      case Instruction::BitCast:
> +    case Instruction::AddrSpaceCast:
>        break;
>  
>      default:
>        Inst->dump();
> -      llvm_unreachable("Do not know how to replace this instruction "
> -                       "with vector op");
> +      llvm_unreachable("Inconsistency in instructions promotable to vector");
>      }
>    }
>    return true;
> Index: test/CodeGen/R600/private-memory-broken.ll
> ===================================================================
> --- /dev/null
> +++ test/CodeGen/R600/private-memory-broken.ll
> @@ -0,0 +1,20 @@
> +; RUN: not llc -verify-machineinstrs -march=r600 -mcpu=SI %s 2>&1 | FileCheck %s
> +
> +; Make sure promote alloca pass doesn't crash
> +
> +; CHECK: unsupported call
> +
> +declare i32 @foo(i32*) nounwind
> +
> +define void @call_private(i32 addrspace(1)* %out, i32 %in) nounwind {
> +entry:
> +  %tmp = alloca [2 x i32]
> +  %tmp1 = getelementptr [2 x i32]* %tmp, i32 0, i32 0
> +  %tmp2 = getelementptr [2 x i32]* %tmp, i32 0, i32 1
> +  store i32 0, i32* %tmp1
> +  store i32 1, i32* %tmp2
> +  %tmp3 = getelementptr [2 x i32]* %tmp, i32 0, i32 %in
> +  %val = call i32 @foo(i32* %tmp3) nounwind
> +  store i32 %val, i32 addrspace(1)* %out
> +  ret void
> +}
> Index: test/CodeGen/R600/private-memory.ll
> ===================================================================
> --- test/CodeGen/R600/private-memory.ll
> +++ test/CodeGen/R600/private-memory.ll
> @@ -269,3 +269,45 @@
>    ret void
>  
>  }
> +
> +define void @select_private(i32 addrspace(1)* %out, i32 %in) nounwind {
> +entry:
> +  %tmp = alloca [2 x i32]
> +  %tmp1 = getelementptr [2 x i32]* %tmp, i32 0, i32 0
> +  %tmp2 = getelementptr [2 x i32]* %tmp, i32 0, i32 1
> +  store i32 0, i32* %tmp1
> +  store i32 1, i32* %tmp2
> +  %cmp = icmp eq i32 %in, 0
> +  %sel = select i1 %cmp, i32* %tmp1, i32* %tmp2
> +  %load = load i32* %sel
> +  store i32 %load, i32 addrspace(1)* %out
> +  ret void
> +}
> +
> +; Private atomics have no real use, but at least shouldn't crash on it.
> +define void @atomicrmw_private(i32 addrspace(1)* %out, i32 %in) nounwind {
> +entry:
> +  %tmp = alloca [2 x i32]
> +  %tmp1 = getelementptr [2 x i32]* %tmp, i32 0, i32 0
> +  %tmp2 = getelementptr [2 x i32]* %tmp, i32 0, i32 1
> +  store i32 0, i32* %tmp1
> +  store i32 1, i32* %tmp2
> +  %tmp3 = getelementptr [2 x i32]* %tmp, i32 0, i32 %in
> +  %tmp4 = atomicrmw add i32* %tmp3, i32 7 acq_rel
> +  store i32 %tmp4, i32 addrspace(1)* %out
> +  ret void
> +}
> +
> +define void @cmpxchg_private(i32 addrspace(1)* %out, i32 %in) nounwind {
> +entry:
> +  %tmp = alloca [2 x i32]
> +  %tmp1 = getelementptr [2 x i32]* %tmp, i32 0, i32 0
> +  %tmp2 = getelementptr [2 x i32]* %tmp, i32 0, i32 1
> +  store i32 0, i32* %tmp1
> +  store i32 1, i32* %tmp2
> +  %tmp3 = getelementptr [2 x i32]* %tmp, i32 0, i32 %in
> +  %tmp4 = cmpxchg i32* %tmp3, i32 0, i32 1 acq_rel monotonic
> +  %val = extractvalue { i32, i1 } %tmp4, 0
> +  store i32 %val, i32 addrspace(1)* %out
> +  ret void
> +}

> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits