[llvm] r211906 - R600: Don't crash on unhandled instruction in promote alloca

Matt Arsenault Matthew.Arsenault at amd.com
Fri Jun 27 09:52:49 PDT 2014


Author: arsenm
Date: Fri Jun 27 11:52:49 2014
New Revision: 211906

URL: http://llvm.org/viewvc/llvm-project?rev=211906&view=rev
Log:
R600: Don't crash on unhandled instruction in promote alloca

Added:
    llvm/trunk/test/CodeGen/R600/private-memory-atomics.ll
    llvm/trunk/test/CodeGen/R600/private-memory-broken.ll
Modified:
    llvm/trunk/lib/Target/R600/AMDGPUPromoteAlloca.cpp
    llvm/trunk/test/CodeGen/R600/private-memory.ll

Modified: llvm/trunk/lib/Target/R600/AMDGPUPromoteAlloca.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/AMDGPUPromoteAlloca.cpp?rev=211906&r1=211905&r2=211906&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/AMDGPUPromoteAlloca.cpp (original)
+++ llvm/trunk/lib/Target/R600/AMDGPUPromoteAlloca.cpp Fri Jun 27 11:52:49 2014
@@ -129,6 +129,22 @@ static Value* GEPToVectorIndex(GetElemen
   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 @@ static bool tryPromoteAllocaToVector(All
   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 @@ static bool tryPromoteAllocaToVector(All
 
     GEPVectorIdx[GEP] = Index;
     for (User *GEPUser : AllocaUser->users()) {
+      if (!canVectorizeInst(cast<Instruction>(GEPUser)))
+        return false;
+
       WorkList.push_back(GEPUser);
     }
   }
@@ -201,12 +223,12 @@ static bool tryPromoteAllocaToVector(All
       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;

Added: llvm/trunk/test/CodeGen/R600/private-memory-atomics.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/private-memory-atomics.ll?rev=211906&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/R600/private-memory-atomics.ll (added)
+++ llvm/trunk/test/CodeGen/R600/private-memory-atomics.ll Fri Jun 27 11:52:49 2014
@@ -0,0 +1,31 @@
+; RUN: llc -verify-machineinstrs -march=r600 -mcpu=SI < %s
+
+; This works because promote allocas pass replaces these with LDS atomics.
+
+; 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
+}

Added: llvm/trunk/test/CodeGen/R600/private-memory-broken.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/private-memory-broken.ll?rev=211906&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/R600/private-memory-broken.ll (added)
+++ llvm/trunk/test/CodeGen/R600/private-memory-broken.ll Fri Jun 27 11:52:49 2014
@@ -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
+}

Modified: llvm/trunk/test/CodeGen/R600/private-memory.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/private-memory.ll?rev=211906&r1=211905&r2=211906&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/R600/private-memory.ll (original)
+++ llvm/trunk/test/CodeGen/R600/private-memory.ll Fri Jun 27 11:52:49 2014
@@ -267,5 +267,19 @@ entry:
   %load = load i32* %gep2
   store i32 %load, i32 addrspace(1)* %out
   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
 }
+





More information about the llvm-commits mailing list