diff options
| -rw-r--r-- | llvm/lib/CodeGen/MachineBlockPlacement.cpp | 26 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/block-placement.ll | 102 |
2 files changed, 119 insertions, 9 deletions
diff --git a/llvm/lib/CodeGen/MachineBlockPlacement.cpp b/llvm/lib/CodeGen/MachineBlockPlacement.cpp index e4ddb043505..ae0f5bd0bcc 100644 --- a/llvm/lib/CodeGen/MachineBlockPlacement.cpp +++ b/llvm/lib/CodeGen/MachineBlockPlacement.cpp @@ -68,6 +68,13 @@ ExitBlockBias("block-placement-exit-block-bias", "over the original exit to be considered the new exit."), cl::init(0), cl::Hidden); +static cl::opt<bool> PlaceLastSuccessor( + "place-last-successor", + cl::desc("When selecting a non-successor block, choose the last block to " + "have been a successor. This represents the block whose " + "predecessor was most recently placed."), + cl::init(false), cl::Hidden); + static cl::opt<bool> OutlineOptionalBranches( "outline-optional-branches", cl::desc("Put completely optional branches, i.e. branches with a common " @@ -443,6 +450,25 @@ MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor( MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock( BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList, const BlockFilterSet *BlockFilter) { + if (PlaceLastSuccessor) { + // If we're just placing the last successor as the best candidate, the + // logic is super simple. We skip the already placed entries on the + // worklist and return the most recently added entry that isn't placed. + while (!WorkList.empty()) { + MachineBasicBlock *SuccBB = WorkList.pop_back_val(); + BlockChain &SuccChain = *BlockToChain.lookup(SuccBB); + if (&SuccChain == &Chain) { + DEBUG(dbgs() << " " << getBlockName(SuccBB) + << " -> Already merged!\n"); + continue; + } + assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block"); + return SuccBB; + } + + return nullptr; + } + // Once we need to walk the worklist looking for a candidate, cleanup the // worklist of already placed entries. // FIXME: If this shows up on profiles, it could be folded (at the cost of diff --git a/llvm/test/CodeGen/X86/block-placement.ll b/llvm/test/CodeGen/X86/block-placement.ll index e0276e42d4d..49ee1b2c162 100644 --- a/llvm/test/CodeGen/X86/block-placement.ll +++ b/llvm/test/CodeGen/X86/block-placement.ll @@ -1,4 +1,4 @@ -; RUN: llc -mtriple=i686-linux -pre-RA-sched=source < %s | FileCheck %s +; RUN: llc -mtriple=i686-linux -place-last-successor -pre-RA-sched=source < %s | FileCheck %s declare void @error(i32 %i, i32 %a, i32 %b) @@ -17,11 +17,11 @@ define i32 @test_ifchains(i32 %i, i32* %a, i32 %b) { ; CHECK: %else4 ; CHECK-NOT: .align ; CHECK: %exit -; CHECK: %then1 -; CHECK: %then2 -; CHECK: %then3 -; CHECK: %then4 ; CHECK: %then5 +; CHECK: %then4 +; CHECK: %then3 +; CHECK: %then2 +; CHECK: %then1 entry: %gep1 = getelementptr i32, i32* %a, i32 1 @@ -82,9 +82,9 @@ define i32 @test_loop_cold_blocks(i32 %i, i32* %a) { ; CHECK-LABEL: test_loop_cold_blocks: ; CHECK: %entry ; CHECK-NOT: .align -; CHECK: %unlikely1 -; CHECK-NOT: .align ; CHECK: %unlikely2 +; CHECK-NOT: .align +; CHECK: %unlikely1 ; CHECK: .align ; CHECK: %body1 ; CHECK: %body2 @@ -135,9 +135,9 @@ define i32 @test_loop_early_exits(i32 %i, i32* %a) { ; CHECK: %body3 ; CHECK: %body4 ; CHECK: %exit -; CHECK: %bail1 -; CHECK: %bail2 ; CHECK: %bail3 +; CHECK: %bail2 +; CHECK: %bail1 entry: br label %body1 @@ -1083,3 +1083,87 @@ exit: %ret = phi i32 [ %val1, %then ], [ %val2, %else ] ret i32 %ret } + +define void @test_outlined() { +; This test ends up with diamond control flow in outlined optional regions. +; These diamonds should still be locally cohensive even when out-of-line due to +; being cold. +; CHECK-LABEL: test_outlined: +; CHECK: %a1 +; CHECK: %a2 +; CHECK: %done +; CHECK: %b2 +; CHECK: %c2 +; CHECK: %d2 +; CHECK: %f2 +; CHECK: %b1 +; CHECK: %c1 +; CHECK: %d1 +; CHECK: %f1 + +a1: + %call.a1 = call i1 @a1() + br i1 %call.a1, label %b1, label %a2, !prof !0 + +b1: + %call.b1 = call i1 @b1() + br i1 %call.b1, label %c1, label %d1 + +c1: + call void @c1() + br label %f1 + +d1: + call void @d1() + br label %f1 + +f1: + call void @f1() + br label %a2 + +a2: + %call.a2 = call i1 @a2() + br i1 %call.a2, label %b2, label %done, !prof !0 + +b2: + %call.b2 = call i1 @b2() + br i1 %call.b2, label %c2, label %d2 + +c2: + call void @c2() + br label %f2 + +d2: + call void @d2() + br label %f2 + +f2: + call void @f2() + br label %done + +done: + call void @done() + ret void +} + +declare i1 @a1() + +declare i1 @b1() + +declare void @c1() + +declare void @d1() + +declare void @f1() + +declare i1 @a2() + +declare i1 @b2() + +declare void @c2() + +declare void @d2() + +declare void @f2() + +declare void @done() |
