diff options
author | Dan Gohman <gohman@apple.com> | 2009-08-18 15:18:18 +0000 |
---|---|---|
committer | Dan Gohman <gohman@apple.com> | 2009-08-18 15:18:18 +0000 |
commit | da65822cfc938594f8fb7840947c1eb77e057a48 (patch) | |
tree | 825ce41886300ad659c1105bd376e0663c276488 | |
parent | 6874a2ae033b7b5e1d0c10714e01d9c87480956a (diff) | |
download | external_llvm-da65822cfc938594f8fb7840947c1eb77e057a48.zip external_llvm-da65822cfc938594f8fb7840947c1eb77e057a48.tar.gz external_llvm-da65822cfc938594f8fb7840947c1eb77e057a48.tar.bz2 |
Make tail merging handle blocks with repeated predecessors correctly, and
remove RemoveDuplicateSuccessor, as it is no longer necessary, and because
it breaks assumptions made in
MachineBasicBlock::isOnlyReachableByFallthrough.
Convert test/CodeGen/X86/omit-label.ll to FileCheck and add a testcase
for PR4732.
test/CodeGen/Thumb2/thumb2-ifcvt2.ll sees a diff with this commit due to
it being bugpoint-reduced to the point where it doesn't matter what the
condition for the branch is.
Add some more interesting code to
test/CodeGen/X86/2009-08-06-branchfolder-crash.ll, which is the testcase
that originally motivated the RemoveDuplicateSuccessor code, to help
verify that the original problem isn't being re-broken.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@79338 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r-- | lib/CodeGen/BranchFolding.cpp | 29 | ||||
-rw-r--r-- | test/CodeGen/Thumb2/thumb2-ifcvt2.ll | 2 | ||||
-rw-r--r-- | test/CodeGen/X86/2009-08-06-branchfolder-crash.ll | 52 | ||||
-rw-r--r-- | test/CodeGen/X86/omit-label.ll | 38 |
4 files changed, 93 insertions, 28 deletions
diff --git a/lib/CodeGen/BranchFolding.cpp b/lib/CodeGen/BranchFolding.cpp index 1ab3df2..1ce83f8 100644 --- a/lib/CodeGen/BranchFolding.cpp +++ b/lib/CodeGen/BranchFolding.cpp @@ -700,6 +700,7 @@ bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { + SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; MachineBasicBlock *IBB = I; MachineBasicBlock *PredBB = prior(I); MergePotentials.clear(); @@ -710,6 +711,9 @@ bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { // Skip blocks that loop to themselves, can't tail merge these. if (PBB==IBB) continue; + // Visit each predecessor only once. + if (!UniquePreds.insert(PBB)) + continue; MachineBasicBlock *TBB = 0, *FBB = 0; SmallVector<MachineOperand, 4> Cond; if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { @@ -850,27 +854,6 @@ bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); } -/// RemoveDuplicateSuccessor - make sure block Pred has at most one -/// successor edge leading to Succ. This is only called in one place, -/// but Chris prefers that it be a separate function. -static void RemoveDuplicateSuccessor(MachineBasicBlock *Pred, - MachineBasicBlock *Succ) { - MachineBasicBlock::succ_iterator SI = Pred->succ_begin(); - bool found = false; - while (SI != Pred->succ_end()) { - if (*SI == Succ) { - if (!found) { - found = true; - ++SI; - } else { - SI = Pred->removeSuccessor(SI); - } - } else { - ++SI; - } - } -} - /// IsBetterFallthrough - Return true if it would be clearly better to /// fall-through to MBB1 than to fall through into MBB2. This has to return /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will @@ -914,10 +897,6 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { while (!MBB->pred_empty()) { MachineBasicBlock *Pred = *(MBB->pred_end()-1); Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); - // If this resulted in a predecessor with true and false edges - // both going to the fallthrough block, clean up; - // BranchFolding doesn't like this. - RemoveDuplicateSuccessor(Pred, FallThrough); } // If MBB was the target of a jump table, update jump tables to go to the // fallthrough instead. diff --git a/test/CodeGen/Thumb2/thumb2-ifcvt2.ll b/test/CodeGen/Thumb2/thumb2-ifcvt2.ll index 15ccb2d..299ab4d 100644 --- a/test/CodeGen/Thumb2/thumb2-ifcvt2.ll +++ b/test/CodeGen/Thumb2/thumb2-ifcvt2.ll @@ -82,7 +82,7 @@ define fastcc void @t2() nounwind { entry: ; CHECK: t2: ; CHECK: cmp r0, #0 -; CHECK: bne.n +; CHECK: beq.n br i1 undef, label %bb.i.i3, label %growMapping.exit bb.i.i3: ; preds = %entry diff --git a/test/CodeGen/X86/2009-08-06-branchfolder-crash.ll b/test/CodeGen/X86/2009-08-06-branchfolder-crash.ll index 5e58725..2c8cf0c 100644 --- a/test/CodeGen/X86/2009-08-06-branchfolder-crash.ll +++ b/test/CodeGen/X86/2009-08-06-branchfolder-crash.ll @@ -88,3 +88,55 @@ for.inc: ; preds = %for.inc, %lor.end.i, %lor.rhs.i, %land.lhs.true3.i } declare i32 @safe() + +define i32 @func_35(i8 signext %p_35) nounwind readonly { +entry: + %tobool = icmp eq i8 %p_35, 0 ; <i1> [#uses=1] + br i1 %tobool, label %lor.lhs.false, label %if.then + +lor.lhs.false: ; preds = %entry + %tmp1 = load i8* @g_3 ; <i8> [#uses=1] + %tobool3 = icmp eq i8 %tmp1, 0 ; <i1> [#uses=1] + br i1 %tobool3, label %return, label %if.then + +if.then: ; preds = %lor.lhs.false, %entry + %tmp4 = load i8* @g_3 ; <i8> [#uses=1] + %conv5 = sext i8 %tmp4 to i32 ; <i32> [#uses=1] + ret i32 %conv5 + +return: ; preds = %lor.lhs.false + ret i32 0 +} + +define void @bar(i32 %p_5) noreturn nounwind { +entry: + %cmp = icmp sgt i32 %p_5, 0 ; <i1> [#uses=2] + %call = tail call i32 @safe() nounwind ; <i32> [#uses=1] + %conv1 = trunc i32 %call to i8 ; <i8> [#uses=3] + %tobool.i = xor i1 %cmp, true ; <i1> [#uses=3] + %cmp.i = icmp sgt i8 %conv1, 0 ; <i1> [#uses=3] + %or.cond.i = or i1 %cmp.i, %tobool.i ; <i1> [#uses=1] + br i1 %or.cond.i, label %lor.rhs.i, label %land.lhs.true3.i + +land.lhs.true3.i: ; preds = %entry + %xor = zext i1 %cmp to i32 ; <i32> [#uses=1] + %conv5.i = sext i8 %conv1 to i32 ; <i32> [#uses=1] + %cmp7.i = icmp slt i32 %conv5.i, %xor ; <i1> [#uses=1] + %cmp7.i.not = xor i1 %cmp7.i, true ; <i1> [#uses=1] + %or.cond23.i = and i1 %cmp.i, %tobool.i ; <i1> [#uses=1] + %or.cond = and i1 %cmp7.i.not, %or.cond23.i ; <i1> [#uses=1] + br i1 %or.cond, label %lor.end.i, label %for.inc + +lor.rhs.i: ; preds = %entry + %or.cond23.i.old = and i1 %cmp.i, %tobool.i ; <i1> [#uses=1] + br i1 %or.cond23.i.old, label %lor.end.i, label %for.inc + +lor.end.i: ; preds = %lor.rhs.i, %land.lhs.true3.i + %tobool19.i = icmp eq i8 %conv1, 0 ; <i1> [#uses=0] + br label %for.inc + +for.inc: ; preds = %for.inc, %lor.end.i, %lor.rhs.i, %land.lhs.true3.i + br label %for.inc +} + +declare i32 @safe() diff --git a/test/CodeGen/X86/omit-label.ll b/test/CodeGen/X86/omit-label.ll index 457b66b..6450f95 100644 --- a/test/CodeGen/X86/omit-label.ll +++ b/test/CodeGen/X86/omit-label.ll @@ -1,7 +1,11 @@ -; RUN: llvm-as < %s | llc -march=x86-64 | grep BB1_1: +; RUN: llvm-as < %s | llc -march=x86-64 -asm-verbose=false | FileCheck %s ; PR4126 +; PR4732 -; Don't omit this label's definition. +; Don't omit these labels' definitions. + +; CHECK: bux: +; CHECK: .LBB1_1: define void @bux(i32 %p_53) nounwind optsize { entry: @@ -21,3 +25,33 @@ bb3: ; preds = %bb.i, %entry } declare i32 @baz(...) + +; Don't omit this label in the assembly output. +; CHECK: int321: +; CHECK: jne .LBB2_1 +; CHECK: jle .LBB2_1 +; CHECK: .LBB2_1: + +define void @int321(i8 signext %p_103, i32 %uint8p_104) nounwind readnone { +entry: + %tobool = icmp eq i8 %p_103, 0 ; <i1> [#uses=1] + %cmp.i = icmp sgt i8 %p_103, 0 ; <i1> [#uses=1] + %or.cond = and i1 %tobool, %cmp.i ; <i1> [#uses=1] + br i1 %or.cond, label %land.end.i, label %for.cond.preheader + +land.end.i: ; preds = %entry + %conv3.i = sext i8 %p_103 to i32 ; <i32> [#uses=1] + %div.i = sdiv i32 1, %conv3.i ; <i32> [#uses=1] + %tobool.i = icmp eq i32 %div.i, -2147483647 ; <i1> [#uses=0] + br label %for.cond.preheader + +for.cond.preheader: ; preds = %land.end.i, %entry + %cmp = icmp sgt i8 %p_103, 1 ; <i1> [#uses=1] + br i1 %cmp, label %for.end.split, label %for.cond + +for.cond: ; preds = %for.cond.preheader, %for.cond + br label %for.cond + +for.end.split: ; preds = %for.cond.preheader + ret void +} |