[llvm] r323355 - Revert r321751, "StructurizeCFG: Fix broken backedge detection"
Nicolai Haehnle via llvm-commits
llvm-commits at lists.llvm.org
Wed Jan 24 10:02:05 PST 2018
Author: nha
Date: Wed Jan 24 10:02:05 2018
New Revision: 323355
URL: http://llvm.org/viewvc/llvm-project?rev=323355&view=rev
Log:
Revert r321751, "StructurizeCFG: Fix broken backedge detection"
It causes regressions in various OpenGL test suites.
Keep the test cases introduced by r321751 as XFAIL, and add a test case
for the regression.
Change-Id: I90b4cc354f68cebe5fcef1f2422dc8fe1c6d3514
Bugzilla: https://bugs.llvm.org/show_bug.cgi?id=36015
Added:
llvm/trunk/test/Transforms/StructurizeCFG/bug36015.ll
Modified:
llvm/trunk/lib/Transforms/Scalar/StructurizeCFG.cpp
llvm/trunk/test/CodeGen/AMDGPU/multilevel-break.ll
llvm/trunk/test/CodeGen/AMDGPU/nested-loop-conditions.ll
llvm/trunk/test/Transforms/StructurizeCFG/AMDGPU/backedge-id-bug.ll
llvm/trunk/test/Transforms/StructurizeCFG/nested-loop-order.ll
Modified: llvm/trunk/lib/Transforms/Scalar/StructurizeCFG.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/StructurizeCFG.cpp?rev=323355&r1=323354&r2=323355&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Scalar/StructurizeCFG.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/StructurizeCFG.cpp Wed Jan 24 10:02:05 2018
@@ -14,6 +14,7 @@
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/Analysis/DivergenceAnalysis.h"
+#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/RegionInfo.h"
#include "llvm/Analysis/RegionIterator.h"
#include "llvm/Analysis/RegionPass.h"
@@ -176,8 +177,9 @@ class StructurizeCFG : public RegionPass
Region *ParentRegion;
DominatorTree *DT;
+ LoopInfo *LI;
- std::deque<RegionNode *> Order;
+ SmallVector<RegionNode *, 8> Order;
BBSet Visited;
BBPhiMap DeletedPhis;
@@ -202,7 +204,7 @@ class StructurizeCFG : public RegionPass
void gatherPredicates(RegionNode *N);
- void analyzeNode(RegionNode *N);
+ void collectInfos();
void insertConditions(bool Loops);
@@ -256,6 +258,7 @@ public:
AU.addRequired<DivergenceAnalysis>();
AU.addRequiredID(LowerSwitchID);
AU.addRequired<DominatorTreeWrapperPass>();
+ AU.addRequired<LoopInfoWrapperPass>();
AU.addPreserved<DominatorTreeWrapperPass>();
RegionPass::getAnalysisUsage(AU);
@@ -289,17 +292,55 @@ bool StructurizeCFG::doInitialization(Re
/// \brief Build up the general order of nodes
void StructurizeCFG::orderNodes() {
- assert(Visited.empty());
- assert(Predicates.empty());
- assert(Loops.empty());
- assert(LoopPreds.empty());
-
- // This must be RPO order for the back edge detection to work
- for (RegionNode *RN : ReversePostOrderTraversal<Region*>(ParentRegion)) {
- // FIXME: Is there a better order to use for structurization?
- Order.push_back(RN);
- analyzeNode(RN);
+ ReversePostOrderTraversal<Region*> RPOT(ParentRegion);
+ SmallDenseMap<Loop*, unsigned, 8> LoopBlocks;
+
+ // The reverse post-order traversal of the list gives us an ordering close
+ // to what we want. The only problem with it is that sometimes backedges
+ // for outer loops will be visited before backedges for inner loops.
+ for (RegionNode *RN : RPOT) {
+ BasicBlock *BB = RN->getEntry();
+ Loop *Loop = LI->getLoopFor(BB);
+ ++LoopBlocks[Loop];
+ }
+
+ unsigned CurrentLoopDepth = 0;
+ Loop *CurrentLoop = nullptr;
+ for (auto I = RPOT.begin(), E = RPOT.end(); I != E; ++I) {
+ BasicBlock *BB = (*I)->getEntry();
+ unsigned LoopDepth = LI->getLoopDepth(BB);
+
+ if (is_contained(Order, *I))
+ continue;
+
+ if (LoopDepth < CurrentLoopDepth) {
+ // Make sure we have visited all blocks in this loop before moving back to
+ // the outer loop.
+
+ auto LoopI = I;
+ while (unsigned &BlockCount = LoopBlocks[CurrentLoop]) {
+ LoopI++;
+ BasicBlock *LoopBB = (*LoopI)->getEntry();
+ if (LI->getLoopFor(LoopBB) == CurrentLoop) {
+ --BlockCount;
+ Order.push_back(*LoopI);
+ }
+ }
+ }
+
+ CurrentLoop = LI->getLoopFor(BB);
+ if (CurrentLoop)
+ LoopBlocks[CurrentLoop]--;
+
+ CurrentLoopDepth = LoopDepth;
+ Order.push_back(*I);
}
+
+ // This pass originally used a post-order traversal and then operated on
+ // the list in reverse. Now that we are using a reverse post-order traversal
+ // rather than re-working the whole pass to operate on the list in order,
+ // we just reverse the list and continue to operate on it in reverse.
+ std::reverse(Order.begin(), Order.end());
}
/// \brief Determine the end of the loops
@@ -425,19 +466,32 @@ void StructurizeCFG::gatherPredicates(Re
}
/// \brief Collect various loop and predicate infos
-void StructurizeCFG::analyzeNode(RegionNode *RN) {
- DEBUG(dbgs() << "Visiting: "
- << (RN->isSubRegion() ? "SubRegion with entry: " : "")
- << RN->getEntry()->getName() << '\n');
+void StructurizeCFG::collectInfos() {
+ // Reset predicate
+ Predicates.clear();
+
+ // and loop infos
+ Loops.clear();
+ LoopPreds.clear();
+
+ // Reset the visited nodes
+ Visited.clear();
- // Analyze all the conditions leading to a node
- gatherPredicates(RN);
+ for (RegionNode *RN : reverse(Order)) {
+ DEBUG(dbgs() << "Visiting: "
+ << (RN->isSubRegion() ? "SubRegion with entry: " : "")
+ << RN->getEntry()->getName() << " Loop Depth: "
+ << LI->getLoopDepth(RN->getEntry()) << "\n");
- // Remember that we've seen this node
- Visited.insert(RN->getEntry());
+ // Analyze all the conditions leading to a node
+ gatherPredicates(RN);
- // Find the last back edges
- analyzeLoops(RN);
+ // Remember that we've seen this node
+ Visited.insert(RN->getEntry());
+
+ // Find the last back edges
+ analyzeLoops(RN);
+ }
}
/// \brief Insert the missing branch conditions
@@ -610,7 +664,7 @@ void StructurizeCFG::changeExit(RegionNo
BasicBlock *StructurizeCFG::getNextFlow(BasicBlock *Dominator) {
LLVMContext &Context = Func->getContext();
BasicBlock *Insert = Order.empty() ? ParentRegion->getExit() :
- Order.front()->getEntry();
+ Order.back()->getEntry();
BasicBlock *Flow = BasicBlock::Create(Context, FlowBlockName,
Func, Insert);
DT->addNewBlock(Flow, Dominator);
@@ -690,8 +744,7 @@ bool StructurizeCFG::isPredictableTrue(R
/// Take one node from the order vector and wire it up
void StructurizeCFG::wireFlow(bool ExitUseAllowed,
BasicBlock *LoopEnd) {
- RegionNode *Node = Order.front();
- Order.pop_front();
+ RegionNode *Node = Order.pop_back_val();
Visited.insert(Node->getEntry());
if (isPredictableTrue(Node)) {
@@ -715,7 +768,7 @@ void StructurizeCFG::wireFlow(bool ExitU
PrevNode = Node;
while (!Order.empty() && !Visited.count(LoopEnd) &&
- dominatesPredicates(Entry, Order.front())) {
+ dominatesPredicates(Entry, Order.back())) {
handleLoops(false, LoopEnd);
}
@@ -726,7 +779,7 @@ void StructurizeCFG::wireFlow(bool ExitU
void StructurizeCFG::handleLoops(bool ExitUseAllowed,
BasicBlock *LoopEnd) {
- RegionNode *Node = Order.front();
+ RegionNode *Node = Order.back();
BasicBlock *LoopStart = Node->getEntry();
if (!Loops.count(LoopStart)) {
@@ -871,9 +924,10 @@ bool StructurizeCFG::runOnRegion(Region
ParentRegion = R;
DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+ LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
orderNodes();
-
+ collectInfos();
createFlow();
insertConditions(false);
insertConditions(true);
Modified: llvm/trunk/test/CodeGen/AMDGPU/multilevel-break.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AMDGPU/multilevel-break.ll?rev=323355&r1=323354&r2=323355&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/multilevel-break.ll (original)
+++ llvm/trunk/test/CodeGen/AMDGPU/multilevel-break.ll Wed Jan 24 10:02:05 2018
@@ -66,10 +66,9 @@ ENDIF:
; OPT-LABEL: define amdgpu_kernel void @multi_if_break_loop(
; OPT: llvm.amdgcn.break
-; OPT: llvm.amdgcn.break
+; OPT: llvm.amdgcn.loop
; OPT: llvm.amdgcn.if.break
; OPT: llvm.amdgcn.if.break
-; OPT: llvm.amdgcn.loop
; OPT: llvm.amdgcn.end.cf
; GCN-LABEL: {{^}}multi_if_break_loop:
Modified: llvm/trunk/test/CodeGen/AMDGPU/nested-loop-conditions.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AMDGPU/nested-loop-conditions.ll?rev=323355&r1=323354&r2=323355&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/nested-loop-conditions.ll (original)
+++ llvm/trunk/test/CodeGen/AMDGPU/nested-loop-conditions.ll Wed Jan 24 10:02:05 2018
@@ -123,100 +123,55 @@ bb23:
; Earlier version of above, before a run of the structurizer.
; IR-LABEL: @nested_loop_conditions(
-; IR: %tmp1235 = icmp slt i32 %tmp1134, 9
-; IR: br i1 %tmp1235, label %bb14.lr.ph, label %Flow
-
-; IR: bb14.lr.ph:
-; IR: br label %bb14
-
-; IR: Flow3:
-; IR: call void @llvm.amdgcn.end.cf(i64 %18)
-; IR: %0 = call { i1, i64 } @llvm.amdgcn.if(i1 %17)
-; IR: %1 = extractvalue { i1, i64 } %0, 0
-; IR: %2 = extractvalue { i1, i64 } %0, 1
-; IR: br i1 %1, label %bb4.bb13_crit_edge, label %Flow4
-
-; IR: bb4.bb13_crit_edge:
-; IR: br label %Flow4
-
-; IR: Flow4:
-; IR: %3 = phi i1 [ true, %bb4.bb13_crit_edge ], [ false, %Flow3 ]
-; IR: call void @llvm.amdgcn.end.cf(i64 %2)
-; IR: br label %Flow
-
-; IR: bb13:
-; IR: br label %bb31
-
-; IR: Flow:
-; IR: %4 = phi i1 [ %3, %Flow4 ], [ true, %bb ]
-; IR: %5 = call { i1, i64 } @llvm.amdgcn.if(i1 %4)
-; IR: %6 = extractvalue { i1, i64 } %5, 0
-; IR: %7 = extractvalue { i1, i64 } %5, 1
-; IR: br i1 %6, label %bb13, label %bb31
-
-; IR: bb14:
-; IR: %phi.broken = phi i64 [ %18, %Flow2 ], [ 0, %bb14.lr.ph ]
-; IR: %tmp1037 = phi i32 [ %tmp1033, %bb14.lr.ph ], [ %16, %Flow2 ]
-; IR: %tmp936 = phi <4 x i32> [ %tmp932, %bb14.lr.ph ], [ %15, %Flow2 ]
-; IR: %tmp15 = icmp eq i32 %tmp1037, 1
-; IR: %8 = xor i1 %tmp15, true
-; IR: %9 = call { i1, i64 } @llvm.amdgcn.if(i1 %8)
-; IR: %10 = extractvalue { i1, i64 } %9, 0
-; IR: %11 = extractvalue { i1, i64 } %9, 1
-; IR: br i1 %10, label %bb31.loopexit, label %Flow1
+; IR: Flow7:
+; IR-NEXT: call void @llvm.amdgcn.end.cf(i64 %17)
+; IR-NEXT: %0 = call { i1, i64 } @llvm.amdgcn.if(i1 %15)
+; IR-NEXT: %1 = extractvalue { i1, i64 } %0, 0
+; IR-NEXT: %2 = extractvalue { i1, i64 } %0, 1
+; IR-NEXT: br i1 %1, label %bb4.bb13_crit_edge, label %Flow8
; IR: Flow1:
-; IR: %12 = call { i1, i64 } @llvm.amdgcn.else(i64 %11)
-; IR: %13 = extractvalue { i1, i64 } %12, 0
-; IR: %14 = extractvalue { i1, i64 } %12, 1
-; IR: br i1 %13, label %bb16, label %Flow2
-
-; IR: bb16:
-; IR: %tmp17 = bitcast i64 %tmp3 to <2 x i32>
-; IR: br label %bb18
+; IR-NEXT: %loop.phi = phi i64 [ %loop.phi9, %Flow6 ], [ %phi.broken, %bb14 ]
+; IR-NEXT: %13 = phi <4 x i32> [ %29, %Flow6 ], [ undef, %bb14 ]
+; IR-NEXT: %14 = phi i32 [ %30, %Flow6 ], [ undef, %bb14 ]
+; IR-NEXT: %15 = phi i1 [ %31, %Flow6 ], [ false, %bb14 ]
+; IR-NEXT: %16 = phi i1 [ false, %Flow6 ], [ %8, %bb14 ]
+; IR-NEXT: %17 = call i64 @llvm.amdgcn.else.break(i64 %11, i64 %loop.phi)
+; IR-NEXT: call void @llvm.amdgcn.end.cf(i64 %11)
+; IR-NEXT: %18 = call i1 @llvm.amdgcn.loop(i64 %17)
+; IR-NEXT: br i1 %18, label %Flow7, label %bb14
; IR: Flow2:
-; IR: %loop.phi = phi i64 [ %21, %bb21 ], [ %phi.broken, %Flow1 ]
-; IR: %15 = phi <4 x i32> [ %tmp9, %bb21 ], [ undef, %Flow1 ]
-; IR: %16 = phi i32 [ %tmp10, %bb21 ], [ undef, %Flow1 ]
-; IR: %17 = phi i1 [ %20, %bb21 ], [ false, %Flow1 ]
-; IR: %18 = call i64 @llvm.amdgcn.else.break(i64 %14, i64 %loop.phi)
-; IR: call void @llvm.amdgcn.end.cf(i64 %14)
-; IR: %19 = call i1 @llvm.amdgcn.loop(i64 %18)
-; IR: br i1 %19, label %Flow3, label %bb14
-
-; IR: bb18:
-; IR: %tmp19 = load volatile i32, i32 addrspace(1)* undef
-; IR: %tmp20 = icmp slt i32 %tmp19, 9
-; IR: br i1 %tmp20, label %bb21, label %bb18
+; IR-NEXT: %loop.phi10 = phi i64 [ %loop.phi11, %Flow5 ], [ %12, %bb16 ]
+; IR-NEXT: %19 = phi <4 x i32> [ %29, %Flow5 ], [ undef, %bb16 ]
+; IR-NEXT: %20 = phi i32 [ %30, %Flow5 ], [ undef, %bb16 ]
+; IR-NEXT: %21 = phi i1 [ %31, %Flow5 ], [ false, %bb16 ]
+; IR-NEXT: %22 = phi i1 [ false, %Flow5 ], [ false, %bb16 ]
+; IR-NEXT: %23 = phi i1 [ false, %Flow5 ], [ %8, %bb16 ]
+; IR-NEXT: %24 = call { i1, i64 } @llvm.amdgcn.if(i1 %23)
+; IR-NEXT: %25 = extractvalue { i1, i64 } %24, 0
+; IR-NEXT: %26 = extractvalue { i1, i64 } %24, 1
+; IR-NEXT: br i1 %25, label %bb21, label %Flow3
; IR: bb21:
-; IR: %tmp22 = extractelement <2 x i32> %tmp17, i64 1
-; IR: %tmp23 = lshr i32 %tmp22, 16
-; IR: %tmp24 = select i1 undef, i32 undef, i32 %tmp23
-; IR: %tmp25 = uitofp i32 %tmp24 to float
-; IR: %tmp26 = fmul float %tmp25, 0x3EF0001000000000
-; IR: %tmp27 = fsub float %tmp26, undef
-; IR: %tmp28 = fcmp olt float %tmp27, 5.000000e-01
-; IR: %tmp29 = select i1 %tmp28, i64 1, i64 2
-; IR: %tmp30 = extractelement <4 x i32> %tmp936, i64 %tmp29
-; IR: %tmp7 = zext i32 %tmp30 to i64
-; IR: %tmp8 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* undef, i64 %tmp7
-; IR: %tmp9 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp8, align 16
-; IR: %tmp10 = extractelement <4 x i32> %tmp9, i64 0
-; IR: %tmp11 = load volatile i32, i32 addrspace(1)* undef
-; IR: %tmp12 = icmp slt i32 %tmp11, 9
-; IR: %20 = xor i1 %tmp12, true
-; IR: %21 = call i64 @llvm.amdgcn.if.break(i1 %20, i64 %phi.broken)
-; IR: br label %Flow2
+; IR: %tmp12 = icmp slt i32 %tmp11, 9
+; IR-NEXT: %27 = xor i1 %tmp12, true
+; IR-NEXT: %28 = call i64 @llvm.amdgcn.if.break(i1 %27, i64 %phi.broken)
+; IR-NEXT: br label %Flow3
-; IR: bb31.loopexit:
-; IR: br label %Flow1
+; IR: Flow3:
+; IR-NEXT: %loop.phi11 = phi i64 [ %phi.broken, %bb21 ], [ %phi.broken, %Flow2 ]
+; IR-NEXT: %loop.phi9 = phi i64 [ %28, %bb21 ], [ %loop.phi10, %Flow2 ]
+; IR-NEXT: %29 = phi <4 x i32> [ %tmp9, %bb21 ], [ %19, %Flow2 ]
+; IR-NEXT: %30 = phi i32 [ %tmp10, %bb21 ], [ %20, %Flow2 ]
+; IR-NEXT: %31 = phi i1 [ %27, %bb21 ], [ %21, %Flow2 ]
+; IR-NEXT: call void @llvm.amdgcn.end.cf(i64 %26)
+; IR-NEXT: br i1 %22, label %bb31.loopexit, label %Flow4
; IR: bb31:
-; IR: call void @llvm.amdgcn.end.cf(i64 %7)
-; IR: store volatile i32 0, i32 addrspace(1)* undef
-; IR: ret void
+; IR-NEXT: call void @llvm.amdgcn.end.cf(i64 %7)
+; IR-NEXT: store volatile i32 0, i32 addrspace(1)* undef
+; IR-NEXT: ret void
; GCN-LABEL: {{^}}nested_loop_conditions:
Modified: llvm/trunk/test/Transforms/StructurizeCFG/AMDGPU/backedge-id-bug.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/StructurizeCFG/AMDGPU/backedge-id-bug.ll?rev=323355&r1=323354&r2=323355&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/StructurizeCFG/AMDGPU/backedge-id-bug.ll (original)
+++ llvm/trunk/test/Transforms/StructurizeCFG/AMDGPU/backedge-id-bug.ll Wed Jan 24 10:02:05 2018
@@ -1,3 +1,4 @@
+; XFAIL: *
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -mtriple=amdgcn-amd-amdhsa -S -structurizecfg %s | FileCheck %s
Added: llvm/trunk/test/Transforms/StructurizeCFG/bug36015.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/StructurizeCFG/bug36015.ll?rev=323355&view=auto
==============================================================================
--- llvm/trunk/test/Transforms/StructurizeCFG/bug36015.ll (added)
+++ llvm/trunk/test/Transforms/StructurizeCFG/bug36015.ll Wed Jan 24 10:02:05 2018
@@ -0,0 +1,53 @@
+; RUN: opt -S -structurizecfg %s | FileCheck %s
+
+; r321751 introduced a bug where control flow branching from if to exit was
+; not handled properly and instead ended up in an infinite loop.
+define void @bug36015(i32 %cmp0, i32 %count) {
+entry:
+ br label %loop.outer
+
+loop.outer:
+ %ctr.loop.outer = phi i32 [ 0, %entry ], [ %ctr.else, %else ]
+ call void @foo(i32 0)
+ br label %loop.inner
+
+loop.inner:
+ %ctr.loop.inner = phi i32 [ %ctr.loop.outer, %loop.outer ], [ %ctr.if, %if ]
+ call void @foo(i32 1)
+ %cond.inner = icmp eq i32 %cmp0, %ctr.loop.inner
+ br i1 %cond.inner, label %if, label %else
+
+; CHECK: if:
+; CHECK: %0 = xor i1 %cond.if, true
+; CHECK: br label %Flow
+if:
+ %ctr.if = add i32 %ctr.loop.inner, 1
+ call void @foo(i32 2)
+ %cond.if = icmp slt i32 %ctr.if, %count
+ br i1 %cond.if, label %loop.inner, label %exit
+
+; CHECK: Flow:
+; CHECK: %2 = phi i1 [ %0, %if ], [ true, %loop.inner ]
+; CHECK: %3 = phi i1 [ false, %if ], [ true, %loop.inner ]
+; CHECK: br i1 %2, label %Flow1, label %loop.inner
+
+; CHECK: Flow1:
+; CHECK: br i1 %3, label %else, label %Flow2
+
+; CHECK: else:
+; CHECK: br label %Flow2
+else:
+ %ctr.else = add i32 %ctr.loop.inner, 1
+ call void @foo(i32 3)
+ %cond.else = icmp slt i32 %ctr.else, %count
+ br i1 %cond.else, label %loop.outer, label %exit
+
+; CHECK: Flow2:
+; CHECK: %6 = phi i1 [ %4, %else ], [ true, %Flow1 ]
+; CHECK: br i1 %6, label %exit, label %loop.outer
+
+exit:
+ ret void
+}
+
+declare void @foo(i32)
Modified: llvm/trunk/test/Transforms/StructurizeCFG/nested-loop-order.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/StructurizeCFG/nested-loop-order.ll?rev=323355&r1=323354&r2=323355&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/StructurizeCFG/nested-loop-order.ll (original)
+++ llvm/trunk/test/Transforms/StructurizeCFG/nested-loop-order.ll Wed Jan 24 10:02:05 2018
@@ -1,76 +1,32 @@
-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -S -structurizecfg %s -o - | FileCheck %s
define void @main(float addrspace(1)* %out) {
-; CHECK-LABEL: @main(
-; CHECK-NEXT: main_body:
-; CHECK-NEXT: br label [[LOOP_OUTER:%.*]]
-; CHECK: LOOP.outer:
-; CHECK-NEXT: [[TEMP8_0_PH:%.*]] = phi float [ 0.000000e+00, [[MAIN_BODY:%.*]] ], [ [[TMP13:%.*]], [[FLOW3:%.*]] ]
-; CHECK-NEXT: [[TEMP4_0_PH:%.*]] = phi i32 [ 0, [[MAIN_BODY]] ], [ [[TMP12:%.*]], [[FLOW3]] ]
-; CHECK-NEXT: br label [[LOOP:%.*]]
-; CHECK: LOOP:
-; CHECK-NEXT: [[TMP0:%.*]] = phi i32 [ undef, [[LOOP_OUTER]] ], [ [[TMP12]], [[FLOW:%.*]] ]
-; CHECK-NEXT: [[TMP1:%.*]] = phi float [ undef, [[LOOP_OUTER]] ], [ [[TMP13]], [[FLOW]] ]
-; CHECK-NEXT: [[TEMP4_0:%.*]] = phi i32 [ [[TEMP4_0_PH]], [[LOOP_OUTER]] ], [ [[TMP15:%.*]], [[FLOW]] ]
-; CHECK-NEXT: [[TMP20:%.*]] = add i32 [[TEMP4_0]], 1
-; CHECK-NEXT: [[TMP22:%.*]] = icmp sgt i32 [[TMP20]], 3
-; CHECK-NEXT: [[TMP2:%.*]] = xor i1 [[TMP22]], true
-; CHECK-NEXT: br i1 [[TMP2]], label [[ENDIF:%.*]], label [[FLOW]]
-; CHECK: Flow2:
-; CHECK-NEXT: [[TMP3:%.*]] = phi float [ [[TEMP8_0_PH]], [[IF29:%.*]] ], [ [[TMP9:%.*]], [[FLOW1:%.*]] ]
-; CHECK-NEXT: [[TMP4:%.*]] = phi i32 [ [[TMP20]], [[IF29]] ], [ undef, [[FLOW1]] ]
-; CHECK-NEXT: [[TMP5:%.*]] = phi i1 [ [[TMP32:%.*]], [[IF29]] ], [ true, [[FLOW1]] ]
-; CHECK-NEXT: br label [[FLOW]]
-; CHECK: Flow3:
-; CHECK-NEXT: br i1 [[TMP16:%.*]], label [[ENDLOOP:%.*]], label [[LOOP_OUTER]]
-; CHECK: ENDLOOP:
-; CHECK-NEXT: [[TEMP8_1:%.*]] = phi float [ [[TMP14:%.*]], [[FLOW3]] ]
-; CHECK-NEXT: [[TMP23:%.*]] = icmp eq i32 [[TMP20]], 3
-; CHECK-NEXT: [[DOT45:%.*]] = select i1 [[TMP23]], float 0.000000e+00, float 1.000000e+00
-; CHECK-NEXT: store float [[DOT45]], float addrspace(1)* [[OUT:%.*]]
-; CHECK-NEXT: ret void
-; CHECK: ENDIF:
-; CHECK-NEXT: [[TMP31:%.*]] = icmp sgt i32 [[TMP20]], 1
-; CHECK-NEXT: [[TMP6:%.*]] = xor i1 [[TMP31]], true
-; CHECK-NEXT: br i1 [[TMP6]], label [[ENDIF28:%.*]], label [[FLOW1]]
-; CHECK: Flow1:
-; CHECK-NEXT: [[TMP7:%.*]] = phi i32 [ [[TMP20]], [[ENDIF28]] ], [ [[TMP0]], [[ENDIF]] ]
-; CHECK-NEXT: [[TMP8:%.*]] = phi float [ [[TMP35:%.*]], [[ENDIF28]] ], [ [[TMP1]], [[ENDIF]] ]
-; CHECK-NEXT: [[TMP9]] = phi float [ [[TMP35]], [[ENDIF28]] ], [ [[TEMP8_0_PH]], [[ENDIF]] ]
-; CHECK-NEXT: [[TMP10:%.*]] = phi i1 [ [[TMP36:%.*]], [[ENDIF28]] ], [ true, [[ENDIF]] ]
-; CHECK-NEXT: [[TMP11:%.*]] = phi i1 [ false, [[ENDIF28]] ], [ true, [[ENDIF]] ]
-; CHECK-NEXT: br i1 [[TMP11]], label [[IF29]], label [[FLOW2:%.*]]
-; CHECK: IF29:
-; CHECK-NEXT: [[TMP32]] = icmp sgt i32 [[TMP20]], 2
-; CHECK-NEXT: br label [[FLOW2]]
-; CHECK: Flow:
-; CHECK-NEXT: [[TMP12]] = phi i32 [ [[TMP7]], [[FLOW2]] ], [ [[TMP0]], [[LOOP]] ]
-; CHECK-NEXT: [[TMP13]] = phi float [ [[TMP8]], [[FLOW2]] ], [ [[TMP1]], [[LOOP]] ]
-; CHECK-NEXT: [[TMP14]] = phi float [ [[TMP3]], [[FLOW2]] ], [ [[TEMP8_0_PH]], [[LOOP]] ]
-; CHECK-NEXT: [[TMP15]] = phi i32 [ [[TMP4]], [[FLOW2]] ], [ undef, [[LOOP]] ]
-; CHECK-NEXT: [[TMP16]] = phi i1 [ [[TMP10]], [[FLOW2]] ], [ true, [[LOOP]] ]
-; CHECK-NEXT: [[TMP17:%.*]] = phi i1 [ [[TMP5]], [[FLOW2]] ], [ true, [[LOOP]] ]
-; CHECK-NEXT: br i1 [[TMP17]], label [[FLOW3]], label [[LOOP]]
-; CHECK: ENDIF28:
-; CHECK-NEXT: [[TMP35]] = fadd float [[TEMP8_0_PH]], 1.000000e+00
-; CHECK-NEXT: [[TMP36]] = icmp sgt i32 [[TMP20]], 2
-; CHECK-NEXT: br label [[FLOW1]]
-;
+
+; CHECK: main_body:
+; CHECK: br label %LOOP.outer
main_body:
br label %LOOP.outer
+; CHECK: LOOP.outer:
+; CHECK: br label %LOOP
LOOP.outer: ; preds = %ENDIF28, %main_body
%temp8.0.ph = phi float [ 0.000000e+00, %main_body ], [ %tmp35, %ENDIF28 ]
%temp4.0.ph = phi i32 [ 0, %main_body ], [ %tmp20, %ENDIF28 ]
br label %LOOP
+; CHECK: LOOP:
+; br i1 %{{[0-9]+}}, label %ENDIF, label %Flow
LOOP: ; preds = %IF29, %LOOP.outer
%temp4.0 = phi i32 [ %temp4.0.ph, %LOOP.outer ], [ %tmp20, %IF29 ]
%tmp20 = add i32 %temp4.0, 1
%tmp22 = icmp sgt i32 %tmp20, 3
br i1 %tmp22, label %ENDLOOP, label %ENDIF
+; CHECK: Flow3
+; CHECK: br i1 %{{[0-9]+}}, label %ENDLOOP, label %LOOP.outer
+
+; CHECK: ENDLOOP:
+; CHECK: ret void
ENDLOOP: ; preds = %ENDIF28, %IF29, %LOOP
%temp8.1 = phi float [ %temp8.0.ph, %LOOP ], [ %temp8.0.ph, %IF29 ], [ %tmp35, %ENDIF28 ]
%tmp23 = icmp eq i32 %tmp20, 3
@@ -78,14 +34,29 @@ ENDLOOP:
store float %.45, float addrspace(1)* %out
ret void
+; CHECK: ENDIF:
+; CHECK: br i1 %tmp31, label %IF29, label %Flow1
ENDIF: ; preds = %LOOP
%tmp31 = icmp sgt i32 %tmp20, 1
br i1 %tmp31, label %IF29, label %ENDIF28
+; CHECK: Flow:
+; CHECK: br i1 %{{[0-9]+}}, label %Flow2, label %LOOP
+
+; CHECK: IF29:
+; CHECK: br label %Flow1
IF29: ; preds = %ENDIF
%tmp32 = icmp sgt i32 %tmp20, 2
br i1 %tmp32, label %ENDLOOP, label %LOOP
+; CHECK: Flow1:
+; CHECK: br label %Flow
+
+; CHECK: Flow2:
+; CHECK: br i1 %{{[0-9]+}}, label %ENDIF28, label %Flow3
+
+; CHECK: ENDIF28:
+; CHECK: br label %Flow3
ENDIF28: ; preds = %ENDIF
%tmp35 = fadd float %temp8.0.ph, 1.0
%tmp36 = icmp sgt i32 %tmp20, 2
More information about the llvm-commits
mailing list