[polly] r264118 - [ScopInfo] Fix domains after loops.
Michael Kruse via llvm-commits
llvm-commits at lists.llvm.org
Thu Mar 24 06:55:04 PDT 2016
2016-03-24 13:57 GMT+01:00 Johannes Doerfert <doerfert at cs.uni-saarland.de>:
> Hey Michael,
>
> I inlined some comments which we might want to address. If you want I
> can give you feedback on such changes also prior to the commit.
>
> 1) It seems to me there are actually two things done here
> a) The removal of the line
> Domain = isl_set_lower_bound_si(Domain,...)
> fits the "fix" part in the commit message (2nd paragraph).
> b) The bigger rewrite of the dimension adjustment part. Here I am not
> sure how it is connected to commit message. Additionally, I do not
> understand the need for some of the changes here. Especially since
> I only need to remove the line mentioned above and add one
> additional line (see below) to pass all tests after your change:
>
> if (BBLoopDepth < PredBBLoopDepth)
> PredBBDom = isl_set_project_out(
> PredBBDom, isl_dim_set, isl_set_n_dim(PredBBDom) - LoopDepthDiff,
> LoopDepthDiff);
> else if (PredBBLoopDepth < BBLoopDepth) {
> assert(LoopDepthDiff == 1);
> PredBBDom = isl_set_add_dims(PredBBDom, isl_dim_set, 1);
> + PredBBDom = addDomainDimId(PredBBDom, BBLoopDepth, BBLoop);
> } else if (BBLoop != PredBBLoop && BBLoopDepth >= 0) {
> assert(LoopDepthDiff <= 1);
> PredBBDom = isl_set_drop_constraints_involving_dims(
> PredBBDom, isl_dim_set, BBLoopDepth, 1);
> }
At least the isl_set_drop_constraints_involving_dims is also wrong. It
does not adapt the isl_id from the old to the new Loop and is
different to project_out and adding a dimension again. I encountered
this during LNT runs, not check-polly.
I am not convinced that the previous code was working in unusual cases
such as when we leave 2 or more loops and enter one, which is
indistinguishable from leaving just one when looking only in the depth
difference. Also for some reasons the 3 conditions are structurally
different from the code with the same purpose in
buildDomainsWithBranchConstraints.
When isl_set_drop_constraints_involving_dims, I still encountered LNT
fails. It did not seem worth fixing more bugs in this code.
>
> 2) I think that the effect of this change would could have been achieved
> differently, namely if we would use the DomTree and PostDomTree to
> simplify domains in general. That was proposed several times already
> in the context of our current domain generation compile-time bugs and
> it should takle the issue in a general way.
I think these two are orthogonal as the complexity of the domain is
not the problem. The domains in the loop are logically correct.
Could you maybe explain how you think (Post-)DomTree information would
keep ISL from deducing %p_0 > 0 in loop-succ-cond.ll ?
> Cheers,
> Johannes
>
> On 03/22, Michael Kruse via llvm-commits wrote:
>> Author: meinersbur
>> Date: Tue Mar 22 18:27:42 2016
>> New Revision: 264118
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=264118&view=rev
>> Log:
>> [ScopInfo] Fix domains after loops.
>>
>> ISL can conclude additional conditions on parameters from restrictions
>> on loop variables. Such conditions persist when leaving the loop and the
>> loop variable is projected out. This results in a narrower domain for
>> exiting the loop than entering it and is logically impossible for
>> non-infinite loops.
>>
>> We fix this by not adding a lower bound i>=0 when constructing BB
>> domains, but defer it to when also the upper bound it computed, which
>> was done redundantly even before this patch.
>>
>> This reduces the number of LNT fails with -polly-process-unprofitable
>> -polly-position=before-vectorizer from 8 to 6.
>>
>> Added:
>> polly/trunk/test/ScopInfo/loop-succ-cond.ll
>> Modified:
>> polly/trunk/lib/Analysis/ScopInfo.cpp
>> polly/trunk/test/Isl/CodeGen/phi_scalar_simple_1.ll
>> polly/trunk/test/ScopInfo/intra_and_inter_bb_scalar_dep.ll
>> polly/trunk/test/ScopInfo/isl_trip_count_03.ll
>> polly/trunk/test/ScopInfo/phi_scalar_simple_1.ll
>> polly/trunk/test/ScopInfo/remarks.ll
>> polly/trunk/test/ScopInfo/switch-1.ll
>> polly/trunk/test/ScopInfo/switch-2.ll
>> polly/trunk/test/ScopInfo/switch-3.ll
>> polly/trunk/test/ScopInfo/switch-4.ll
>>
>> Modified: polly/trunk/lib/Analysis/ScopInfo.cpp
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/lib/Analysis/ScopInfo.cpp?rev=264118&r1=264117&r2=264118&view=diff
>> ==============================================================================
>> --- polly/trunk/lib/Analysis/ScopInfo.cpp (original)
>> +++ polly/trunk/lib/Analysis/ScopInfo.cpp Tue Mar 22 18:27:42 2016
>> @@ -2101,7 +2101,6 @@ static bool containsErrorBlock(RegionNod
>>
>> static inline __isl_give isl_set *addDomainDimId(__isl_take isl_set *Domain,
>> unsigned Dim, Loop *L) {
>> - Domain = isl_set_lower_bound_si(Domain, isl_dim_set, Dim, -1);
>> isl_id *DimId =
>> isl_id_alloc(isl_set_get_ctx(Domain), nullptr, static_cast<void *>(L));
>> return isl_set_set_dim_id(Domain, isl_dim_set, Dim, DimId);
>> @@ -2366,7 +2365,6 @@ void Scop::propagateDomainConstraints(Re
>> }
>>
>> Loop *BBLoop = getRegionNodeLoop(RN, LI);
>> - int BBLoopDepth = getRelativeLoopDepth(BBLoop);
>>
>> isl_set *PredDom = isl_set_empty(isl_set_get_space(Domain));
>> for (auto *PredBB : predecessors(BB)) {
>> @@ -2383,32 +2381,30 @@ void Scop::propagateDomainConstraints(Re
>>
>> if (!PredBBDom) {
>> // Determine the loop depth of the predecessor and adjust its domain to
>> - // the domain of the current block. This can mean we have to:
>> - // o) Drop a dimension if this block is the exit of a loop, not the
>> - // header of a new loop and the predecessor was part of the loop.
>> - // o) Add an unconstrainted new dimension if this block is the header
>> - // of a loop and the predecessor is not part of it.
>> - // o) Drop the information about the innermost loop dimension when the
>> - // predecessor and the current block are surrounded by different
>> - // loops in the same depth.
>> + // the domain of the current block. This means we have to:
>> + // o) Drop all loop dimension of loops we are leaving.
>> + // o) Add a dimension for each loop we are entering.
> I think the comments before were more descriptive and as they still
> apply [at least hey should] I would not have changed them.
I concluded that the three bullet point would be coined to the the
conditions below. As these are removed, the bullet point would not
explain the code accurately anymore.
>> PredBBDom = getDomainForBlock(PredBB, DomainMap, *R->getRegionInfo());
>> Loop *PredBBLoop = LI.getLoopFor(PredBB);
>> while (BoxedLoops.count(PredBBLoop))
>> PredBBLoop = PredBBLoop->getParentLoop();
>>
>> - int PredBBLoopDepth = getRelativeLoopDepth(PredBBLoop);
>> - unsigned LoopDepthDiff = std::abs(BBLoopDepth - PredBBLoopDepth);
>> - if (BBLoopDepth < PredBBLoopDepth)
>> - PredBBDom = isl_set_project_out(
>> - PredBBDom, isl_dim_set, isl_set_n_dim(PredBBDom) - LoopDepthDiff,
>> - LoopDepthDiff);
>> - else if (PredBBLoopDepth < BBLoopDepth) {
>> - assert(LoopDepthDiff == 1);
>> - PredBBDom = isl_set_add_dims(PredBBDom, isl_dim_set, 1);
>> - } else if (BBLoop != PredBBLoop && BBLoopDepth >= 0) {
>> - assert(LoopDepthDiff <= 1);
>> - PredBBDom = isl_set_drop_constraints_involving_dims(
>> - PredBBDom, isl_dim_set, BBLoopDepth, 1);
>> + Loop *LeaveL = PredBBLoop;
>> + while (getRegion().contains(LeaveL) &&
>> + (!BBLoop || !LeaveL->contains(BBLoop))) {
>> + PredBBDom = isl_set_project_out(PredBBDom, isl_dim_set,
>> + isl_set_n_dim(PredBBDom) - 1, 1);
>> + LeaveL = LeaveL->getParentLoop();
>> + }
> Why do you project out loops one by one now? Before we project all
> dimensions out at once.
Simpler code and symmetry to the second while-loop. The possible
performance gains from project_out of multiple dimensions at once are
comparably insignificant.
>> + unsigned CommonDepth = isl_set_n_dim(PredBBDom);
>> +
>> + Loop *EnterL = BBLoop;
>> + while (getRegion().contains(EnterL) &&
>> + (!PredBBLoop || !EnterL->contains(PredBBLoop))) {
>> + PredBBDom =
>> + isl_set_insert_dims(PredBBDom, isl_dim_set, CommonDepth, 1);
> Why do you insert dimensions here? Before we "added" them which seems
> still fine and less specific.
Because the loop runs in reverse. The first encountered loop (i.e. the
innermost) must end up as the last one, so we insert the outermore
ones before the ones already inserted.
>> + PredBBDom = addDomainDimId(PredBBDom, CommonDepth, EnterL);
>> + EnterL = EnterL->getParentLoop();
>> }
>> }
>>
>>
>> Modified: polly/trunk/test/Isl/CodeGen/phi_scalar_simple_1.ll
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/Isl/CodeGen/phi_scalar_simple_1.ll?rev=264118&r1=264117&r2=264118&view=diff
>> ==============================================================================
>> --- polly/trunk/test/Isl/CodeGen/phi_scalar_simple_1.ll (original)
>> +++ polly/trunk/test/Isl/CodeGen/phi_scalar_simple_1.ll Tue Mar 22 18:27:42 2016
>> @@ -28,7 +28,7 @@ entry:
>> ; CHECK-LABEL: polly.start:
>> ; CHECK: store i32 %x, i32* %x.addr.0.phiops
>>
>> -; CHECK-LABEL: polly.merge:
>> +; CHECK-LABEL: polly.exiting:
>> ; CHECK: %x.addr.0.final_reload = load i32, i32* %x.addr.0.s2a
>>
>> for.cond: ; preds = %for.inc4, %entry
>>
>> Modified: polly/trunk/test/ScopInfo/intra_and_inter_bb_scalar_dep.ll
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/ScopInfo/intra_and_inter_bb_scalar_dep.ll?rev=264118&r1=264117&r2=264118&view=diff
>> ==============================================================================
>> --- polly/trunk/test/ScopInfo/intra_and_inter_bb_scalar_dep.ll (original)
>> +++ polly/trunk/test/ScopInfo/intra_and_inter_bb_scalar_dep.ll Tue Mar 22 18:27:42 2016
>> @@ -15,7 +15,7 @@
>> ; CHECK: Invariant Accesses: {
>> ; CHECK-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 0]
>> ; CHECK-NEXT: [N] -> { Stmt_for_j[i0, i1] -> MemRef_init_ptr[0] };
>> -; CHECK-NEXT: Execution Context: [N] -> { : N < 0 or N > 0 }
>> +; CHECK-NEXT: Execution Context: [N] -> { : N > 0 }
>> ; CHECK-NEXT: }
>> ;
>> ; CHECK: Statements {
>>
>> Modified: polly/trunk/test/ScopInfo/isl_trip_count_03.ll
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/ScopInfo/isl_trip_count_03.ll?rev=264118&r1=264117&r2=264118&view=diff
>> ==============================================================================
>> --- polly/trunk/test/ScopInfo/isl_trip_count_03.ll (original)
>> +++ polly/trunk/test/ScopInfo/isl_trip_count_03.ll Tue Mar 22 18:27:42 2016
>> @@ -12,7 +12,7 @@
>> ; CHECK: Statements {
>> ; CHECK-NEXT: Stmt_for_next
>> ; CHECK-NEXT: Domain :=
>> -; CHECK-NEXT: [n] -> { Stmt_for_next[i0] : i0 >= 0 and 2i0 <= -3 + n };
>> +; CHECK-NEXT: [n] -> { Stmt_for_next[i0] : n > 0 and i0 >= 0 and 2i0 <= -3 + n };
>> ; CHECK-NEXT: Schedule :=
>> ; CHECK-NEXT: [n] -> { Stmt_for_next[i0] -> [i0] };
>> ; CHECK-NEXT: ReadAccess := [Reduction Type: +] [Scalar: 0]
>>
>> Added: polly/trunk/test/ScopInfo/loop-succ-cond.ll
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/ScopInfo/loop-succ-cond.ll?rev=264118&view=auto
>> ==============================================================================
>> --- polly/trunk/test/ScopInfo/loop-succ-cond.ll (added)
>> +++ polly/trunk/test/ScopInfo/loop-succ-cond.ll Tue Mar 22 18:27:42 2016
>> @@ -0,0 +1,77 @@
>> +; RUN: opt %loadPolly -polly-scops -analyze < %s | FileCheck %s
>> +; RUN: opt %loadPolly -polly-codegen -S < %s
>> +
>> +; Bugpoint-reduced from
>> +; test-suite/SingleSource/Benchmarks/Adobe-C++/loop_unroll.cpp
>> +;
>> +; Check that the loop %loop_start does not narrow the domain (the domain when
>> +; entering and leaving the loop must be identical)
>> +;
>> +; What happened in detail:
>> +;
>> +; 1) if.end5 checks %count whether there will be at least one iteration of
>> +; loop_start. The domain condition added to loop_start therefore is
>> +; [count] -> { [] : count > 0 }
>> +; 2) The loop exit condition of loop_start tests whether
>> +; %indvars.iv.next64 == %0 (which is zext i32 %count to i64, NOT %count
>> +; itself). %0 and %count have to be treated as independent parameters. The
>> +; loop exit condition is
>> +; [p_0] -> { [i0] : i0 = p_0 - 1 }
>> +; 3) Normalized loop induction variables are always non-negative. The domain
>> +; condition for this is loop
>> +; { [i0] : i0 >= 0 }
>> +; 4) The intersection of all three sets (condition of executing/entering loop,
>> +; non-negative induction variables, loop exit condition) is
>> +; [count, p_0] -> { [i0] : count > 0 and i0 >= 0 and i0 = p_0 - 1 }
>> +; 5) from which ISL can derive
>> +; [count, p_0] -> { [i0] : p_0 > 0 }
>> +; 6) if.end5 is either executed when skipping the loop
>> +; (domain [count] -> { [] : count <= 0 })
>> +; or though the loop.
>> +; 7) Assuming the loop is guaranteed to exit, Polly computes the after-the-loop
>> +; domain by taking the loop exit condition and projecting-out the induction
>> +; variable. This yields
>> +; [count, p_0] -> { [] : count > 0 and p_0 > 0 }
>> +; 8) The disjunction of both cases, 6) and 7)
>> +; (the two incoming edges of if.end12) is
>> +; [count, p_0] -> { [] : count <= 0 or (count > 0 and p_0 > 0) }
>> +; 9) Notice that if.end12 is logically _always_ executed in every scop
>> +; execution. Both cases of if.end5 will eventually land in if.end12
>> +
>> +define void @func(i32 %count, float* %A) {
>> +entry:
>> + %0 = zext i32 %count to i64
>> + br i1 undef, label %if.end5.preheader, label %for.end
>> +
>> +if.end5.preheader:
>> + %cmp6 = icmp sgt i32 %count, 0
>> + br label %if.end5
>> +
>> +if.end5:
>> + br i1 %cmp6, label %loop_start, label %if.end12
>> +
>> +loop_start:
>> + %indvars.iv63 = phi i64 [ %indvars.iv.next64, %loop_start ], [ 0, %if.end5 ]
>> + %add8 = add i32 undef, undef
>> + %indvars.iv.next64 = add nuw nsw i64 %indvars.iv63, 1
>> + %cmp9 = icmp eq i64 %indvars.iv.next64, %0
>> + br i1 %cmp9, label %if.end12, label %loop_start
>> +
>> +if.end12:
>> + store float 0.0, float* %A
>> + br label %for.end
>> +
>> +for.end:
>> + ret void
>> +}
>> +
>> +
>> +; CHECK: Statements {
>> +; CHECK-NEXT: Stmt_if_end12
>> +; CHECK-NEXT: Domain :=
>> +; CHECK-NEXT: [count, p_1] -> { Stmt_if_end12[] };
>> +; CHECK-NEXT: Schedule :=
>> +; CHECK-NEXT: [count, p_1] -> { Stmt_if_end12[] -> [] };
>> +; CHECK-NEXT: MustWriteAccess := [Reduction Type: NONE] [Scalar: 0]
>> +; CHECK-NEXT: [count, p_1] -> { Stmt_if_end12[] -> MemRef_A[0] };
>> +; CHECK-NEXT: }
>>
>> Modified: polly/trunk/test/ScopInfo/phi_scalar_simple_1.ll
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/ScopInfo/phi_scalar_simple_1.ll?rev=264118&r1=264117&r2=264118&view=diff
>> ==============================================================================
>> --- polly/trunk/test/ScopInfo/phi_scalar_simple_1.ll (original)
>> +++ polly/trunk/test/ScopInfo/phi_scalar_simple_1.ll Tue Mar 22 18:27:42 2016
>> @@ -17,16 +17,16 @@
>> ; CHECK: Statements {
>> ; CHECK-NEXT: Stmt_for_cond
>> ; CHECK-NEXT: Domain :=
>> -; CHECK-NEXT: [N] -> { Stmt_for_cond[i0] : N >= 2 and 0 <= i0 < N; Stmt_for_cond[0] : N <= 1 };
>> +; CHECK-NEXT: [N] -> { Stmt_for_cond[i0] : 0 <= i0 < N; Stmt_for_cond[0] : N <= 0 };
>> ; CHECK-NEXT: Schedule :=
>> -; CHECK-NEXT: [N] -> { Stmt_for_cond[i0] -> [i0, 0, 0, 0] : N >= 2 and i0 < N; Stmt_for_cond[0] -> [0, 0, 0, 0] : N <= 1 };
>> +; CHECK-NEXT: [N] -> { Stmt_for_cond[i0] -> [i0, 0, 0, 0] : i0 < N; Stmt_for_cond[0] -> [0, 0, 0, 0] : N <= 0 };
>> ; CHECK-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 1]
>> ; CHECK-NEXT: [N] -> { Stmt_for_cond[i0] -> MemRef_x_addr_0__phi[] };
>> ; CHECK-NEXT: MustWriteAccess := [Reduction Type: NONE] [Scalar: 1]
>> ; CHECK-NEXT: [N] -> { Stmt_for_cond[i0] -> MemRef_x_addr_0[] };
>> ; CHECK-NEXT: Stmt_for_body
>> ; CHECK-NEXT: Domain :=
>> -; CHECK-NEXT: [N] -> { Stmt_for_body[i0] : N >= 2 and 0 <= i0 <= -2 + N };
>> +; CHECK-NEXT: [N] -> { Stmt_for_body[i0] : 0 <= i0 <= -2 + N };
>> ; CHECK-NEXT: Schedule :=
>> ; CHECK-NEXT: [N] -> { Stmt_for_body[i0] -> [i0, 1, 0, 0] };
>> ; CHECK-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 1]
>>
>> Modified: polly/trunk/test/ScopInfo/remarks.ll
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/ScopInfo/remarks.ll?rev=264118&r1=264117&r2=264118&view=diff
>> ==============================================================================
>> --- polly/trunk/test/ScopInfo/remarks.ll (original)
>> +++ polly/trunk/test/ScopInfo/remarks.ll Tue Mar 22 18:27:42 2016
>> @@ -1,10 +1,10 @@
>> ; RUN: opt %loadPolly -pass-remarks-analysis="polly-scops" -polly-scops -disable-output < %s 2>&1 | FileCheck %s
>> ;
>> ; CHECK: remark: test/ScopInfo/remarks.c:4:7: SCoP begins here.
>> -; CHECK: remark: test/ScopInfo/remarks.c:8:5: Finite loop restriction: [M, N, Debug] -> { : N > 0 and (M <= -2 or M = -1) }
>> -; CHECK: remark: test/ScopInfo/remarks.c:13:7: No-error restriction: [M, N, Debug] -> { : M >= 0 and N > 0 and (Debug < 0 or Debug > 0) }
>> -; CHECK: remark: test/ScopInfo/remarks.c:9:7: Inbounds assumption: [M, N, Debug] -> { : M <= 100 or (M > 0 and N <= 0) }
>> -; CHECK: remark: <unknown>:0:0: No-overflows restriction: [N, M, Debug] -> { : M <= -2147483649 - N or M >= 2147483648 - N }
>> +; CHECK: remark: test/ScopInfo/remarks.c:8:5: Finite loop restriction: [M, N, Debug] -> { : M < 0 and N > 0 }
>> +; CHECK: remark: test/ScopInfo/remarks.c:13:7: No-error restriction: [M, N, Debug] -> { : M >= 0 and N > 0 and (Debug < 0 or Debug > 0) }
>> +; CHECK: remark: test/ScopInfo/remarks.c:9:7: Inbounds assumption: [M, N, Debug] -> { : M <= 100 or (M > 0 and N <= 0) }
>> +; CHECK: remark: <unknown>:0:0: No-overflows restriction: [N, M, Debug] -> { : M <= -2147483649 - N or M >= 2147483648 - N }
>> ; CHECK: remark: test/ScopInfo/remarks.c:9:18: Possibly aliasing pointer, use restrict keyword.
>> ; CHECK: remark: test/ScopInfo/remarks.c:9:33: Possibly aliasing pointer, use restrict keyword.
>> ; CHECK: remark: test/ScopInfo/remarks.c:9:15: Possibly aliasing pointer, use restrict keyword.
>>
>> Modified: polly/trunk/test/ScopInfo/switch-1.ll
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/ScopInfo/switch-1.ll?rev=264118&r1=264117&r2=264118&view=diff
>> ==============================================================================
>> --- polly/trunk/test/ScopInfo/switch-1.ll (original)
>> +++ polly/trunk/test/ScopInfo/switch-1.ll Tue Mar 22 18:27:42 2016
>> @@ -21,7 +21,7 @@
>> ; CHECK: Statements {
>> ; CHECK-NEXT: Stmt_sw_bb_1
>> ; CHECK-NEXT: Domain :=
>> -; CHECK-NEXT: [N] -> { Stmt_sw_bb_1[i0] : 4*floor((-1 + i0)/4) = -1 + i0 and 0 < i0 < N };
>> +; CHECK-NEXT: [N] -> { Stmt_sw_bb_1[i0] : 4*floor((-1 + i0)/4) = -1 + i0 and 0 <= i0 < N };
>> ; CHECK-NEXT: Schedule :=
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb_1[i0] -> [i0, 2] };
>> ; CHECK-NEXT: ReadAccess := [Reduction Type: +] [Scalar: 0]
>> @@ -30,7 +30,7 @@
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb_1[i0] -> MemRef_A[i0] };
>> ; CHECK-NEXT: Stmt_sw_bb_2
>> ; CHECK-NEXT: Domain :=
>> -; CHECK-NEXT: [N] -> { Stmt_sw_bb_2[i0] : 4*floor((-2 + i0)/4) = -2 + i0 and 2 <= i0 < N };
>> +; CHECK-NEXT: [N] -> { Stmt_sw_bb_2[i0] : 4*floor((-2 + i0)/4) = -2 + i0 and 0 <= i0 < N };
>> ; CHECK-NEXT: Schedule :=
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb_2[i0] -> [i0, 1] };
>> ; CHECK-NEXT: ReadAccess := [Reduction Type: +] [Scalar: 0]
>> @@ -50,18 +50,13 @@
>>
>> ; AST: if (1)
>> ;
>> -; AST: {
>> -; AST-NEXT: for (int c0 = 1; c0 < N - 2; c0 += 4) {
>> -; AST-NEXT: Stmt_sw_bb_1(c0);
>> +; AST: for (int c0 = 1; c0 < N; c0 += 4) {
>> +; AST-NEXT: Stmt_sw_bb_1(c0);
>> +; AST-NEXT: if (N >= c0 + 2) {
>> ; AST-NEXT: Stmt_sw_bb_2(c0 + 1);
>> -; AST-NEXT: Stmt_sw_bb_6(c0 + 2);
>> +; AST-NEXT: if (N >= c0 + 3)
>> +; AST-NEXT: Stmt_sw_bb_6(c0 + 2);
>> ; AST-NEXT: }
>> -; AST-NEXT: if (N >= 2)
>> -; AST-NEXT: if (N % 4 >= 2) {
>> -; AST-NEXT: Stmt_sw_bb_1(-(N % 4) + N + 1);
>> -; AST-NEXT: if ((N - 3) % 4 == 0)
>> -; AST-NEXT: Stmt_sw_bb_2(N - 1);
>> -; AST-NEXT: }
>> ; AST-NEXT: }
>> ;
>> ; AST: else
>>
>> Modified: polly/trunk/test/ScopInfo/switch-2.ll
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/ScopInfo/switch-2.ll?rev=264118&r1=264117&r2=264118&view=diff
>> ==============================================================================
>> --- polly/trunk/test/ScopInfo/switch-2.ll (original)
>> +++ polly/trunk/test/ScopInfo/switch-2.ll Tue Mar 22 18:27:42 2016
>> @@ -29,7 +29,7 @@
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb[i0] -> MemRef_A[i0] };
>> ; CHECK-NEXT: Stmt_sw_bb_2
>> ; CHECK-NEXT: Domain :=
>> -; CHECK-NEXT: [N] -> { Stmt_sw_bb_2[i0] : 4*floor((-2 + i0)/4) = -2 + i0 and 2 <= i0 < N };
>> +; CHECK-NEXT: [N] -> { Stmt_sw_bb_2[i0] : 4*floor((-2 + i0)/4) = -2 + i0 and 0 <= i0 < N };
>> ; CHECK-NEXT: Schedule :=
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb_2[i0] -> [i0, 0] };
>> ; CHECK-NEXT: ReadAccess := [Reduction Type: +] [Scalar: 0]
>> @@ -40,10 +40,13 @@
>>
>> ; AST: if (1)
>> ;
>> -; AST: for (int c0 = 0; c0 < N; c0 += 4) {
>> -; AST-NEXT: Stmt_sw_bb(c0);
>> -; AST-NEXT: if (N >= c0 + 3)
>> +; AST: {
>> +; AST-NEXT: for (int c0 = 0; c0 < N - 2; c0 += 4) {
>> +; AST-NEXT: Stmt_sw_bb(c0);
>> ; AST-NEXT: Stmt_sw_bb_2(c0 + 2);
>> +; AST-NEXT: }
>> +; AST-NEXT: if (N >= 1 && (N + 1) % 4 >= 2)
>> +; AST-NEXT: Stmt_sw_bb(-((N + 1) % 4) + N + 1);
>> ; AST-NEXT: }
>> ;
>> ; AST: else
>>
>> Modified: polly/trunk/test/ScopInfo/switch-3.ll
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/ScopInfo/switch-3.ll?rev=264118&r1=264117&r2=264118&view=diff
>> ==============================================================================
>> --- polly/trunk/test/ScopInfo/switch-3.ll (original)
>> +++ polly/trunk/test/ScopInfo/switch-3.ll Tue Mar 22 18:27:42 2016
>> @@ -38,7 +38,7 @@
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb_1[i0] -> MemRef_A[i0] };
>> ; CHECK-NEXT: Stmt_sw_bb_5
>> ; CHECK-NEXT: Domain :=
>> -; CHECK-NEXT: [N] -> { Stmt_sw_bb_5[i0] : 4*floor((-2 + i0)/4) = -2 + i0 and 2 <= i0 < N };
>> +; CHECK-NEXT: [N] -> { Stmt_sw_bb_5[i0] : 4*floor((-2 + i0)/4) = -2 + i0 and 0 <= i0 < N };
>> ; CHECK-NEXT: Schedule :=
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb_5[i0] -> [i0, 0] };
>> ; CHECK-NEXT: ReadAccess := [Reduction Type: +] [Scalar: 0]
>>
>> Modified: polly/trunk/test/ScopInfo/switch-4.ll
>> URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/ScopInfo/switch-4.ll?rev=264118&r1=264117&r2=264118&view=diff
>> ==============================================================================
>> --- polly/trunk/test/ScopInfo/switch-4.ll (original)
>> +++ polly/trunk/test/ScopInfo/switch-4.ll Tue Mar 22 18:27:42 2016
>> @@ -33,7 +33,7 @@
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb[i0] -> MemRef_A[i0] };
>> ; CHECK-NEXT: Stmt_sw_bb_1
>> ; CHECK-NEXT: Domain :=
>> -; CHECK-NEXT: [N] -> { Stmt_sw_bb_1[i0] : 4*floor((-1 + i0)/4) = -1 + i0 and 0 < i0 < N };
>> +; CHECK-NEXT: [N] -> { Stmt_sw_bb_1[i0] : 4*floor((-1 + i0)/4) = -1 + i0 and 0 <= i0 < N };
>> ; CHECK-NEXT: Schedule :=
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb_1[i0] -> [i0, 2] };
>> ; CHECK-NEXT: ReadAccess := [Reduction Type: +] [Scalar: 0]
>> @@ -42,7 +42,7 @@
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb_1[i0] -> MemRef_A[i0] };
>> ; CHECK-NEXT: Stmt_sw_bb_5
>> ; CHECK-NEXT: Domain :=
>> -; CHECK-NEXT: [N] -> { Stmt_sw_bb_5[i0] : 4*floor((-2 + i0)/4) = -2 + i0 and 2 <= i0 < N };
>> +; CHECK-NEXT: [N] -> { Stmt_sw_bb_5[i0] : 4*floor((-2 + i0)/4) = -2 + i0 and 0 <= i0 < N };
>> ; CHECK-NEXT: Schedule :=
>> ; CHECK-NEXT: [N] -> { Stmt_sw_bb_5[i0] -> [i0, 1] };
>> ; CHECK-NEXT: ReadAccess := [Reduction Type: +] [Scalar: 0]
>> @@ -62,22 +62,16 @@
>>
>> ; AST: if (1)
>> ;
>> -; AST: {
>> -; AST-NEXT: for (int c0 = 0; c0 < N - 3; c0 += 4) {
>> -; AST-NEXT: Stmt_sw_bb(c0);
>> +; AST: for (int c0 = 0; c0 < N; c0 += 4) {
>> +; AST-NEXT: Stmt_sw_bb(c0);
>> +; AST-NEXT: if (N >= c0 + 2) {
>> ; AST-NEXT: Stmt_sw_bb_1(c0 + 1);
>> -; AST-NEXT: Stmt_sw_bb_5(c0 + 2);
>> -; AST-NEXT: Stmt_sw_bb_9(c0 + 3);
>> -; AST-NEXT: }
>> -; AST-NEXT: if (N >= 1)
>> -; AST-NEXT: if (N % 4 >= 1) {
>> -; AST-NEXT: Stmt_sw_bb(-(N % 4) + N);
>> -; AST-NEXT: if (N % 4 >= 2) {
>> -; AST-NEXT: Stmt_sw_bb_1(-(N % 4) + N + 1);
>> -; AST-NEXT: if ((N - 3) % 4 == 0)
>> -; AST-NEXT: Stmt_sw_bb_5(N - 1);
>> -; AST-NEXT: }
>> +; AST-NEXT: if (N >= c0 + 3) {
>> +; AST-NEXT: Stmt_sw_bb_5(c0 + 2);
>> +; AST-NEXT: if (N >= c0 + 4)
>> +; AST-NEXT: Stmt_sw_bb_9(c0 + 3);
>> ; AST-NEXT: }
>> +; AST-NEXT: }
>> ; AST-NEXT: }
>> ;
>> ; AST: else
>>
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
> --
>
> Johannes Doerfert
> Researcher / PhD Student
>
> Compiler Design Lab (Prof. Hack)
> Saarland University, Computer Science
> Building E1.3, Room 4.31
>
> Tel. +49 (0)681 302-57521 : doerfert at cs.uni-saarland.de
> Fax. +49 (0)681 302-3065 : http://www.cdl.uni-saarland.de/people/doerfert
More information about the llvm-commits
mailing list