[clang] [polly] [llvm] [polly][ScheduleOptimizer] Fix long compile time(hang) reported in polly (PR #75141)
Karthika Devi C via cfe-commits
cfe-commits at lists.llvm.org
Tue Dec 19 00:03:15 PST 2023
https://github.com/kartcq updated https://github.com/llvm/llvm-project/pull/75141
>From 98745d850c588eaf9b5d2ac6f71c79873107501f Mon Sep 17 00:00:00 2001
From: kartcq <quic_kartc at quicinc.com>
Date: Mon, 11 Dec 2023 05:22:33 -0800
Subject: [PATCH 1/4] [polly][ScheduleOptimizer] Bail out on exceeding Schedule
compute's ISL quota
There is no upper cap set on current Schedule Optimizer to compute schedule. In
some cases a very long compile time taken to compute the schedule resulting in
hang kind of behavior. This patch introduces a flag 'polly-schedule-computeout'
to pass the cap which is initialized to 300000. This patch handles the compute
out cases by bailing out and exiting gracefully.
Change-Id: Id506832df4ae8d3f140579ba10cf570e18efac62
---
polly/lib/Transform/ScheduleOptimizer.cpp | 25 +++++
.../ScheduleOptimizer/schedule_computeout.ll | 97 +++++++++++++++++++
2 files changed, 122 insertions(+)
create mode 100644 polly/test/ScheduleOptimizer/schedule_computeout.ll
diff --git a/polly/lib/Transform/ScheduleOptimizer.cpp b/polly/lib/Transform/ScheduleOptimizer.cpp
index 35a0a4def0403d..8ee2b66339adbc 100644
--- a/polly/lib/Transform/ScheduleOptimizer.cpp
+++ b/polly/lib/Transform/ScheduleOptimizer.cpp
@@ -96,6 +96,13 @@ static cl::opt<std::string>
cl::desc("Maximize the band depth (yes/no)"), cl::Hidden,
cl::init("yes"), cl::cat(PollyCategory));
+static cl::opt<int>
+ ScheduleComputeOut("polly-schedule-computeout",
+ cl::desc("Bound the scheduler by maximal amount"
+ "of computational steps. "),
+ cl::Hidden, cl::init(300000), cl::ZeroOrMore,
+ cl::cat(PollyCategory));
+
static cl::opt<bool>
GreedyFusion("polly-loopfusion-greedy",
cl::desc("Aggressively try to fuse everything"), cl::Hidden,
@@ -860,7 +867,25 @@ static void runIslScheduleOptimizer(
SC = SC.set_proximity(Proximity);
SC = SC.set_validity(Validity);
SC = SC.set_coincidence(Validity);
+
+ // Save error handling behavior
+ long MaxOperations = isl_ctx_get_max_operations(Ctx);
+ isl_ctx_set_max_operations(Ctx, ScheduleComputeOut);
Schedule = SC.compute_schedule();
+ bool ScheduleQuota = false;
+ if (isl_ctx_last_error(Ctx) == isl_error_quota) {
+ isl_ctx_reset_error(Ctx);
+ LLVM_DEBUG(
+ dbgs() << "Schedule optimizer calculation exceeds ISL quota\n");
+ ScheduleQuota = true;
+ }
+ isl_options_set_on_error(Ctx, ISL_ON_ERROR_ABORT);
+ isl_ctx_reset_operations(Ctx);
+ isl_ctx_set_max_operations(Ctx, MaxOperations);
+
+ if (ScheduleQuota)
+ return;
+
isl_options_set_on_error(Ctx, OnErrorStatus);
ScopsRescheduled++;
diff --git a/polly/test/ScheduleOptimizer/schedule_computeout.ll b/polly/test/ScheduleOptimizer/schedule_computeout.ll
new file mode 100644
index 00000000000000..3e768e02c8a740
--- /dev/null
+++ b/polly/test/ScheduleOptimizer/schedule_computeout.ll
@@ -0,0 +1,97 @@
+; RUN: opt -S -polly-optree -polly-delicm -polly-opt-isl -polly-schedule-computeout=100000 -debug-only="polly-opt-isl" < %s 2>&1 | FileCheck %s
+; Bailout if the computations of schedule compute exceeds the max scheduling quota.
+; Max compute out is initialized to 300000, Here it is set to 100000 for test purpose.
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64-unknown-linux-gnu"
+
+ at a = dso_local local_unnamed_addr global ptr null, align 8
+ at b = dso_local local_unnamed_addr global ptr null, align 8
+ at c = dso_local local_unnamed_addr global ptr null, align 8
+
+define dso_local void @foo(i32 noundef %I, i32 noundef %J, i32 noundef %K1, i32 noundef %K2, i32 noundef %L1, i32 noundef %L2) local_unnamed_addr {
+entry:
+ %j = alloca i32, align 4
+ store volatile i32 0, ptr %j, align 4
+ %j.0.j.0.j.0.54 = load volatile i32, ptr %j, align 4
+ %cmp55 = icmp slt i32 %j.0.j.0.j.0.54, %J
+ br i1 %cmp55, label %for.body.lr.ph, label %for.cond.cleanup
+
+for.body.lr.ph: ; preds = %entry
+ %0 = load ptr, ptr @a, align 8
+ %1 = load ptr, ptr @b, align 8
+ %2 = load ptr, ptr %1, align 8
+ %cmp352 = icmp slt i32 %L1, %L2
+ %cmp750 = icmp slt i32 %K1, %K2
+ %3 = sext i32 %K1 to i64
+ %4 = sext i32 %L1 to i64
+ br label %for.body
+
+for.cond.cleanup: ; preds = %for.cond.cleanup4, %entry
+ ret void
+
+for.body: ; preds = %for.cond.cleanup4, %for.body.lr.ph
+ br i1 %cmp352, label %for.cond6.preheader.preheader, label %for.cond.cleanup4
+
+for.cond6.preheader.preheader: ; preds = %for.body
+ %wide.trip.count66 = sext i32 %L2 to i64
+ br label %for.cond6.preheader
+
+for.cond6.preheader: ; preds = %for.cond.cleanup8, %for.cond6.preheader.preheader
+ %indvars.iv61 = phi i64 [ %4, %for.cond6.preheader.preheader ], [ %indvars.iv.next62, %for.cond.cleanup8 ]
+ br i1 %cmp750, label %for.cond10.preheader.lr.ph, label %for.cond.cleanup8
+
+for.cond10.preheader.lr.ph: ; preds = %for.cond6.preheader
+ %5 = mul nsw i64 %indvars.iv61, 516
+ %6 = mul nsw i64 %indvars.iv61, 516
+ %wide.trip.count = sext i32 %K2 to i64
+ br label %for.cond10.preheader
+
+for.cond.cleanup4: ; preds = %for.cond.cleanup8, %for.body
+ %j.0.j.0.j.0.45 = load volatile i32, ptr %j, align 4
+ %inc34 = add nsw i32 %j.0.j.0.j.0.45, 1
+ store volatile i32 %inc34, ptr %j, align 4
+ %j.0.j.0.j.0. = load volatile i32, ptr %j, align 4
+ %cmp = icmp slt i32 %j.0.j.0.j.0., %J
+ br i1 %cmp, label %for.body, label %for.cond.cleanup
+
+for.cond10.preheader: ; preds = %for.cond.cleanup12, %for.cond10.preheader.lr.ph
+ %indvars.iv = phi i64 [ %3, %for.cond10.preheader.lr.ph ], [ %indvars.iv.next, %for.cond.cleanup12 ]
+ %7 = getelementptr float, ptr %0, i64 %indvars.iv
+ %arrayidx18 = getelementptr float, ptr %7, i64 %5
+ %8 = load float, ptr %arrayidx18, align 4
+ br label %for.cond14.preheader
+
+for.cond.cleanup8: ; preds = %for.cond.cleanup12, %for.cond6.preheader
+ %indvars.iv.next62 = add nsw i64 %indvars.iv61, 1
+ %exitcond67.not = icmp eq i64 %indvars.iv.next62, %wide.trip.count66
+ br i1 %exitcond67.not, label %for.cond.cleanup4, label %for.cond6.preheader
+
+for.cond14.preheader: ; preds = %for.cond.cleanup16, %for.cond10.preheader
+ %m.049 = phi i32 [ -2, %for.cond10.preheader ], [ %inc21, %for.cond.cleanup16 ]
+ %sum.048 = phi float [ 0.000000e+00, %for.cond10.preheader ], [ %add19, %for.cond.cleanup16 ]
+ br label %for.body17
+
+for.cond.cleanup12: ; preds = %for.cond.cleanup16
+ %9 = getelementptr float, ptr %2, i64 %indvars.iv
+ %arrayidx26 = getelementptr float, ptr %9, i64 %6
+ store float %add19, ptr %arrayidx26, align 4
+ %indvars.iv.next = add nsw i64 %indvars.iv, 1
+ %exitcond60.not = icmp eq i64 %indvars.iv.next, %wide.trip.count
+ br i1 %exitcond60.not, label %for.cond.cleanup8, label %for.cond10.preheader
+
+for.cond.cleanup16: ; preds = %for.body17
+ %inc21 = add nsw i32 %m.049, 1
+ %exitcond56.not = icmp eq i32 %inc21, 3
+ br i1 %exitcond56.not, label %for.cond.cleanup12, label %for.cond14.preheader
+
+for.body17: ; preds = %for.body17, %for.cond14.preheader
+ %n.047 = phi i32 [ -2, %for.cond14.preheader ], [ %inc, %for.body17 ]
+ %sum.146 = phi float [ %sum.048, %for.cond14.preheader ], [ %add19, %for.body17 ]
+ %add19 = fadd float %sum.146, %8
+ %inc = add nsw i32 %n.047, 1
+ %exitcond.not = icmp eq i32 %inc, 3
+ br i1 %exitcond.not, label %for.cond.cleanup16, label %for.body17
+}
+
+; CHECK: Schedule optimizer calculation exceeds ISL quota
>From c1d5ca08f06fc2e2b55269ffbbf7dedad00a8bdb Mon Sep 17 00:00:00 2001
From: Karthika Devi C <quic_kartc at quicinc.com>
Date: Wed, 13 Dec 2023 00:32:10 +0530
Subject: [PATCH 2/4] Removed target details in test file.
---
polly/test/ScheduleOptimizer/schedule_computeout.ll | 3 ---
1 file changed, 3 deletions(-)
diff --git a/polly/test/ScheduleOptimizer/schedule_computeout.ll b/polly/test/ScheduleOptimizer/schedule_computeout.ll
index 3e768e02c8a740..db46ef5021932c 100644
--- a/polly/test/ScheduleOptimizer/schedule_computeout.ll
+++ b/polly/test/ScheduleOptimizer/schedule_computeout.ll
@@ -2,9 +2,6 @@
; Bailout if the computations of schedule compute exceeds the max scheduling quota.
; Max compute out is initialized to 300000, Here it is set to 100000 for test purpose.
-target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
-target triple = "aarch64-unknown-linux-gnu"
-
@a = dso_local local_unnamed_addr global ptr null, align 8
@b = dso_local local_unnamed_addr global ptr null, align 8
@c = dso_local local_unnamed_addr global ptr null, align 8
>From 7c7497d3217691cf8e4801a770698e12090e06a2 Mon Sep 17 00:00:00 2001
From: Karthika Devi C <quic_kartc at quicinc.com>
Date: Wed, 13 Dec 2023 20:44:59 +0530
Subject: [PATCH 3/4] Update ScheduleOptimizer.cpp
---
polly/lib/Transform/ScheduleOptimizer.cpp | 25 ++++++++---------------
1 file changed, 9 insertions(+), 16 deletions(-)
diff --git a/polly/lib/Transform/ScheduleOptimizer.cpp b/polly/lib/Transform/ScheduleOptimizer.cpp
index 8ee2b66339adbc..bd29e80e8792a7 100644
--- a/polly/lib/Transform/ScheduleOptimizer.cpp
+++ b/polly/lib/Transform/ScheduleOptimizer.cpp
@@ -868,23 +868,16 @@ static void runIslScheduleOptimizer(
SC = SC.set_validity(Validity);
SC = SC.set_coincidence(Validity);
- // Save error handling behavior
- long MaxOperations = isl_ctx_get_max_operations(Ctx);
- isl_ctx_set_max_operations(Ctx, ScheduleComputeOut);
- Schedule = SC.compute_schedule();
- bool ScheduleQuota = false;
- if (isl_ctx_last_error(Ctx) == isl_error_quota) {
- isl_ctx_reset_error(Ctx);
- LLVM_DEBUG(
- dbgs() << "Schedule optimizer calculation exceeds ISL quota\n");
- ScheduleQuota = true;
+ {
+ IslMaxOperationsGuard MaxOpGuard(Ctx, ScheduleComputeOut);
+ Schedule = SC.compute_schedule();
+
+ if (MaxOpGuard.hasQuotaExceeded()) {
+ LLVM_DEBUG(
+ dbgs() << "Schedule optimizer calculation exceeds ISL quota\n");
+ return;
+ }
}
- isl_options_set_on_error(Ctx, ISL_ON_ERROR_ABORT);
- isl_ctx_reset_operations(Ctx);
- isl_ctx_set_max_operations(Ctx, MaxOperations);
-
- if (ScheduleQuota)
- return;
isl_options_set_on_error(Ctx, OnErrorStatus);
>From 7d21eabcc58b5cf04433ab56968be54da906d03b Mon Sep 17 00:00:00 2001
From: Karthika Devi C <quic_kartc at quicinc.com>
Date: Tue, 19 Dec 2023 13:33:06 +0530
Subject: [PATCH 4/4] Update ScheduleOptimizer.cpp
---
polly/lib/Transform/ScheduleOptimizer.cpp | 4 +---
1 file changed, 1 insertion(+), 3 deletions(-)
diff --git a/polly/lib/Transform/ScheduleOptimizer.cpp b/polly/lib/Transform/ScheduleOptimizer.cpp
index bd29e80e8792a7..5a0ea3b4067542 100644
--- a/polly/lib/Transform/ScheduleOptimizer.cpp
+++ b/polly/lib/Transform/ScheduleOptimizer.cpp
@@ -872,11 +872,9 @@ static void runIslScheduleOptimizer(
IslMaxOperationsGuard MaxOpGuard(Ctx, ScheduleComputeOut);
Schedule = SC.compute_schedule();
- if (MaxOpGuard.hasQuotaExceeded()) {
+ if (MaxOpGuard.hasQuotaExceeded())
LLVM_DEBUG(
dbgs() << "Schedule optimizer calculation exceeds ISL quota\n");
- return;
- }
}
isl_options_set_on_error(Ctx, OnErrorStatus);
More information about the cfe-commits
mailing list