[llvm] 234cc40 - [LAA] Limit no-overlap check to at least one loop-invariant accesses.

[Florian Hahn via llvm-commits]

Author: Florian Hahn
Date: 2024-05-28T09:23:02-07:00
New Revision: 234cc40adc610a55d1a5a2fe798a9dd07b993f0c

URL: https://github.com/llvm/llvm-project/commit/234cc40adc610a55d1a5a2fe798a9dd07b993f0c
DIFF: https://github.com/llvm/llvm-project/commit/234cc40adc610a55d1a5a2fe798a9dd07b993f0c.diff

LOG: [LAA] Limit no-overlap check to at least one loop-invariant accesses.

Limit the logic added in https://github.com/llvm/llvm-project/pull/9230
to cases where either sink or source are loop-invariant, to avoid
compile-time increases. This is not needed for correctness.

I am working on follow-up changes to reduce the compile-time impact in
general to allow us to enable this again for any source/sink.

This should fix the compile-time regression introduced by this change:

* compile-time improvement with this change:
  https://llvm-compile-time-tracker.com/compare.php?from=4351787fb650da6d1bfb8d6e58753c90dcd4c418&to=b89010a2eb5f98494787c1c3b77f25208c59090c&stat=instructions:u

* compile-time improvement with original patch reverted on top of this
  change:
  https://llvm-compile-time-tracker.com/compare.php?from=b89010a2eb5f98494787c1c3b77f25208c59090c&to=19a1103fe68115cfd7d6472c6961f4fabe81a593&stat=instructions:u

Added: 
    

Modified: 
    llvm/lib/Analysis/LoopAccessAnalysis.cpp
    llvm/test/Analysis/LoopAccessAnalysis/depend_diff_types.ll
    llvm/test/Analysis/LoopAccessAnalysis/non-constant-strides-backward.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Analysis/LoopAccessAnalysis.cpp b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
index bc8b9b8479e4f..bd4c2a35ebf2c 100644
--- a/llvm/lib/Analysis/LoopAccessAnalysis.cpp
+++ b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
@@ -1983,20 +1983,25 @@ getDependenceDistanceStrideAndSize(
     return MemoryDepChecker::Dependence::IndirectUnsafe;
 
   // Check if we can prove that Sink only accesses memory after Src's end or
-  // vice versa.
-  const auto &[SrcStart, SrcEnd] =
-      getStartAndEndForAccess(InnermostLoop, Src, ATy, PSE);
-  const auto &[SinkStart, SinkEnd] =
-      getStartAndEndForAccess(InnermostLoop, Sink, BTy, PSE);
-
-  if (!isa<SCEVCouldNotCompute>(SrcStart) &&
-      !isa<SCEVCouldNotCompute>(SrcEnd) &&
-      !isa<SCEVCouldNotCompute>(SinkStart) &&
-      !isa<SCEVCouldNotCompute>(SinkEnd)) {
-    if (SE.isKnownPredicate(CmpInst::ICMP_ULE, SrcEnd, SinkStart))
-      return MemoryDepChecker::Dependence::NoDep;
-    if (SE.isKnownPredicate(CmpInst::ICMP_ULE, SinkEnd, SrcStart))
-      return MemoryDepChecker::Dependence::NoDep;
+  // vice versa. At the moment this is limited to cases where either source or
+  // sink are loop invariant to avoid compile-time increases. This is not
+  // required for correctness.
+  if (SE.isLoopInvariant(Src, InnermostLoop) ||
+      SE.isLoopInvariant(Sink, InnermostLoop)) {
+    const auto &[SrcStart, SrcEnd] =
+        getStartAndEndForAccess(InnermostLoop, Src, ATy, PSE);
+    const auto &[SinkStart, SinkEnd] =
+        getStartAndEndForAccess(InnermostLoop, Sink, BTy, PSE);
+
+    if (!isa<SCEVCouldNotCompute>(SrcStart) &&
+        !isa<SCEVCouldNotCompute>(SrcEnd) &&
+        !isa<SCEVCouldNotCompute>(SinkStart) &&
+        !isa<SCEVCouldNotCompute>(SinkEnd)) {
+      if (SE.isKnownPredicate(CmpInst::ICMP_ULE, SrcEnd, SinkStart))
+        return MemoryDepChecker::Dependence::NoDep;
+      if (SE.isKnownPredicate(CmpInst::ICMP_ULE, SinkEnd, SrcStart))
+        return MemoryDepChecker::Dependence::NoDep;
+    }
   }
 
   // Need accesses with constant strides and the same direction. We don't want

diff  --git a/llvm/test/Analysis/LoopAccessAnalysis/depend_
diff _types.ll b/llvm/test/Analysis/LoopAccessAnalysis/depend_
diff _types.ll
index 809b15b200495..81d8b01fe7fb7 100644
--- a/llvm/test/Analysis/LoopAccessAnalysis/depend_
diff _types.ll
+++ b/llvm/test/Analysis/LoopAccessAnalysis/depend_
diff _types.ll
@@ -130,8 +130,16 @@ define void @neg_dist_dep_type_size_equivalence(ptr nocapture %vec, i64 %n) {
 ; CHECK-LABEL: 'neg_dist_dep_type_size_equivalence'
 ; CHECK-NEXT:    loop:
 ; CHECK-NEXT:      Report: unsafe dependent memory operations in loop. Use #pragma clang loop distribute(enable) to allow loop distribution to attempt to isolate the offending operations into a separate loop
-; CHECK-NEXT:  Backward loop carried data dependence that prevents store-to-load forwarding.
+; CHECK-NEXT:  Unknown data dependence.
 ; CHECK-NEXT:      Dependences:
+; CHECK-NEXT:        Unknown:
+; CHECK-NEXT:            %ld.f64 = load double, ptr %gep.iv, align 8 ->
+; CHECK-NEXT:            store i32 %ld.i64.i32, ptr %gep.iv.n.i64, align 8
+; CHECK-EMPTY:
+; CHECK-NEXT:        Unknown:
+; CHECK-NEXT:            %ld.i64 = load i64, ptr %gep.iv, align 8 ->
+; CHECK-NEXT:            store i32 %ld.i64.i32, ptr %gep.iv.n.i64, align 8
+; CHECK-EMPTY:
 ; CHECK-NEXT:        BackwardVectorizableButPreventsForwarding:
 ; CHECK-NEXT:            %ld.f64 = load double, ptr %gep.iv, align 8 ->
 ; CHECK-NEXT:            store double %val, ptr %gep.iv.101.i64, align 8

diff  --git a/llvm/test/Analysis/LoopAccessAnalysis/non-constant-strides-backward.ll b/llvm/test/Analysis/LoopAccessAnalysis/non-constant-strides-backward.ll
index 845ff078ee0eb..416742a94e0d3 100644
--- a/llvm/test/Analysis/LoopAccessAnalysis/non-constant-strides-backward.ll
+++ b/llvm/test/Analysis/LoopAccessAnalysis/non-constant-strides-backward.ll
@@ -45,8 +45,13 @@ exit:
 define void @
diff erent_non_constant_strides_known_backward_distance_larger_than_trip_count(ptr %A) {
 ; CHECK-LABEL: '
diff erent_non_constant_strides_known_backward_distance_larger_than_trip_count'
 ; CHECK-NEXT:    loop:
-; CHECK-NEXT:      Memory dependences are safe
+; CHECK-NEXT:      Report: unsafe dependent memory operations in loop. Use #pragma clang loop distribute(enable) to allow loop distribution to attempt to isolate the offending operations into a separate loop
+; CHECK-NEXT:  Unknown data dependence.
 ; CHECK-NEXT:      Dependences:
+; CHECK-NEXT:        Unknown:
+; CHECK-NEXT:            %l = load i32, ptr %gep, align 4 ->
+; CHECK-NEXT:            store i32 %add, ptr %gep.mul.2, align 4
+; CHECK-EMPTY:
 ; CHECK-NEXT:      Run-time memory checks:
 ; CHECK-NEXT:      Grouped accesses:
 ; CHECK-EMPTY: