[llvm] b06f10d - [DA] Add check for base pointer invariance (#148241)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Jul 25 11:25:04 PDT 2025
Author: Ryotaro Kasuga
Date: 2025-07-26T03:25:01+09:00
New Revision: b06f10d96c6d0fb89253c75c7f1d75c4cf519339
URL: https://github.com/llvm/llvm-project/commit/b06f10d96c6d0fb89253c75c7f1d75c4cf519339
DIFF: https://github.com/llvm/llvm-project/commit/b06f10d96c6d0fb89253c75c7f1d75c4cf519339.diff
LOG: [DA] Add check for base pointer invariance (#148241)
As specified in #53942, DA assumes base pointer invariance in its
process. Some cases were fixed by #116628. However, that PR only
addressed the parts related to AliasAnalysis, so the original issue
persists in later stages, especially when the AliasAnalysis results in
`MustAlias`.
This patch insert an explicit loop-invariant checks for the base pointer
and skips analysis when it is not loop-invariant.
Fix the cases added in #148240.
Added:
Modified:
llvm/lib/Analysis/DependenceAnalysis.cpp
llvm/test/Analysis/DependenceAnalysis/Banerjee.ll
llvm/test/Analysis/DependenceAnalysis/FlipFlopBaseAddress.ll
llvm/test/Analysis/DependenceAnalysis/GCD.ll
llvm/test/Analysis/DependenceAnalysis/NonAffineExpr.ll
llvm/test/Analysis/DependenceAnalysis/Preliminary.ll
llvm/test/Analysis/DependenceAnalysis/PreliminaryNoValidityCheckFixedSize.ll
llvm/test/Analysis/DependenceAnalysis/SymbolicRDIV.ll
Removed:
################################################################################
diff --git a/llvm/lib/Analysis/DependenceAnalysis.cpp b/llvm/lib/Analysis/DependenceAnalysis.cpp
index dd9a44b9aecac..f1473b2694ca4 100644
--- a/llvm/lib/Analysis/DependenceAnalysis.cpp
+++ b/llvm/lib/Analysis/DependenceAnalysis.cpp
@@ -3383,6 +3383,10 @@ bool DependenceInfo::tryDelinearize(Instruction *Src, Instruction *Dst,
SrcSubscripts, DstSubscripts))
return false;
+ assert(isLoopInvariant(SrcBase, SrcLoop) &&
+ isLoopInvariant(DstBase, DstLoop) &&
+ "Expected SrcBase and DstBase to be loop invariant");
+
int Size = SrcSubscripts.size();
LLVM_DEBUG({
dbgs() << "\nSrcSubscripts: ";
@@ -3666,6 +3670,19 @@ DependenceInfo::depends(Instruction *Src, Instruction *Dst,
SCEVUnionPredicate(Assume, *SE));
}
+ // Even if the base pointers are the same, they may not be loop-invariant. It
+ // could lead to incorrect results, as we're analyzing loop-carried
+ // dependencies. Src and Dst can be in
diff erent loops, so we need to check
+ // the base pointer is invariant in both loops.
+ Loop *SrcLoop = LI->getLoopFor(Src->getParent());
+ Loop *DstLoop = LI->getLoopFor(Dst->getParent());
+ if (!isLoopInvariant(SrcBase, SrcLoop) ||
+ !isLoopInvariant(DstBase, DstLoop)) {
+ LLVM_DEBUG(dbgs() << "The base pointer is not loop invariant.\n");
+ return std::make_unique<Dependence>(Src, Dst,
+ SCEVUnionPredicate(Assume, *SE));
+ }
+
uint64_t EltSize = SrcLoc.Size.toRaw();
const SCEV *SrcEv = SE->getMinusSCEV(SrcSCEV, SrcBase);
const SCEV *DstEv = SE->getMinusSCEV(DstSCEV, DstBase);
diff --git a/llvm/test/Analysis/DependenceAnalysis/Banerjee.ll b/llvm/test/Analysis/DependenceAnalysis/Banerjee.ll
index d3301520fd107..e0def901d1759 100644
--- a/llvm/test/Analysis/DependenceAnalysis/Banerjee.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/Banerjee.ll
@@ -113,7 +113,7 @@ define void @banerjee1(ptr %A, ptr %B, i64 %m, i64 %n) nounwind uwtable ssp {
; CHECK-NEXT: Src: %2 = load i64, ptr %arrayidx6, align 8 --> Dst: store i64 %2, ptr %B.addr.12, align 8
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store i64 %2, ptr %B.addr.12, align 8 --> Dst: store i64 %2, ptr %B.addr.12, align 8
-; CHECK-NEXT: da analyze - output [* *]!
+; CHECK-NEXT: da analyze - confused!
;
; NORMALIZE-LABEL: 'banerjee1'
; NORMALIZE-NEXT: Src: store i64 0, ptr %arrayidx, align 8 --> Dst: store i64 0, ptr %arrayidx, align 8
@@ -127,7 +127,7 @@ define void @banerjee1(ptr %A, ptr %B, i64 %m, i64 %n) nounwind uwtable ssp {
; NORMALIZE-NEXT: Src: %2 = load i64, ptr %arrayidx6, align 8 --> Dst: store i64 %2, ptr %B.addr.12, align 8
; NORMALIZE-NEXT: da analyze - confused!
; NORMALIZE-NEXT: Src: store i64 %2, ptr %B.addr.12, align 8 --> Dst: store i64 %2, ptr %B.addr.12, align 8
-; NORMALIZE-NEXT: da analyze - output [* *]!
+; NORMALIZE-NEXT: da analyze - confused!
;
; DELIN-LABEL: 'banerjee1'
; DELIN-NEXT: Src: store i64 0, ptr %arrayidx, align 8 --> Dst: store i64 0, ptr %arrayidx, align 8
@@ -141,7 +141,7 @@ define void @banerjee1(ptr %A, ptr %B, i64 %m, i64 %n) nounwind uwtable ssp {
; DELIN-NEXT: Src: %2 = load i64, ptr %arrayidx6, align 8 --> Dst: store i64 %2, ptr %B.addr.12, align 8
; DELIN-NEXT: da analyze - confused!
; DELIN-NEXT: Src: store i64 %2, ptr %B.addr.12, align 8 --> Dst: store i64 %2, ptr %B.addr.12, align 8
-; DELIN-NEXT: da analyze - output [* *]!
+; DELIN-NEXT: da analyze - confused!
;
entry:
%cmp4 = icmp sgt i64 %n, 0
diff --git a/llvm/test/Analysis/DependenceAnalysis/FlipFlopBaseAddress.ll b/llvm/test/Analysis/DependenceAnalysis/FlipFlopBaseAddress.ll
index 3e3426afab0f7..bf2268b746a6f 100644
--- a/llvm/test/Analysis/DependenceAnalysis/FlipFlopBaseAddress.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/FlipFlopBaseAddress.ll
@@ -8,11 +8,11 @@
define float @bug41488_test1(float %f) {
; CHECK-LABEL: 'bug41488_test1'
; CHECK-NEXT: Src: %0 = load float, ptr %p, align 4 --> Dst: %0 = load float, ptr %p, align 4
-; CHECK-NEXT: da analyze - input [*]!
+; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: %0 = load float, ptr %p, align 4 --> Dst: store float %f, ptr %q, align 4
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store float %f, ptr %q, align 4 --> Dst: store float %f, ptr %q, align 4
-; CHECK-NEXT: da analyze - output [*]!
+; CHECK-NEXT: da analyze - confused!
;
entry:
%g = alloca float, align 4
@@ -34,11 +34,11 @@ for.cond.cleanup:
define void @bug41488_test2(i32 %n) {
; CHECK-LABEL: 'bug41488_test2'
; CHECK-NEXT: Src: %0 = load float, ptr %p, align 4 --> Dst: %0 = load float, ptr %p, align 4
-; CHECK-NEXT: da analyze - input [*]!
+; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: %0 = load float, ptr %p, align 4 --> Dst: store float 0.000000e+00, ptr %q, align 4
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store float 0.000000e+00, ptr %q, align 4 --> Dst: store float 0.000000e+00, ptr %q, align 4
-; CHECK-NEXT: da analyze - output [*]!
+; CHECK-NEXT: da analyze - confused!
;
entry:
%g = alloca float, align 4
@@ -68,7 +68,7 @@ define void @bug53942_foo(i32 noundef %n, ptr noalias nocapture noundef writeonl
; CHECK-NEXT: Src: %.pre = load double, ptr %B, align 8 --> Dst: store double %.pre, ptr %arrayidx2, align 8
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store double %.pre, ptr %arrayidx2, align 8 --> Dst: store double %.pre, ptr %arrayidx2, align 8
-; CHECK-NEXT: da analyze - output [*]!
+; CHECK-NEXT: da analyze - confused!
;
entry:
%cmp8 = icmp sgt i32 %n, 1
@@ -99,11 +99,11 @@ for.body: ; preds = %for.body.preheader,
define void @bug53942_bar(i32 noundef %n, ptr noalias noundef %A, ptr noalias noundef %B) {
; CHECK-LABEL: 'bug53942_bar'
; CHECK-NEXT: Src: %0 = load double, ptr %arrayidx, align 8 --> Dst: %0 = load double, ptr %arrayidx, align 8
-; CHECK-NEXT: da analyze - input [*]!
+; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: %0 = load double, ptr %arrayidx, align 8 --> Dst: store double %0, ptr %arrayidx8, align 8
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store double %0, ptr %arrayidx8, align 8 --> Dst: store double %0, ptr %arrayidx8, align 8
-; CHECK-NEXT: da analyze - output [*]!
+; CHECK-NEXT: da analyze - confused!
;
entry:
br label %for.cond
@@ -166,14 +166,14 @@ for.end: ; preds = %for.cond.cleanup
; (j % 2 == 0 ? A[i][j] : A[i][j+1]) = 1;
; }
;
-; FIXME: There are loop-carried dependencies between the store instruction. For
+; There are loop-carried dependencies between the store instruction. For
; example, the value of %ptr0 when (i, j) = (0, 1) is %A+8, which is the same
; as when (i, j) = (0, 2).
define void @non_invariant_baseptr_with_identical_obj(ptr %A) {
; CHECK-LABEL: 'non_invariant_baseptr_with_identical_obj'
; CHECK-NEXT: Src: store i32 1, ptr %idx, align 4 --> Dst: store i32 1, ptr %idx, align 4
-; CHECK-NEXT: da analyze - none!
+; CHECK-NEXT: da analyze - confused!
;
entry:
br label %loop.i.header
@@ -216,13 +216,13 @@ exit:
; Similar to the above case, but ptr0 is loop-invariant with respsect to the
; k-loop.
;
-; FIXME: Same as the above case, there are loop-carried dependencies between
-; the store.
+; Same as the above case, there are loop-carried dependencies between the
+; store.
define void @non_invariant_baseptr_with_identical_obj2(ptr %A) {
; CHECK-LABEL: 'non_invariant_baseptr_with_identical_obj2'
; CHECK-NEXT: Src: store i32 1, ptr %idx, align 4 --> Dst: store i32 1, ptr %idx, align 4
-; CHECK-NEXT: da analyze - none!
+; CHECK-NEXT: da analyze - confused!
;
entry:
br label %loop.i.header
@@ -259,3 +259,58 @@ loop.i.latch:
exit:
ret void
}
+
+; Pseudo-code that is approximately semantically equivalent to the below IR:
+;
+; void f(int A[][32]) {
+; for (int i = 0; i < 100; i++)
+; for (int j = 0; j < 15; j++) {
+; int offset = (j % 2 == 0) ? 1 : 0;
+; A[i][2 * j + offset + 0] = 1;
+; A[i][2 * j + offset + 1] = 1;
+; }
+; }
+;
+; There are loop-carried dependencies between the two stores. For example,
+; A[0][2] is accessed from both the former one when (i, j) = (0, 1) and the
+; latter one when (i, j) = (0, 0).
+;
+define void @non_invariant_baseptr_with_identical_obj3(ptr %A) {
+; CHECK-LABEL: 'non_invariant_baseptr_with_identical_obj3'
+; CHECK-NEXT: Src: store i32 1, ptr %idx0, align 4 --> Dst: store i32 1, ptr %idx0, align 4
+; CHECK-NEXT: da analyze - confused!
+; CHECK-NEXT: Src: store i32 1, ptr %idx0, align 4 --> Dst: store i32 1, ptr %idx1, align 4
+; CHECK-NEXT: da analyze - confused!
+; CHECK-NEXT: Src: store i32 1, ptr %idx1, align 4 --> Dst: store i32 1, ptr %idx1, align 4
+; CHECK-NEXT: da analyze - confused!
+;
+entry:
+ br label %loop.i.header
+
+loop.i.header:
+ %i = phi i32 [ 0, %entry ], [ %i.inc, %loop.i.latch ]
+ %A1 = getelementptr i32, ptr %A, i32 1
+ br label %loop.j
+
+loop.j:
+ %j = phi i32 [ 0, %loop.i.header ], [ %j.inc, %loop.j ]
+ %ptr0 = phi ptr [ %A1, %loop.i.header ], [ %ptr1, %loop.j ]
+ %ptr1 = phi ptr [ %A, %loop.i.header ], [ %ptr0, %loop.j ]
+ %j2_0 = shl i32 %j, 1
+ %j2_1 = add i32 %j2_0, 1
+ %idx0 = getelementptr [32 x i32], ptr %ptr0, i32 %i, i32 %j2_0
+ %idx1 = getelementptr [32 x i32], ptr %ptr0, i32 %i, i32 %j2_1
+ store i32 1, ptr %idx0
+ store i32 1, ptr %idx1
+ %j.inc = add i32 %j, 1
+ %cmp.j = icmp slt i32 %j.inc, 15
+ br i1 %cmp.j, label %loop.j, label %loop.i.latch
+
+loop.i.latch:
+ %i.inc = add i32 %i, 1
+ %cmp.i = icmp slt i32 %i.inc, 100
+ br i1 %cmp.i, label %loop.i.header, label %exit
+
+exit:
+ ret void
+}
diff --git a/llvm/test/Analysis/DependenceAnalysis/GCD.ll b/llvm/test/Analysis/DependenceAnalysis/GCD.ll
index c0e1362c82b50..03343e7a98211 100644
--- a/llvm/test/Analysis/DependenceAnalysis/GCD.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/GCD.ll
@@ -398,7 +398,7 @@ define void @gcd6(i64 %n, ptr %A, ptr %B) nounwind uwtable ssp {
; CHECK-NEXT: Src: %2 = load i32, ptr %arrayidx9, align 4 --> Dst: store i32 %2, ptr %B.addr.12, align 4
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store i32 %2, ptr %B.addr.12, align 4 --> Dst: store i32 %2, ptr %B.addr.12, align 4
-; CHECK-NEXT: da analyze - output [* *]!
+; CHECK-NEXT: da analyze - confused!
;
entry:
%cmp4 = icmp sgt i64 %n, 0
@@ -475,7 +475,7 @@ define void @gcd7(i32 %n, ptr %A, ptr %B) nounwind uwtable ssp {
; CHECK-NEXT: Src: %11 = load i32, ptr %arrayidx12, align 4 --> Dst: store i32 %11, ptr %B.addr.12, align 4
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store i32 %11, ptr %B.addr.12, align 4 --> Dst: store i32 %11, ptr %B.addr.12, align 4
-; CHECK-NEXT: da analyze - output [* *]!
+; CHECK-NEXT: da analyze - confused!
;
entry:
%0 = zext i32 %n to i64
@@ -566,7 +566,7 @@ define void @gcd8(i32 %n, ptr %A, ptr %B) nounwind uwtable ssp {
; CHECK-NEXT: Src: %5 = load i32, ptr %arrayidx12, align 4 --> Dst: store i32 %5, ptr %B.addr.12, align 4
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store i32 %5, ptr %B.addr.12, align 4 --> Dst: store i32 %5, ptr %B.addr.12, align 4
-; CHECK-NEXT: da analyze - output [* *]!
+; CHECK-NEXT: da analyze - confused!
;
entry:
%cmp4 = icmp sgt i32 %n, 0
@@ -650,7 +650,7 @@ define void @gcd9(i32 %n, ptr %A, ptr %B) nounwind uwtable ssp {
; CHECK-NEXT: Src: %11 = load i32, ptr %arrayidx12, align 4 --> Dst: store i32 %11, ptr %B.addr.12, align 4
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store i32 %11, ptr %B.addr.12, align 4 --> Dst: store i32 %11, ptr %B.addr.12, align 4
-; CHECK-NEXT: da analyze - output [* *]!
+; CHECK-NEXT: da analyze - confused!
;
entry:
%0 = zext i32 %n to i64
diff --git a/llvm/test/Analysis/DependenceAnalysis/NonAffineExpr.ll b/llvm/test/Analysis/DependenceAnalysis/NonAffineExpr.ll
index d983bd49d49d6..3e110acbefb20 100644
--- a/llvm/test/Analysis/DependenceAnalysis/NonAffineExpr.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/NonAffineExpr.ll
@@ -12,7 +12,7 @@ define void @f(ptr %a, i32 %n, i1 %arg) align 2 {
; CHECK-NEXT: Src: %t.2 = load ptr, ptr %a, align 4 --> Dst: %t.4 = load i32, ptr %t.3, align 4
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: %t.4 = load i32, ptr %t.3, align 4 --> Dst: %t.4 = load i32, ptr %t.3, align 4
-; CHECK-NEXT: da analyze - input [* *]!
+; CHECK-NEXT: da analyze - confused!
;
for.preheader:
%t.0 = ashr exact i32 %n, 3
diff --git a/llvm/test/Analysis/DependenceAnalysis/Preliminary.ll b/llvm/test/Analysis/DependenceAnalysis/Preliminary.ll
index 4ab87771417e5..8cb0e2ac770dc 100644
--- a/llvm/test/Analysis/DependenceAnalysis/Preliminary.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/Preliminary.ll
@@ -69,7 +69,7 @@ define void @p2(i64 %n, ptr %A, ptr %B) nounwind uwtable ssp {
; CHECK-NEXT: Src: %0 = load i64, ptr %arrayidx17, align 8 --> Dst: store i64 %0, ptr %B.addr.24, align 8
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store i64 %0, ptr %B.addr.24, align 8 --> Dst: store i64 %0, ptr %B.addr.24, align 8
-; CHECK-NEXT: da analyze - output [* * *]!
+; CHECK-NEXT: da analyze - confused!
;
entry:
%cmp10 = icmp sgt i64 %n, 0
diff --git a/llvm/test/Analysis/DependenceAnalysis/PreliminaryNoValidityCheckFixedSize.ll b/llvm/test/Analysis/DependenceAnalysis/PreliminaryNoValidityCheckFixedSize.ll
index 404018707c0a5..e67cae7d39a75 100644
--- a/llvm/test/Analysis/DependenceAnalysis/PreliminaryNoValidityCheckFixedSize.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/PreliminaryNoValidityCheckFixedSize.ll
@@ -28,7 +28,7 @@ define void @p2(i64 %n, ptr %A, ptr %B) nounwind uwtable ssp {
; CHECK-NEXT: Src: %0 = load i64, ptr %arrayidx17, align 8 --> Dst: store i64 %0, ptr %B.addr.24, align 8
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store i64 %0, ptr %B.addr.24, align 8 --> Dst: store i64 %0, ptr %B.addr.24, align 8
-; CHECK-NEXT: da analyze - output [* * *]!
+; CHECK-NEXT: da analyze - confused!
;
; LIN-LABEL: 'p2'
; LIN-NEXT: Src: store i64 %i.011, ptr %arrayidx8, align 8 --> Dst: store i64 %i.011, ptr %arrayidx8, align 8
@@ -42,7 +42,7 @@ define void @p2(i64 %n, ptr %A, ptr %B) nounwind uwtable ssp {
; LIN-NEXT: Src: %0 = load i64, ptr %arrayidx17, align 8 --> Dst: store i64 %0, ptr %B.addr.24, align 8
; LIN-NEXT: da analyze - confused!
; LIN-NEXT: Src: store i64 %0, ptr %B.addr.24, align 8 --> Dst: store i64 %0, ptr %B.addr.24, align 8
-; LIN-NEXT: da analyze - output [* * *]!
+; LIN-NEXT: da analyze - confused!
;
entry:
%cmp10 = icmp sgt i64 %n, 0
diff --git a/llvm/test/Analysis/DependenceAnalysis/SymbolicRDIV.ll b/llvm/test/Analysis/DependenceAnalysis/SymbolicRDIV.ll
index f64a7483d7b66..8b9aa257a7c57 100644
--- a/llvm/test/Analysis/DependenceAnalysis/SymbolicRDIV.ll
+++ b/llvm/test/Analysis/DependenceAnalysis/SymbolicRDIV.ll
@@ -437,7 +437,7 @@ define void @symbolicrdiv6(ptr %A, ptr %B, i64 %n1, i64 %n2) nounwind uwtable ss
; CHECK-NEXT: Src: %0 = load i32, ptr %arrayidx4, align 4 --> Dst: store i32 %0, ptr %B.addr.12, align 4
; CHECK-NEXT: da analyze - confused!
; CHECK-NEXT: Src: store i32 %0, ptr %B.addr.12, align 4 --> Dst: store i32 %0, ptr %B.addr.12, align 4
-; CHECK-NEXT: da analyze - output [* *]!
+; CHECK-NEXT: da analyze - confused!
;
entry:
%cmp4 = icmp eq i64 %n1, 0
More information about the llvm-commits
mailing list