[llvm] [LAA] Analyze pointers forked by a phi (PR #65834)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Sep 8 22:12:55 PDT 2023
https://github.com/vfdff created https://github.com/llvm/llvm-project/pull/65834:
Given a function like the following: https://godbolt.org/z/T9c99fr88
```
1161_noReadWrite(int *Preds) {
for (int i = 0; i < LEN_1D-1; ++i) {
if (Preds[i] != 0)
b[i] = c[i] + 1;
else
a[i] = i * i;
}
}
```
LLVM will optimize the IR to a single store by a phi instruction:
```
%1 = load ptr, ptr @a, align 64
%2 = load ptr, ptr @b, align 64
...
for.inc:
%.sink = phi ptr [ %1, %if.then ], [ %2, %if.else ]
%add.sink = phi double [ %add, %if.then ], [ %conv8, %if.else ]
%arrayidx7 = getelementptr inbounds double, ptr %.sink, i64 %indvars.iv
store double %add.sink, ptr %arrayidx7, align 8
```
LAA is currently unable to analyze such IR, since ScalarEvolution will return a SCEVUnknown for the forked pointer operand of the store.
This patch adds initial optional support for analyzing both possibilities for the pointer and allowing LAA to generate runtime checks for the bounds if required, refers to D108699, but here address the phi node.
Fixes https://github.com/llvm/llvm-project/issues/64888
>From be921127972f133c801c76ec3ba1e2b063f11c43 Mon Sep 17 00:00:00 2001
From: Zhongyunde <zhongyunde at huawei.com>
Date: Sat, 9 Sep 2023 13:09:36 +0800
Subject: [PATCH] [LAA] Analyze pointers forked by a phi
Given a function like the following: https://godbolt.org/z/T9c99fr88
```
1161_noReadWrite(int *Preds) {
for (int i = 0; i < LEN_1D-1; ++i) {
if (Preds[i] != 0)
b[i] = c[i] + 1;
else
a[i] = i * i;
}
}
```
LLVM will optimize the IR to a single store by a phi instruction:
```
%1 = load ptr, ptr @a, align 64
%2 = load ptr, ptr @b, align 64
...
for.inc:
%.sink = phi ptr [ %1, %if.then ], [ %2, %if.else ]
%add.sink = phi double [ %add, %if.then ], [ %conv8, %if.else ]
%arrayidx7 = getelementptr inbounds double, ptr %.sink, i64 %indvars.iv
store double %add.sink, ptr %arrayidx7, align 8
```
LAA is currently unable to analyze such IR, since ScalarEvolution
will return a SCEVUnknown for the forked pointer operand of the store.
This patch adds initial optional support for analyzing both possibilities for the pointer
and allowing LAA to generate runtime checks for the bounds if required, refers to D108699, but here address the phi node.
Fixes https://github.com/llvm/llvm-project/issues/64888
---
llvm/lib/Analysis/LoopAccessAnalysis.cpp | 16 +++
.../LoopAccessAnalysis/forked-pointers.ll | 123 ++++++++++++++++++
2 files changed, 139 insertions(+)
diff --git a/llvm/lib/Analysis/LoopAccessAnalysis.cpp b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
index 4dd150492453f72..565cbfdc22d8466 100644
--- a/llvm/lib/Analysis/LoopAccessAnalysis.cpp
+++ b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
@@ -942,6 +942,22 @@ static void findForkedSCEVs(
ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(Ptr));
break;
}
+ case Instruction::PHI: {
+ SmallVector<PointerIntPair<const SCEV *, 1, bool>, 2> ChildScevs;
+ // A phi means we've found a forked pointer, but we currently only
+ // support a single phi per pointer so if there's another behind this
+ // then we just bail out and return the generic SCEV.
+ if (I->getNumOperands() == 2) {
+ findForkedSCEVs(SE, L, I->getOperand(0), ChildScevs, Depth);
+ findForkedSCEVs(SE, L, I->getOperand(1), ChildScevs, Depth);
+ }
+ if (ChildScevs.size() == 2) {
+ ScevList.push_back(ChildScevs[0]);
+ ScevList.push_back(ChildScevs[1]);
+ } else
+ ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(Ptr));
+ break;
+ }
case Instruction::Add:
case Instruction::Sub: {
SmallVector<PointerIntPair<const SCEV *, 1, bool>> LScevs;
diff --git a/llvm/test/Analysis/LoopAccessAnalysis/forked-pointers.ll b/llvm/test/Analysis/LoopAccessAnalysis/forked-pointers.ll
index 3f5c8e4c1c72ff1..30ffc4cb5e1e85f 100644
--- a/llvm/test/Analysis/LoopAccessAnalysis/forked-pointers.ll
+++ b/llvm/test/Analysis/LoopAccessAnalysis/forked-pointers.ll
@@ -936,3 +936,126 @@ for.body:
exit:
ret void
}
+
+define void @forked_ptrs_with_different_base(ptr nocapture readonly %Preds, ptr nocapture %a, ptr nocapture %b, ptr nocapture readonly %c) {
+; CHECK: for.body:
+; CHECK-NEXT: Memory dependences are safe with run-time checks
+; CHECK-NEXT: Dependences:
+; CHECK-NEXT: Run-time memory checks:
+; CHECK-NEXT: Check 0:
+; CHECK-NEXT: Comparing group ([[G1:.+]]):
+; CHECK-NEXT: %arrayidx7 = getelementptr inbounds double, ptr %.sink, i64 %indvars.iv
+; CHECK-NEXT: Against group ([[G2:.+]]):
+; CHECK-NEXT: %arrayidx = getelementptr inbounds i32, ptr %Preds, i64 %indvars.iv
+; CHECK-NEXT: Check 1:
+; CHECK-NEXT: Comparing group ([[G1]]):
+; CHECK-NEXT: %arrayidx7 = getelementptr inbounds double, ptr %.sink, i64 %indvars.iv
+; CHECK-NEXT: Against group ([[G4:.+]]):
+; CHECK-NEXT: %arrayidx5 = getelementptr inbounds double, ptr %0, i64 %indvars.iv
+; CHECK-NEXT: Check 2:
+; CHECK-NEXT: Comparing group ([[G3:.+]]):
+; CHECK-NEXT: %arrayidx7 = getelementptr inbounds double, ptr %.sink, i64 %indvars.iv
+; CHECK-NEXT: Against group ([[G2]]):
+; CHECK-NEXT: %arrayidx = getelementptr inbounds i32, ptr %Preds, i64 %indvars.iv
+; CHECK-NEXT: Check 3:
+; CHECK-NEXT: Comparing group ([[G3]]):
+; CHECK-NEXT: %arrayidx7 = getelementptr inbounds double, ptr %.sink, i64 %indvars.iv
+; CHECK-NEXT: Against group ([[G4]]):
+; CHECK-NEXT: %arrayidx5 = getelementptr inbounds double, ptr %0, i64 %indvars.iv
+; CHECK-NEXT: Grouped accesses:
+; CHECK-NEXT: Group [[G1]]:
+; CHECK-NEXT: (Low: %1 High: (63992 + %1))
+; CHECK-NEXT: Member: {%1,+,8}<nw><%for.body>
+; CHECK-NEXT: Group [[G3]]:
+; CHECK-NEXT: (Low: %2 High: (63992 + %2))
+; CHECK-NEXT: Member: {%2,+,8}<nw><%for.body>
+; CHECK-NEXT: Group [[G2]]:
+; CHECK-NEXT: (Low: %Preds High: (31996 + %Preds))
+; CHECK-NEXT: Member: {%Preds,+,4}<nuw><%for.body>
+; CHECK-NEXT: Group [[G4]]:
+; CHECK-NEXT: (Low: %0 High: (63992 + %0))
+; CHECK-NEXT: Member: {%0,+,8}<nw><%for.body>
+; CHECK-EMPTY:
+; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
+entry:
+ %0 = load ptr, ptr %c, align 64
+ %1 = load ptr, ptr %a, align 64
+ %2 = load ptr, ptr %b, align 64
+ br label %for.body
+
+for.cond.cleanup: ; preds = %for.inc
+ ret void
+
+for.body: ; preds = %entry, %for.inc
+ %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.inc ]
+ %arrayidx = getelementptr inbounds i32, ptr %Preds, i64 %indvars.iv
+ %3 = load i32, ptr %arrayidx, align 4
+ %cmp2.not = icmp eq i32 %3, 0
+ br i1 %cmp2.not, label %if.else, label %if.then
+
+if.then: ; preds = %for.body
+ %arrayidx5 = getelementptr inbounds double, ptr %0, i64 %indvars.iv
+ %4 = load double, ptr %arrayidx5, align 8
+ %add = fadd fast double %4, 1.000000e+00
+ br label %for.inc
+
+if.else: ; preds = %for.body
+ %5 = mul nuw nsw i64 %indvars.iv, %indvars.iv
+ %6 = trunc i64 %5 to i32
+ %conv8 = sitofp i32 %6 to double
+ br label %for.inc
+
+for.inc: ; preds = %if.then, %if.else
+ %.sink = phi ptr [ %1, %if.then ], [ %2, %if.else ]
+ %add.sink = phi double [ %add, %if.then ], [ %conv8, %if.else ]
+ %arrayidx7 = getelementptr inbounds double, ptr %.sink, i64 %indvars.iv
+ store double %add.sink, ptr %arrayidx7, align 8
+ %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+ %exitcond.not = icmp eq i64 %indvars.iv.next, 7999
+ br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
+}
+
+; Negative test: the operator number of PhiNode is not 2.
+define void @forked_ptrs_with_different_base3(ptr nocapture readonly %Preds, ptr nocapture %a, ptr nocapture %b, ptr nocapture readonly %c) {
+; CHECK: for.body:
+; CHECK-NEXT: Report: cannot identify array bounds
+; CHECK-NEXT: Dependences:
+; CHECK-NEXT: Run-time memory checks:
+; CHECK-NEXT: Grouped accesses:
+; CHECK-EMPTY:
+; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
+entry:
+ %ld.c = load ptr, ptr %c, align 64
+ %ld.a = load ptr, ptr %a, align 64
+ %ld.b = load ptr, ptr %b, align 64
+ br label %for.body
+
+for.body: ; preds = %entry, %for.inc
+ %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.inc ]
+ %arrayidx = getelementptr inbounds i32, ptr %Preds, i64 %indvars.iv
+ %ld.preds = load i32, ptr %arrayidx, align 4
+ switch i32 %ld.preds, label %if.else [
+ i32 0, label %if.br0
+ i32 1, label %if.br1
+ ]
+
+if.br0: ; preds = %for.body
+ br label %for.inc
+
+if.br1: ; preds = %for.body
+ br label %for.inc
+
+if.else: ; preds = %for.body
+ br label %for.inc
+
+for.inc: ; preds = %if.br1, %if.br0
+ %.sink = phi ptr [ %ld.a, %if.br0 ], [ %ld.b, %if.br1 ], [ %ld.c, %if.else ]
+ %arrayidx7 = getelementptr inbounds double, ptr %.sink, i64 %indvars.iv
+ store double 1.000000e+00, ptr %arrayidx7, align 8
+ %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+ %exitcond.not = icmp eq i64 %indvars.iv.next, 7999
+ br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
+
+for.cond.cleanup: ; preds = %for.inc
+ ret void
+}
\ No newline at end of file
More information about the llvm-commits
mailing list