[llvm] [HashRecognize] Rewrite arePHIsIntertwined (PR #144878)
Ramkumar Ramachandra via llvm-commits
llvm-commits at lists.llvm.org
Mon Jun 23 04:17:01 PDT 2025
https://github.com/artagnon updated https://github.com/llvm/llvm-project/pull/144878
>From 85a8fdd91f5fd74c8ded09648d736249461c57c1 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Thu, 19 Jun 2025 11:29:47 +0100
Subject: [PATCH 1/3] [HashRecognize] Pre-commit tests for PHIsIntertwined
---
.../HashRecognize/cyclic-redundancy-check.ll | 58 ++++++++++++++++++-
1 file changed, 56 insertions(+), 2 deletions(-)
diff --git a/llvm/test/Analysis/HashRecognize/cyclic-redundancy-check.ll b/llvm/test/Analysis/HashRecognize/cyclic-redundancy-check.ll
index 0366684a13b54..1fa76cb23173b 100644
--- a/llvm/test/Analysis/HashRecognize/cyclic-redundancy-check.ll
+++ b/llvm/test/Analysis/HashRecognize/cyclic-redundancy-check.ll
@@ -144,6 +144,34 @@ exit: ; preds = %loop
ret i16 %crc.next
}
+define i8 @crc8.le.tc16(i16 %msg, i8 %checksum) {
+; CHECK-LABEL: 'crc8.le.tc16'
+; CHECK-NEXT: Did not find a hash algorithm
+; CHECK-NEXT: Reason: Simple recurrence doesn't use conditional recurrence with XOR
+;
+entry:
+ br label %loop
+
+loop: ; preds = %loop, %entry
+ %iv = phi i8 [ 0, %entry ], [ %iv.next, %loop ]
+ %crc = phi i8 [ %checksum, %entry ], [ %crc.next, %loop ]
+ %data = phi i16 [ %msg, %entry ], [ %data.next, %loop ]
+ %data.trunc = trunc i16 %data to i8
+ %xor.crc.data = xor i8 %crc, %data.trunc
+ %and.crc.data = and i8 %xor.crc.data, 1
+ %data.next = lshr i16 %data, 1
+ %check.sb = icmp eq i8 %and.crc.data, 0
+ %crc.lshr = lshr i8 %crc, 1
+ %crc.xor = xor i8 %crc.lshr, 29
+ %crc.next = select i1 %check.sb, i8 %crc.lshr, i8 %crc.xor
+ %iv.next = add nuw nsw i8 %iv, 1
+ %exit.cond = icmp samesign ult i8 %iv, 15
+ br i1 %exit.cond, label %loop, label %exit
+
+exit: ; preds = %loop
+ ret i8 %crc.next
+}
+
define i16 @crc16.be.tc8.crc.init.li(i16 %checksum, i8 %msg) {
; CHECK-LABEL: 'crc16.be.tc8.crc.init.li'
; CHECK-NEXT: Found big-endian CRC-16 loop with trip count 8
@@ -610,9 +638,8 @@ loop: ; preds = %loop, %entry
%iv = phi i8 [ 0, %entry ], [ %iv.next, %loop ]
%data = phi i8 [ %msg, %entry ], [ %data.next, %loop ]
%crc = phi i16 [ %checksum, %entry ], [ %crc.next, %loop ]
- %crc.lshr = lshr i16 %crc, 8
%data.ext = zext i8 %data to i16
- %xor.crc.data = xor i16 %crc.lshr, %data.ext
+ %xor.crc.data = xor i16 %crc, %data.ext
%check.sb = icmp samesign ult i16 %xor.crc.data, 128
%crc.shl = shl i16 %crc, 1
%crc.xor = xor i16 %crc.shl, 258
@@ -838,6 +865,33 @@ exit: ; preds = %loop
ret i16 %crc.next
}
+define i16 @not.crc.bad.cast(i8 %msg, i16 %checksum) {
+; CHECK-LABEL: 'not.crc.bad.cast'
+; CHECK-NEXT: Did not find a hash algorithm
+; CHECK-NEXT: Reason: Simple recurrence doesn't use conditional recurrence with XOR
+;
+entry:
+ br label %loop
+
+loop: ; preds = %loop, %entry
+ %iv = phi i8 [ 0, %entry ], [ %iv.next, %loop ]
+ %data = phi i8 [ %msg, %entry ], [ %data.next, %loop ]
+ %crc = phi i16 [ %checksum, %entry ], [ %crc.next, %loop ]
+ %data.ext = zext i8 %data to i16
+ %xor.crc.data = xor i16 %crc, %data.ext
+ %check.sb = icmp slt i16 %xor.crc.data, 0
+ %crc.shl = shl i16 %crc, 1
+ %crc.xor = xor i16 %crc.shl, 29
+ %crc.next = select i1 %check.sb, i16 %crc.shl, i16 %crc.xor
+ %data.next = shl i8 %data, 1
+ %iv.next = add nuw nsw i8 %iv, 1
+ %exit.cond = icmp samesign ult i8 %iv, 7
+ br i1 %exit.cond, label %loop, label %exit
+
+exit: ; preds = %loop
+ ret i16 %crc.next
+}
+
define i32 @not.crc.dead.msg.bad.use(i32 %checksum, i32 %msg) {
; CHECK-LABEL: 'not.crc.dead.msg.bad.use'
; CHECK-NEXT: Did not find a hash algorithm
>From c0c26be2bf35c443f732e8d4e978ebbc2155b641 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Thu, 19 Jun 2025 12:07:55 +0100
Subject: [PATCH 2/3] [HashRecognize] Rewrite arePHIsIntertwined
The test crc8.le.tc16 is a valid CRC algorithm, but isn't recognized as
such due to a buggy arePHIsIntertwined, which is asymmetric in its PHI
node arguments. Another issue is that arePHIsIntertwined is
unnecessarily complicated: the core functionality is to match a BinOp
that's a recurrence in both PHI nodes, ignoring zext and trunc casts.
Hence, rewrite it for clarity, and make it symmetric in both PHI
operands.
crc8.le.tc16 is still not recognized as a valid CRC algorithm, due to an
incorrect check for loop iterations exceeding the bitwidth of the
result: in reality, it should not exceed the bitwidth of LHSAux, but we
leave this fix to a follow-up.
---
llvm/include/llvm/IR/PatternMatch.h | 7 ++
llvm/lib/Analysis/HashRecognize.cpp | 66 ++++++++-----------
.../HashRecognize/cyclic-redundancy-check.ll | 12 ++--
3 files changed, 42 insertions(+), 43 deletions(-)
diff --git a/llvm/include/llvm/IR/PatternMatch.h b/llvm/include/llvm/IR/PatternMatch.h
index 1f86cdfd94e17..1c34ce6058b8a 100644
--- a/llvm/include/llvm/IR/PatternMatch.h
+++ b/llvm/include/llvm/IR/PatternMatch.h
@@ -2168,6 +2168,13 @@ m_ZExtOrSelf(const OpTy &Op) {
return m_CombineOr(m_ZExt(Op), Op);
}
+template <typename OpTy>
+inline match_combine_or<match_combine_or<CastInst_match<OpTy, ZExtInst>, OpTy>,
+ CastInst_match<OpTy, TruncInst>>
+m_ZExtOrTruncOrSelf(const OpTy &Op) {
+ return m_CombineOr(m_ZExtOrSelf(Op), m_Trunc(Op));
+}
+
template <typename OpTy>
inline match_combine_or<CastInst_match<OpTy, SExtInst>, OpTy>
m_SExtOrSelf(const OpTy &Op) {
diff --git a/llvm/lib/Analysis/HashRecognize.cpp b/llvm/lib/Analysis/HashRecognize.cpp
index 1a5dfbe4bed6e..6adb5eee4dcca 100644
--- a/llvm/lib/Analysis/HashRecognize.cpp
+++ b/llvm/lib/Analysis/HashRecognize.cpp
@@ -497,45 +497,37 @@ CRCTable HashRecognize::genSarwateTable(const APInt &GenPoly,
return Table;
}
-/// Checks if \p Reference is reachable from \p Needle on the use-def chain, and
-/// that there are no stray PHI nodes while digging the use-def chain. \p
-/// BOToMatch is a CRC peculiarity: at least one of the Users of Needle needs to
-/// match this OpCode, which is XOR for CRC.
+/// Checks that \p Needle and \p Reference are used together in a
+/// BinaryOperator, which is a recurrence over both, ignoring zext and trunc.
+/// Additionally checks that the BinaryOperator has opcode \p BOToMatch, which
+/// is XOR in the case of CRC. In other words, it checks for the following
+/// pattern:
+///
+/// loop:
+/// %needle = phi [_, %entry], [%needle.next, %loop]
+/// %reference = phi [_, %entry], [%reference.next, %loop]
+/// ...
+/// _ = BOTOMatch ((trunc|zext|self) %needle) ((trunc|zext|self) %reference)
+///
static bool arePHIsIntertwined(
const PHINode *Needle, const PHINode *Reference, const Loop &L,
Instruction::BinaryOps BOToMatch = Instruction::BinaryOpsEnd) {
- // Initialize the worklist with Users of the Needle.
- SmallVector<const Instruction *> Worklist;
- for (const User *U : Needle->users()) {
- if (auto *UI = dyn_cast<Instruction>(U))
- if (L.contains(UI))
- Worklist.push_back(UI);
- }
-
- // BOToMatch is usually XOR for CRC.
- if (BOToMatch != Instruction::BinaryOpsEnd) {
- if (count_if(Worklist, [BOToMatch](const Instruction *I) {
- return I->getOpcode() == BOToMatch;
- }) != 1)
- return false;
- }
-
- while (!Worklist.empty()) {
- const Instruction *I = Worklist.pop_back_val();
-
- // Since Needle is never pushed onto the Worklist, I must either be the
- // Reference PHI node (in which case we're done), or a stray PHI node (in
- // which case we abort).
- if (isa<PHINode>(I))
- return I == Reference;
-
- for (const Use &U : I->operands())
- if (auto *UI = dyn_cast<Instruction>(U))
- // Don't push Needle back onto the Worklist.
- if (UI != Needle && L.contains(UI))
- Worklist.push_back(UI);
- }
- return false;
+ return any_of(ArrayRef({Needle, Reference}), [&](const PHINode *S) {
+ return count_if(S->users(), [&](const User *U) {
+ auto *BO = dyn_cast<BinaryOperator>(U);
+ if (!BO)
+ return false;
+ if (BOToMatch != Instruction::BinaryOpsEnd &&
+ BO->getOpcode() != BOToMatch)
+ return false;
+ return all_of(BO->operands(), [Needle, Reference](const Use &U) {
+ return match(
+ U.get(),
+ m_CombineOr(m_ZExtOrTruncOrSelf(m_Specific(Reference)),
+ m_ZExtOrTruncOrSelf(m_Specific(Needle))));
+ });
+ }) == 1;
+ });
}
// Recognizes a multiplication or division by the constant two, using SCEV. By
@@ -588,7 +580,7 @@ HashRecognize::recognizeCRC() const {
return "Loop with non-unit bitshifts";
if (!arePHIsIntertwined(SimpleRecurrence.Phi, ConditionalRecurrence.Phi, L,
Instruction::BinaryOps::Xor))
- return "Simple recurrence doesn't use conditional recurrence with XOR";
+ return "Recurrences not intertwined with XOR";
}
// Make sure that the computed value is used in the exit block: this should be
diff --git a/llvm/test/Analysis/HashRecognize/cyclic-redundancy-check.ll b/llvm/test/Analysis/HashRecognize/cyclic-redundancy-check.ll
index 1fa76cb23173b..c42985037d83a 100644
--- a/llvm/test/Analysis/HashRecognize/cyclic-redundancy-check.ll
+++ b/llvm/test/Analysis/HashRecognize/cyclic-redundancy-check.ll
@@ -147,7 +147,7 @@ exit: ; preds = %loop
define i8 @crc8.le.tc16(i16 %msg, i8 %checksum) {
; CHECK-LABEL: 'crc8.le.tc16'
; CHECK-NEXT: Did not find a hash algorithm
-; CHECK-NEXT: Reason: Simple recurrence doesn't use conditional recurrence with XOR
+; CHECK-NEXT: Reason: Loop iterations exceed bitwidth of result
;
entry:
br label %loop
@@ -629,7 +629,7 @@ exit: ; preds = %loop
define i16 @not.crc.wrong.sb.check.const(i8 %msg, i16 %checksum) {
; CHECK-LABEL: 'not.crc.wrong.sb.check.const'
; CHECK-NEXT: Did not find a hash algorithm
-; CHECK-NEXT: Reason: Simple recurrence doesn't use conditional recurrence with XOR
+; CHECK-NEXT: Reason: Bad RHS of significant-bit-check
;
entry:
br label %loop
@@ -868,7 +868,7 @@ exit: ; preds = %loop
define i16 @not.crc.bad.cast(i8 %msg, i16 %checksum) {
; CHECK-LABEL: 'not.crc.bad.cast'
; CHECK-NEXT: Did not find a hash algorithm
-; CHECK-NEXT: Reason: Simple recurrence doesn't use conditional recurrence with XOR
+; CHECK-NEXT: Reason: Expected bottom 8 bits zero (????????00001011)
;
entry:
br label %loop
@@ -895,7 +895,7 @@ exit: ; preds = %loop
define i32 @not.crc.dead.msg.bad.use(i32 %checksum, i32 %msg) {
; CHECK-LABEL: 'not.crc.dead.msg.bad.use'
; CHECK-NEXT: Did not find a hash algorithm
-; CHECK-NEXT: Reason: Simple recurrence doesn't use conditional recurrence with XOR
+; CHECK-NEXT: Reason: Recurrences not intertwined with XOR
;
entry:
br label %loop
@@ -923,7 +923,7 @@ exit: ; preds = %loop
define i16 @not.crc.dead.msg.no.use(i8 %msg, i16 %checksum) {
; CHECK-LABEL: 'not.crc.dead.msg.no.use'
; CHECK-NEXT: Did not find a hash algorithm
-; CHECK-NEXT: Reason: Simple recurrence doesn't use conditional recurrence with XOR
+; CHECK-NEXT: Reason: Recurrences not intertwined with XOR
;
entry:
br label %loop
@@ -952,7 +952,7 @@ exit: ; preds = %loop
define i32 @not.crc.dead.msg.wrong.op(i32 %checksum, i32 %msg) {
; CHECK-LABEL: 'not.crc.dead.msg.wrong.op'
; CHECK-NEXT: Did not find a hash algorithm
-; CHECK-NEXT: Reason: Simple recurrence doesn't use conditional recurrence with XOR
+; CHECK-NEXT: Reason: Recurrences not intertwined with XOR
;
entry:
br label %loop
>From 55b69f98235c74201b35bb01b0e859264fe74120 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Mon, 23 Jun 2025 12:13:28 +0100
Subject: [PATCH 3/3] [HashRecognize] Strip unnecessary generality
---
llvm/lib/Analysis/HashRecognize.cpp | 22 ++++++++++------------
1 file changed, 10 insertions(+), 12 deletions(-)
diff --git a/llvm/lib/Analysis/HashRecognize.cpp b/llvm/lib/Analysis/HashRecognize.cpp
index 6adb5eee4dcca..09cb6ef4ba4d6 100644
--- a/llvm/lib/Analysis/HashRecognize.cpp
+++ b/llvm/lib/Analysis/HashRecognize.cpp
@@ -497,11 +497,9 @@ CRCTable HashRecognize::genSarwateTable(const APInt &GenPoly,
return Table;
}
-/// Checks that \p Needle and \p Reference are used together in a
-/// BinaryOperator, which is a recurrence over both, ignoring zext and trunc.
-/// Additionally checks that the BinaryOperator has opcode \p BOToMatch, which
-/// is XOR in the case of CRC. In other words, it checks for the following
-/// pattern:
+/// Checks that \p Needle and \p Reference are used together in a BinaryOperator
+/// with opcode \p BOToMatch, which is a recurrence over both, ignoring zext and
+/// trunc. In other words, it checks for the following pattern:
///
/// loop:
/// %needle = phi [_, %entry], [%needle.next, %loop]
@@ -509,16 +507,16 @@ CRCTable HashRecognize::genSarwateTable(const APInt &GenPoly,
/// ...
/// _ = BOTOMatch ((trunc|zext|self) %needle) ((trunc|zext|self) %reference)
///
-static bool arePHIsIntertwined(
- const PHINode *Needle, const PHINode *Reference, const Loop &L,
- Instruction::BinaryOps BOToMatch = Instruction::BinaryOpsEnd) {
+/// where at least one cast is self.
+///
+/// \p BOToMatch is XOR in the case of CRC.
+static bool arePHIsIntertwined(const PHINode *Needle, const PHINode *Reference,
+ const Loop &L,
+ Instruction::BinaryOps BOToMatch) {
return any_of(ArrayRef({Needle, Reference}), [&](const PHINode *S) {
return count_if(S->users(), [&](const User *U) {
auto *BO = dyn_cast<BinaryOperator>(U);
- if (!BO)
- return false;
- if (BOToMatch != Instruction::BinaryOpsEnd &&
- BO->getOpcode() != BOToMatch)
+ if (!BO || BO->getOpcode() != BOToMatch)
return false;
return all_of(BO->operands(), [Needle, Reference](const Use &U) {
return match(
More information about the llvm-commits
mailing list