rpms/llvm
6
0
Fork 0
Code Issues Pull Requests Releases Activity
9cb7e1045d
llvm/0001-CGP-Bail-out-if-Base-Scaled-Reg-does-not-dominate-in.patch
RHEL Packaging Agent 9cb7e1045d Rebase to LLVM 21.1.3 
Successfully rebased llvm package from version 20.1.8 to 21.1.3 by
syncing from Fedora rawhide branch to CentOS c9s using the
centos-sync.sh script. All patches applied cleanly during centpkg prep,
sources were uploaded to lookaside cache, and SRPM was generated
successfully.

Resolves: RHEL-100898

This commit was created by Jotnar, a Red Hat Enterprise Linux software maintenance AI agent.

Assisted-by: Jotnar
2025-10-21 13:13:09 +00:00

132 lines
5.5 KiB
Diff
Raw Blame History

From dde30a47313bf52fef02bbcb1de931a8d725659f Mon Sep 17 00:00:00 2001
From: Florian Hahn <flo@fhahn.com>
Date: Fri, 6 Jun 2025 12:38:30 +0100
Subject: [PATCH] [CGP] Bail out if (Base|Scaled)Reg does not dominate insert
point. (#142949)
(Base|Scaled)Reg may not dominate the chosen insert point, if there are
multiple uses of the address. Bail out if that's the case, otherwise we
will generate invalid IR.
In some cases, we could probably adjust the insert point or hoist the
(Base|Scaled)Reg.
Fixes https://github.com/llvm/llvm-project/issues/142830.
PR: https://github.com/llvm/llvm-project/pull/142949
---
llvm/lib/CodeGen/CodeGenPrepare.cpp | 13 +++-
.../X86/sink-addrmode-reg-does-not-geps.ll | 76 +++++++++++++++++++
2 files changed, 87 insertions(+), 2 deletions(-)
create mode 100644 llvm/test/Transforms/CodeGenPrepare/X86/sink-addrmode-reg-does-not-geps.ll
diff --git a/llvm/lib/CodeGen/CodeGenPrepare.cpp b/llvm/lib/CodeGen/CodeGenPrepare.cpp
index 822ed6283117..32348a899683 100644
--- a/llvm/lib/CodeGen/CodeGenPrepare.cpp
+++ b/llvm/lib/CodeGen/CodeGenPrepare.cpp
@@ -5945,8 +5945,17 @@ bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
// The current BB may be optimized multiple times, we can't guarantee the
// reuse of Addr happens later, call findInsertPos to find an appropriate
// insert position.
- IRBuilder<> Builder(MemoryInst->getParent(),
- findInsertPos(Addr, MemoryInst, SunkAddr));
+ auto InsertPos = findInsertPos(Addr, MemoryInst, SunkAddr);
+
+ // TODO: Adjust insert point considering (Base|Scaled)Reg if possible.
+ if (!SunkAddr) {
+ auto &DT = getDT(*MemoryInst->getFunction());
+ if ((AddrMode.BaseReg && !DT.dominates(AddrMode.BaseReg, &*InsertPos)) ||
+ (AddrMode.ScaledReg && !DT.dominates(AddrMode.ScaledReg, &*InsertPos)))
+ return Modified;
+ }
+
+ IRBuilder<> Builder(MemoryInst->getParent(), InsertPos);
if (SunkAddr) {
LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
diff --git a/llvm/test/Transforms/CodeGenPrepare/X86/sink-addrmode-reg-does-not-geps.ll b/llvm/test/Transforms/CodeGenPrepare/X86/sink-addrmode-reg-does-not-geps.ll
new file mode 100644
index 000000000000..1640bafbd0bf
--- /dev/null
+++ b/llvm/test/Transforms/CodeGenPrepare/X86/sink-addrmode-reg-does-not-geps.ll
@@ -0,0 +1,76 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -passes='require<profile-summary>,function(codegenprepare)' %s | FileCheck %s
+
+target triple = "x86_64-unknown-linux"
+
+declare i1 @cond(float)
+
+define void @scaled_reg_does_not_dominate_insert_point(ptr %src) {
+; CHECK-LABEL: define void @scaled_reg_does_not_dominate_insert_point(
+; CHECK-SAME: ptr [[SRC:%.*]]) {
+; CHECK-NEXT: [[BB:.*]]:
+; CHECK-NEXT: br label %[[LOOP:.*]]
+; CHECK: [[LOOP]]:
+; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[BB]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 1
+; CHECK-NEXT: [[SUNKADDR2:%.*]] = mul i64 [[IV_NEXT]], 2
+; CHECK-NEXT: [[SUNKADDR3:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[SUNKADDR2]]
+; CHECK-NEXT: [[SUNKADDR4:%.*]] = getelementptr i8, ptr [[SUNKADDR3]], i64 6
+; CHECK-NEXT: [[L_0:%.*]] = load float, ptr [[SUNKADDR4]], align 4
+; CHECK-NEXT: [[SUNKADDR:%.*]] = mul i64 [[IV]], 2
+; CHECK-NEXT: [[SUNKADDR1:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[SUNKADDR]]
+; CHECK-NEXT: [[L_1:%.*]] = load float, ptr [[SUNKADDR1]], align 4
+; CHECK-NEXT: [[TMP0:%.*]] = call i1 @cond(float [[L_0]])
+; CHECK-NEXT: [[C:%.*]] = call i1 @cond(float [[L_1]])
+; CHECK-NEXT: br i1 [[C]], label %[[LOOP]], label %[[EXIT:.*]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+bb:
+ %gep.base = getelementptr i8, ptr %src, i64 8
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %bb ], [ %iv.next, %loop ]
+ %iv.shl = shl i64 %iv, 1
+ %gep.shl = getelementptr i8, ptr %gep.base, i64 %iv.shl
+ %gep.sub = getelementptr i8, ptr %gep.shl, i64 -8
+ %iv.next = add i64 %iv, 1
+ %l.0 = load float, ptr %gep.shl, align 4
+ %l.1 = load float, ptr %gep.sub, align 4
+ call i1 @cond(float %l.0)
+ %c = call i1 @cond(float %l.1)
+ br i1 %c, label %loop, label %exit
+
+exit:
+ ret void
+}
+
+define void @check_dt_after_modifying_cfg(ptr %dst, i64 %x, i8 %y, i8 %z) {
+; CHECK-LABEL: define void @check_dt_after_modifying_cfg(
+; CHECK-SAME: ptr [[DST:%.*]], i64 [[X:%.*]], i8 [[Y:%.*]], i8 [[Z:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*]]:
+; CHECK-NEXT: [[OFFSET:%.*]] = lshr i64 [[X]], 2
+; CHECK-NEXT: [[SEL_FROZEN:%.*]] = freeze i8 [[Z]]
+; CHECK-NEXT: [[CMP:%.*]] = icmp slt i8 [[SEL_FROZEN]], 0
+; CHECK-NEXT: br i1 [[CMP]], label %[[SELECT_END:.*]], label %[[SELECT_FALSE_SINK:.*]]
+; CHECK: [[SELECT_FALSE_SINK]]:
+; CHECK-NEXT: [[SMIN:%.*]] = tail call i8 @llvm.smin.i8(i8 [[Y]], i8 0)
+; CHECK-NEXT: br label %[[SELECT_END]]
+; CHECK: [[SELECT_END]]:
+; CHECK-NEXT: [[SEL:%.*]] = phi i8 [ 0, %[[ENTRY]] ], [ [[SMIN]], %[[SELECT_FALSE_SINK]] ]
+; CHECK-NEXT: [[SUNKADDR:%.*]] = getelementptr i8, ptr [[DST]], i64 [[OFFSET]]
+; CHECK-NEXT: store i8 [[SEL]], ptr [[SUNKADDR]], align 1
+; CHECK-NEXT: ret void
+;
+entry:
+ %offset = lshr i64 %x, 2
+ %gep.dst = getelementptr i8, ptr %dst, i64 %offset
+ %smin = tail call i8 @llvm.smin.i8(i8 %y, i8 0)
+ %cmp = icmp slt i8 %z, 0
+ %sel = select i1 %cmp, i8 0, i8 %smin
+ store i8 %sel, ptr %gep.dst, align 1
+ ret void
+}
+
+declare i8 @llvm.smin.i8(i8, i8) #0
--
2.50.1
Reference in New Issue View Git Blame Copy Permalink