[llvm] Ensure propagated constants in the vtable are aligned #136630
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-llvm-transforms Author: None (PiJoules) ChangesIt's possible for virtual constant propagation in whole program devirtualization to create unaligned loads. We originally saw this with 4-byte aligned relative vtables where we could store 8-byte values before/after the vtable. But since the vtable is 4-byte aligned and we unconditionally do an 8-byte load, we can't guarantee that the stored constant will always be aligned to 8 bytes. We can also see this with normal vtables whenever a 1-byte char is stored in the vtable because the offset calculation for the GEP doesn't take into account the original vtable alignment. This patch introduces two changes to virtual constant propagation:
This patch updates tests to reflect this alignment change and adds some cases for relative vtables. Patch is 56.76 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/136630.diff 7 Files Affected:
diff --git a/llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp b/llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp
index 0680b2c558807..b9f5e5d893c9a 100644
--- a/llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp
+++ b/llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp
@@ -226,6 +226,10 @@ struct PatternList {
return false;
}
};
+
+inline uint64_t RoundUp(uint64_t Size, uint64_t Align) {
+ return (Size + (Align - 1)) / Align * Align;
+}
} // namespace
// Find the minimum offset that we may store a value of size Size bits at. If
@@ -298,7 +302,9 @@ wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets,
++Byte;
}
}
- return (MinByte + I) * 8;
+ // Rounding up ensures the constant is always stored at address we
+ // can directly load from without misalignment.
+ return RoundUp((MinByte + I) * 8, Size);
NextI:;
}
}
@@ -1834,9 +1840,19 @@ bool DevirtModule::tryVirtualConstProp(
if (!RetType)
return false;
unsigned BitWidth = RetType->getBitWidth();
+
+ // TODO: Since we can evaluated these constants at compile-time, we can save
+ // some space by calculating the smallest range of values that all these
+ // constants can fit in, then only allocate enough space to fit those values.
+ // At each callsite, we can get the original type by doing a sign/zero
+ // extension. For example, if we would store an i64, but we can see that all
+ // the values fit into an i16, then we can store an i16 before/after the
+ // vtable and at each callsite do a s/zext.
if (BitWidth > 64)
return false;
+ Align TypeAlignment = M.getDataLayout().getPrefTypeAlign(RetType);
+
// Make sure that each function is defined, does not access memory, takes at
// least one argument, does not use its first argument (which we assume is
// 'this'), and has the same return type.
@@ -1861,6 +1877,18 @@ bool DevirtModule::tryVirtualConstProp(
Fn->arg_empty() || !Fn->arg_begin()->use_empty() ||
Fn->getReturnType() != RetType)
return false;
+
+ // This only works if the integer size is at most the alignment of the
+ // vtable. If the table is underaligned, then we can't guarantee that the
+ // constant will always be aligned to the integer type alignment. For
+ // example, if the table is `align 1`, we can never guarantee that an i32
+ // stored before/after the vtable is 32-bit aligned without changing the
+ // alignment of the new global.
+ GlobalVariable *GV = Target.TM->Bits->GV;
+ Align TableAlignment = M.getDataLayout().getValueOrABITypeAlignment(
+ GV->getAlign(), GV->getValueType());
+ if (TypeAlignment > TableAlignment)
+ return false;
}
for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) {
@@ -1880,6 +1908,12 @@ bool DevirtModule::tryVirtualConstProp(
// Find an allocation offset in bits in all vtables associated with the
// type.
+ // TODO: If a series of i1s would be placed after the vtable, it would
+ // help save some space if they were placed at the very end after all
+ // other larger-size constants. Having these i1s anywhere in the middle
+ // of the allocation would mean extra padding is needed for any subsequent
+ // constants, but having them at the end wouldn't require this padding
+ // at the very end.
uint64_t AllocBefore =
findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth);
uint64_t AllocAfter =
@@ -1911,6 +1945,14 @@ bool DevirtModule::tryVirtualConstProp(
setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte,
OffsetBit);
+ // In an earlier check we forbade constant propagation from operating on
+ // tables whose alignment is less than the alignment needed for loading
+ // the constant. Thus, the address we take the offset from will always be
+ // aligned to at least this integer alignment. Now, we need to ensure that
+ // the offset is also aligned to this integer alignment to ensure we always
+ // have an aligned load.
+ assert(OffsetByte % TypeAlignment.value() == 0);
+
if (RemarksEnabled || AreStatisticsEnabled())
for (auto &&Target : TargetsForSlot)
Target.WasDevirt = true;
diff --git a/llvm/test/Transforms/WholeProgramDevirt/virtual-const-prop-begin.ll b/llvm/test/Transforms/WholeProgramDevirt/virtual-const-prop-begin.ll
index 11273804e2081..3ed8027522788 100644
--- a/llvm/test/Transforms/WholeProgramDevirt/virtual-const-prop-begin.ll
+++ b/llvm/test/Transforms/WholeProgramDevirt/virtual-const-prop-begin.ll
@@ -1,34 +1,35 @@
; RUN: opt -S -passes=wholeprogramdevirt -whole-program-visibility %s | FileCheck %s
target datalayout = "e-p:64:64"
-target triple = "x86_64-unknown-linux-gnu"
-; CHECK: [[VT1DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [0 x i8] } { [8 x i8] c"\00\00\00\01\00\00\00\02", [3 x ptr] [ptr @vf0i1, ptr @vf1i1, ptr @vf1i32], [0 x i8] zeroinitializer }, section "vt1sec", !type [[T8:![0-9]+]]
+;; Note that i16 is used here such that we can ensure all constants for "typeid"
+;; can come before the vtable.
+; CHECK: [[VT1DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [0 x i8] } { [8 x i8] c"\00\00\00\00\03\00\00\02", [3 x ptr] [ptr @vf0i1, ptr @vf1i1, ptr @vf1i16], [0 x i8] zeroinitializer }, section "vt1sec", !type [[T8:![0-9]+]]
@vt1 = constant [3 x ptr] [
ptr @vf0i1,
ptr @vf1i1,
-ptr @vf1i32
+ptr @vf1i16
], section "vt1sec", !type !0
-; CHECK: [[VT2DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [0 x i8] } { [8 x i8] c"\00\00\00\02\00\00\00\01", [3 x ptr] [ptr @vf1i1, ptr @vf0i1, ptr @vf2i32], [0 x i8] zeroinitializer }, !type [[T8]]
+; CHECK: [[VT2DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [0 x i8] } { [8 x i8] c"\00\00\00\00\04\00\00\01", [3 x ptr] [ptr @vf1i1, ptr @vf0i1, ptr @vf2i16], [0 x i8] zeroinitializer }, !type [[T8]]
@vt2 = constant [3 x ptr] [
ptr @vf1i1,
ptr @vf0i1,
-ptr @vf2i32
+ptr @vf2i16
], !type !0
-; CHECK: [[VT3DATA:@[^ ]*]] = private constant { [5 x i8], [3 x ptr], [0 x i8] } { [5 x i8] c"\03\00\00\00\02", [3 x ptr] [ptr @vf0i1, ptr @vf1i1, ptr @vf3i32], [0 x i8] zeroinitializer }, align 1, !type [[T5:![0-9]+]]
+; CHECK: [[VT3DATA:@[^ ]*]] = private constant { [4 x i8], [3 x ptr], [0 x i8] } { [4 x i8] c"\05\00\00\02", [3 x ptr] [ptr @vf0i1, ptr @vf1i1, ptr @vf3i16], [0 x i8] zeroinitializer }, align 2, !type [[T5:![0-9]+]]
@vt3 = constant [3 x ptr] [
ptr @vf0i1,
ptr @vf1i1,
-ptr @vf3i32
-], align 1, !type !0
+ptr @vf3i16
+], align 2, !type !0
-; CHECK: [[VT4DATA:@[^ ]*]] = private constant { [16 x i8], [3 x ptr], [0 x i8] } { [16 x i8] c"\00\00\00\00\00\00\00\00\00\00\00\04\00\00\00\01", [3 x ptr] [ptr @vf1i1, ptr @vf0i1, ptr @vf4i32], [0 x i8] zeroinitializer }, align 16, !type [[T16:![0-9]+]]
+; CHECK: [[VT4DATA:@[^ ]*]] = private constant { [16 x i8], [3 x ptr], [0 x i8] } { [16 x i8] c"\00\00\00\00\00\00\00\00\00\00\00\00\06\00\00\01", [3 x ptr] [ptr @vf1i1, ptr @vf0i1, ptr @vf4i16], [0 x i8] zeroinitializer }, align 16, !type [[T16:![0-9]+]]
@vt4 = constant [3 x ptr] [
ptr @vf1i1,
ptr @vf0i1,
-ptr @vf4i32
+ptr @vf4i16
], align 16, !type !0
; CHECK: @vt5 = {{.*}}, !type [[T0:![0-9]+]]
@@ -38,10 +39,35 @@ ptr @__cxa_pure_virtual,
ptr @__cxa_pure_virtual
], !type !0
+;; Test relative vtables
+; CHECK: [[VT6RELDATA:@[^ ]*]] = private constant { [4 x i8], [3 x i32], [0 x i8] } { [4 x i8] c"\00\00\03\00", [3 x i32] [
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf0i1 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32),
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf1i1 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32),
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf1i16 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32)
+; CHECK-SAME: ], [0 x i8] zeroinitializer }, !type [[TREL:![0-9]+]]
+@vt6_rel = constant [3 x i32] [
+i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf0i1 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32),
+i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf1i1 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32),
+i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf1i16 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32)
+], !type !2
+
+; CHECK: [[VT7RELDATA:@[^ ]*]] = private constant { [4 x i8], [3 x i32], [0 x i8] } { [4 x i8] c"\00\00\04\00", [3 x i32] [
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf1i1 to i64), i64 ptrtoint (ptr @vt7_rel to i64)) to i32),
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf0i1 to i64), i64 ptrtoint (ptr @vt7_rel to i64)) to i32),
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf2i16 to i64), i64 ptrtoint (ptr @vt7_rel to i64)) to i32)
+; CHECK-SAME: ], [0 x i8] zeroinitializer }, !type [[TREL]]
+@vt7_rel = constant [3 x i32] [
+i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf1i1 to i64), i64 ptrtoint (ptr @vt7_rel to i64)) to i32),
+i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf0i1 to i64), i64 ptrtoint (ptr @vt7_rel to i64)) to i32),
+i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf2i16 to i64), i64 ptrtoint (ptr @vt7_rel to i64)) to i32)
+], !type !2
+
; CHECK: @vt1 = alias [3 x ptr], getelementptr inbounds ({ [8 x i8], [3 x ptr], [0 x i8] }, ptr [[VT1DATA]], i32 0, i32 1)
; CHECK: @vt2 = alias [3 x ptr], getelementptr inbounds ({ [8 x i8], [3 x ptr], [0 x i8] }, ptr [[VT2DATA]], i32 0, i32 1)
-; CHECK: @vt3 = alias [3 x ptr], getelementptr inbounds ({ [5 x i8], [3 x ptr], [0 x i8] }, ptr [[VT3DATA]], i32 0, i32 1)
+; CHECK: @vt3 = alias [3 x ptr], getelementptr inbounds ({ [4 x i8], [3 x ptr], [0 x i8] }, ptr [[VT3DATA]], i32 0, i32 1)
; CHECK: @vt4 = alias [3 x ptr], getelementptr inbounds ({ [16 x i8], [3 x ptr], [0 x i8] }, ptr [[VT4DATA]], i32 0, i32 1)
+; CHECK: @vt6_rel = alias [3 x i32], getelementptr inbounds ({ [4 x i8], [3 x i32], [0 x i8] }, ptr [[VT6RELDATA]], i32 0, i32 1)
+; CHECK: @vt7_rel = alias [3 x i32], getelementptr inbounds ({ [4 x i8], [3 x i32], [0 x i8] }, ptr [[VT7RELDATA]], i32 0, i32 1)
define i1 @vf0i1(ptr %this) readnone {
ret i1 0
@@ -51,20 +77,20 @@ define i1 @vf1i1(ptr %this) readnone {
ret i1 1
}
-define i32 @vf1i32(ptr %this) readnone {
- ret i32 1
+define i16 @vf1i16(ptr %this) readnone {
+ ret i16 3
}
-define i32 @vf2i32(ptr %this) readnone {
- ret i32 2
+define i16 @vf2i16(ptr %this) readnone {
+ ret i16 4
}
-define i32 @vf3i32(ptr %this) readnone {
- ret i32 3
+define i16 @vf3i16(ptr %this) readnone {
+ ret i16 5
}
-define i32 @vf4i32(ptr %this) readnone {
- ret i32 4
+define i16 @vf4i16(ptr %this) readnone {
+ ret i16 6
}
; CHECK: define i1 @call1(
@@ -87,7 +113,7 @@ define i1 @call2(ptr %obj) {
%vtable = load ptr, ptr %obj
%p = call i1 @llvm.type.test(ptr %vtable, metadata !"typeid")
call void @llvm.assume(i1 %p)
- %fptrptr = getelementptr [3 x ptr], ptr %vtable, i32 0, i32 1
+ %fptrptr = getelementptr [3 x ptr], ptr %vtable, i16 0, i16 1
%fptr = load ptr, ptr %fptrptr
; CHECK: [[VTGEP2:%[^ ]*]] = getelementptr i8, ptr %vtable, i32 -1
; CHECK: [[VTLOAD2:%[^ ]*]] = load i8, ptr [[VTGEP2]]
@@ -98,27 +124,68 @@ define i1 @call2(ptr %obj) {
ret i1 %result
}
-; CHECK: define i32 @call3(
-define i32 @call3(ptr %obj) {
+; CHECK: define i16 @call3(
+define i16 @call3(ptr %obj) {
%vtable = load ptr, ptr %obj
%p = call i1 @llvm.type.test(ptr %vtable, metadata !"typeid")
call void @llvm.assume(i1 %p)
- %fptrptr = getelementptr [3 x ptr], ptr %vtable, i32 0, i32 2
+ %fptrptr = getelementptr [3 x ptr], ptr %vtable, i16 0, i16 2
%fptr = load ptr, ptr %fptrptr
- ; CHECK: [[VTGEP3:%[^ ]*]] = getelementptr i8, ptr %vtable, i32 -5
- ; CHECK: [[VTLOAD3:%[^ ]*]] = load i32, ptr [[VTGEP3]]
- %result = call i32 %fptr(ptr %obj)
- ; CHECK: ret i32 [[VTLOAD3]]
- ret i32 %result
+ ; CHECK: [[VTGEP3:%[^ ]*]] = getelementptr i8, ptr %vtable, i32 -4
+ ; CHECK: [[VTLOAD3:%[^ ]*]] = load i16, ptr [[VTGEP3]]
+ %result = call i16 %fptr(ptr %obj)
+ ; CHECK: ret i16 [[VTLOAD3]]
+ ret i16 %result
+}
+
+; CHECK: define i1 @call1_rel(
+define i1 @call1_rel(ptr %obj) {
+ %vtable = load ptr, ptr %obj
+ %p = call i1 @llvm.type.test(ptr %vtable, metadata !"typeid3")
+ call void @llvm.assume(i1 %p)
+ %fptr = call ptr @llvm.load.relative.i32(ptr %vtable, i32 0)
+ %result = call i1 %fptr(ptr %obj)
+ ret i1 %result
+ ; CHECK: [[RES:%[^ ]*]] = icmp eq ptr %vtable, @vt7_rel
+ ; CHECK: ret i1 [[RES]]
+}
+
+; CHECK: define i1 @call2_rel(
+define i1 @call2_rel(ptr %obj) {
+ %vtable = load ptr, ptr %obj
+ %p = call i1 @llvm.type.test(ptr %vtable, metadata !"typeid3")
+ call void @llvm.assume(i1 %p)
+ %fptr = call ptr @llvm.load.relative.i32(ptr %vtable, i32 4)
+ %result = call i1 %fptr(ptr %obj)
+ ret i1 %result
+ ; CHECK: [[RES:%[^ ]*]] = icmp eq ptr %vtable, @vt6_rel
+ ; CHECK: ret i1 [[RES]]
+}
+
+; CHECK: define i16 @call3_rel(
+define i16 @call3_rel(ptr %obj) {
+ %vtable = load ptr, ptr %obj
+ %p = call i1 @llvm.type.test(ptr %vtable, metadata !"typeid3")
+ call void @llvm.assume(i1 %p)
+ %fptr = call ptr @llvm.load.relative.i32(ptr %vtable, i32 8)
+ ; CHECK: [[VTGEP3:%[^ ]*]] = getelementptr i8, ptr %vtable, i32 -2
+ ; CHECK: [[VTLOAD3:%[^ ]*]] = load i16, ptr [[VTGEP3]]
+ %result = call i16 %fptr(ptr %obj)
+ ; CHECK: ret i16 [[VTLOAD3]]
+ ret i16 %result
}
declare i1 @llvm.type.test(ptr, metadata)
declare void @llvm.assume(i1)
declare void @__cxa_pure_virtual()
+declare ptr @llvm.load.relative.i32(ptr, i32)
; CHECK: [[T8]] = !{i32 8, !"typeid"}
-; CHECK: [[T5]] = !{i32 5, !"typeid"}
+; CHECK: [[T5]] = !{i32 4, !"typeid"}
; CHECK: [[T16]] = !{i32 16, !"typeid"}
; CHECK: [[T0]] = !{i32 0, !"typeid"}
+; CHECK: [[TREL]] = !{i32 4, !"typeid3"}
!0 = !{i32 0, !"typeid"}
+!1 = !{i32 0, !"typeid2"}
+!2 = !{i32 0, !"typeid3"}
diff --git a/llvm/test/Transforms/WholeProgramDevirt/virtual-const-prop-check.ll b/llvm/test/Transforms/WholeProgramDevirt/virtual-const-prop-check.ll
index fb8b6de003fdd..d8f5c912e9a50 100644
--- a/llvm/test/Transforms/WholeProgramDevirt/virtual-const-prop-check.ll
+++ b/llvm/test/Transforms/WholeProgramDevirt/virtual-const-prop-check.ll
@@ -7,13 +7,16 @@
; RUN: opt -S -passes=wholeprogramdevirt -whole-program-visibility -pass-remarks=wholeprogramdevirt -wholeprogramdevirt-skip=vf0i1 %s 2>&1 | FileCheck %s --check-prefix=SKIP
; We have two set of call targets {vf0i1, vf1i1} and {vf1i32, vf2i32, vf3i32, vf4i32}.
; The command below prevents both of them from devirtualization.
-; RUN: opt -S -passes=wholeprogramdevirt -whole-program-visibility -pass-remarks=wholeprogramdevirt -wholeprogramdevirt-skip=vf0i1,vf1i32 %s 2>&1 | FileCheck %s --check-prefix=SKIP-ALL
+; RUN: opt -S -passes=wholeprogramdevirt -whole-program-visibility -pass-remarks=wholeprogramdevirt -wholeprogramdevirt-skip=vf0i1,vf1i32,vf3i32 %s 2>&1 | FileCheck %s --check-prefix=SKIP-ALL
; Check wildcard
; RUN: opt -S -passes=wholeprogramdevirt -whole-program-visibility -pass-remarks=wholeprogramdevirt -wholeprogramdevirt-skip=vf?i1 %s 2>&1 | FileCheck %s --check-prefix=SKIP
target datalayout = "e-p:64:64"
target triple = "x86_64-unknown-linux-gnu"
+; CHECK: remark: <unknown>:0:0: unique-ret-val: devirtualized a call to vf0i1
+; CHECK: remark: <unknown>:0:0: unique-ret-val: devirtualized a call to vf1i1
+; CHECK: remark: <unknown>:0:0: virtual-const-prop: devirtualized a call to vf3i32
; CHECK: remark: <unknown>:0:0: virtual-const-prop-1-bit: devirtualized a call to vf0i1
; CHECK: remark: <unknown>:0:0: virtual-const-prop-1-bit: devirtualized a call to vf1i1
; CHECK: remark: <unknown>:0:0: virtual-const-prop: devirtualized a call to vf1i32
@@ -34,28 +37,28 @@ target triple = "x86_64-unknown-linux-gnu"
; SKIP-ALL-NOT: devirtualized
-; CHECK: [[VT1DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [0 x i8] } { [8 x i8] c"\00\00\00\01\00\00\00\02", [3 x ptr] [ptr @vf0i1, ptr @vf1i1, ptr @vf1i32], [0 x i8] zeroinitializer }, section "vt1sec", !type [[T8:![0-9]+]]
+; CHECK: [[VT1DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [4 x i8] } { [8 x i8] c"\00\00\00\00\00\00\00\02", [3 x ptr] [ptr @vf0i1, ptr @vf1i1, ptr @vf1i32], [4 x i8] c"\01\00\00\00" }, section "vt1sec", !type [[T8:![0-9]+]]
@vt1 = constant [3 x ptr] [
ptr @vf0i1,
ptr @vf1i1,
ptr @vf1i32
], section "vt1sec", !type !0
-; CHECK: [[VT2DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [0 x i8] } { [8 x i8] c"\00\00\00\02\00\00\00\01", [3 x ptr] [ptr @vf1i1, ptr @vf0i1, ptr @vf2i32], [0 x i8] zeroinitializer }, !type [[T8]]
+; CHECK: [[VT2DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [4 x i8] } { [8 x i8] c"\00\00\00\00\00\00\00\01", [3 x ptr] [ptr @vf1i1, ptr @vf0i1, ptr @vf2i32], [4 x i8] c"\02\00\00\00" }, !type [[T8]]
@vt2 = constant [3 x ptr] [
ptr @vf1i1,
ptr @vf0i1,
ptr @vf2i32
], !type !0
-; CHECK: [[VT3DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [0 x i8] } { [8 x i8] c"\00\00\00\03\00\00\00\02", [3 x ptr] [ptr @vf0i1, ptr @vf1i1, ptr @vf3i32], [0 x i8] zeroinitializer }, !type [[T8]]
+; CHECK: [[VT3DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [4 x i8] } { [8 x i8] c"\00\00\00\00\00\00\00\02", [3 x ptr] [ptr @vf0i1, ptr @vf1i1, ptr @vf3i32], [4 x i8] c"\03\00\00\00" }, !type [[T8]]
@vt3 = constant [3 x ptr] [
ptr @vf0i1,
ptr @vf1i1,
ptr @vf3i32
], !type !0
-; CHECK: [[VT4DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [0 x i8] } { [8 x i8] c"\00\00\00\04\00\00\00\01", [3 x ptr] [ptr @vf1i1, ptr @vf0i1, ptr @vf4i32], [0 x i8] zeroinitializer }, !type [[T8]]
+; CHECK: [[VT4DATA:@[^ ]*]] = private constant { [8 x i8], [3 x ptr], [4 x i8] } { [8 x i8] c"\00\00\00\00\00\00\00\01", [3 x ptr] [ptr @vf1i1, ptr @vf0i1, ptr @vf4i32], [4 x i8] c"\04\00\00\00" }, !type [[T8]]
@vt4 = constant [3 x ptr] [
ptr @vf1i1,
ptr @vf0i1,
@@ -69,10 +72,35 @@ ptr @__cxa_pure_virtual,
ptr @__cxa_pure_virtual
], !type !0
-; CHECK: @vt1 = alias [3 x ptr], getelementptr inbounds ({ [8 x i8], [3 x ptr], [0 x i8] }, ptr [[VT1DATA]], i32 0, i32 1)
-; CHECK: @vt2 = alias [3 x ptr], getelementptr inbounds ({ [8 x i8], [3 x ptr], [0 x i8] }, ptr [[VT2DATA]], i32 0, i32 1)
-; CHECK: @vt3 = alias [3 x ptr], getelementptr inbounds ({ [8 x i8], [3 x ptr], [0 x i8] }, ptr [[VT3DATA]], i32 0, i32 1)
-; CHECK: @vt4 = alias [3 x ptr], getelementptr inbounds ({ [8 x i8], [3 x ptr], [0 x i8] }, ptr [[VT4DATA]], i32 0, i32 1)
+;; Test relative vtables
+; CHECK: [[VT6RELDATA:@[^ ]*]] = private constant { [4 x i8], [3 x i32], [0 x i8] } { [4 x i8] c"\03\00\00\00", [3 x i32] [
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf0i1 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32),
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf1i1 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32),
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf3i32 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32)
+; CHECK-SAME: ], [0 x i8] zeroinitializer }, !type [[TREL:![0-9]+]]
+@vt6_rel = constant [3 x i32] [
+i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf0i1 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32),
+i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf1i1 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32),
+i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf3i32 to i64), i64 ptrtoint (ptr @vt6_rel to i64)) to i32)
+], !type !1
+
+; CHECK: [[VT7RELDATA:@[^ ]*]] = private constant { [4 x i8], [3 x i32], [0 x i8] } { [4 x i8] c"\04\00\00\00", [3 x i32] [
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf1i1 to i64), i64 ptrtoint (ptr @vt7_rel to i64)) to i32),
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf0i1 to i64), i64 ptrtoint (ptr @vt7_rel to i64)) to i32),
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf4i32 to i64), i64 ptrtoint (ptr @vt7_rel to i64)) to i32)
+; CHECK-SAME: ], [0 x i8] zeroinitializer }, !type [[TREL]]
+@vt7_rel = constant [3 x i32] [
+i32 trunc (i64 sub (i64 ptrtoint (ptr dso_local_equivalent @vf1i1 to i64), i64 ptrtoint (ptr @vt7_rel to...
[truncated]
|
|
An alternative approach I tried was using Note this approach can also increase the size of vtables slightly, but here are some ideas we can try to reduce the size:
|
pcc
reviewed
Apr 21, 2025
| Align TableAlignment = M.getDataLayout().getValueOrABITypeAlignment( | ||
| GV->getAlign(), GV->getValueType()); | ||
| if (TypeAlignment > TableAlignment) | ||
| return false; |
There was a problem hiding this comment.
Did you consider increasing the alignments of the vtables for the new member?
There was a problem hiding this comment.
I did not since I was under the assumption we wouldn't want to change the underlying table alignment. I think maybe we won't want to do this so we can keep the 4-byte alignment for relative vtables, but lemme see if others have thoughts on this.
There was a problem hiding this comment.
Yeah we want to keep the 4-byte alignment for relative vtables.
There was a problem hiding this comment.
Having a smaller alignment was one of the nice to have features for relative vtables that helped save some space.
There was a problem hiding this comment.
That's fine for now then since it's the simplest approach. It may be a good idea to examine the perf/size tradeoff of adding these constants in the future.
ilovepi
reviewed
Apr 22, 2025
ilovepi
reviewed
Apr 22, 2025
You're referring to having a PR with just |
yes, if possible. That way, when you update the test in this PR, it will be easy to understand what's changing in the new cases. |
PiJoules
added a commit
to PiJoules/llvm-project
that referenced
this pull request
May 7, 2025
…bles This is a patch with precommitted tests to make llvm#136630 easier to review. The `virtual-const-prop-small-alignment-*` tests check the output when the loaded int alignment is less than the vtable alignment. This also changes some constants to make it easier to differentiate between propagated values in vtables.
Uhh is there a way I can upload only the diff between this PR and #138989 in this PR rather than having 2 commits? Locally what I have is and I assumed pushing the branch for this PR would only push the that commit rather than both commits. |
|
Stacking PRs after the fact is hard(at least for me). Maybe land the pre commit change and then rebase this on main? |
pcc
approved these changes
May 8, 2025
| Align TableAlignment = M.getDataLayout().getValueOrABITypeAlignment( | ||
| GV->getAlign(), GV->getValueType()); | ||
| if (TypeAlignment > TableAlignment) | ||
| return false; |
There was a problem hiding this comment.
That's fine for now then since it's the simplest approach. It may be a good idea to examine the perf/size tradeoff of adding these constants in the future.
PiJoules
added a commit
to PiJoules/llvm-project
that referenced
this pull request
May 8, 2025
…bles This is a patch with precommitted tests to make llvm#136630 easier to review. The `virtual-const-prop-small-alignment-*` tests check the output when the loaded int alignment is less than the vtable alignment. This also changes some constants to make it easier to differentiate between propagated values in vtables.
PiJoules
added a commit
to PiJoules/llvm-project
that referenced
this pull request
May 8, 2025
…bles This is a patch with precommitted tests to make llvm#136630 easier to review. The `virtual-const-prop-small-alignment-*` tests check the output when the loaded int alignment is less than the vtable alignment. This also changes some constants to make it easier to differentiate between propagated values in vtables.
PiJoules
added a commit
to PiJoules/llvm-project
that referenced
this pull request
May 9, 2025
…bles This is a patch with precommitted tests to make llvm#136630 easier to review. The `virtual-const-prop-small-alignment-*` tests check the output when the loaded int alignment is less than the vtable alignment. This also changes some constants to make it easier to differentiate between propagated values in vtables.
PiJoules
added a commit
to PiJoules/llvm-project
that referenced
this pull request
May 9, 2025
…bles This is a patch with precommitted tests to make llvm#136630 easier to review. The `virtual-const-prop-small-alignment-*` tests check the output when the loaded int alignment is less than the vtable alignment. This also changes some constants to make it easier to differentiate between propagated values in vtables.
PiJoules
added a commit
that referenced
this pull request
May 13, 2025
…138989) This is a patch with precommitted tests to make #136630 easier to review. The `virtual-const-prop-small-alignment-*` tests check the output when the loaded int alignment is less than the vtable alignment. This also changes some constants to make it easier to differentiate between propagated values in vtables.
llvm-sync bot
pushed a commit
to arm/arm-toolchain
that referenced
this pull request
May 13, 2025
…e vtables (#138989) This is a patch with precommitted tests to make llvm/llvm-project#136630 easier to review. The `virtual-const-prop-small-alignment-*` tests check the output when the loaded int alignment is less than the vtable alignment. This also changes some constants to make it easier to differentiate between propagated values in vtables.
|
This causes lld to crash: Repro attached in https://crbug.com/421938962#comment2. |
|
Looking into this |
|
Forward fix at #142500 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
It's possible for virtual constant propagation in whole program devirtualization to create unaligned loads. We originally saw this with 4-byte aligned relative vtables where we could store 8-byte values before/after the vtable. But since the vtable is 4-byte aligned and we unconditionally do an 8-byte load, we can't guarantee that the stored constant will always be aligned to 8 bytes. We can also see this with normal vtables whenever a 1-byte char is stored in the vtable because the offset calculation for the GEP doesn't take into account the original vtable alignment.
This patch introduces two changes to virtual constant propagation:
This patch updates tests to reflect this alignment change and adds some cases for relative vtables.