[MemProf] Optionally allow transformation of nobuiltin operator new #158396
+70
−35
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
For cases where we can guarantee the application does not override operator new.
|
@llvm/pr-subscribers-llvm-transforms Author: Teresa Johnson (teresajohnson) ChangesFor cases where we can guarantee the application does not override Full diff: https://github.com/llvm/llvm-project/pull/158396.diff 3 Files Affected:
diff --git a/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h b/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
index deb3d6c44ef09..4e7c97194cc59 100644
--- a/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
+++ b/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
@@ -189,7 +189,7 @@ class LibCallSimplifier {
Value *optimizeMemSet(CallInst *CI, IRBuilderBase &B);
Value *optimizeRealloc(CallInst *CI, IRBuilderBase &B);
Value *optimizeNew(CallInst *CI, IRBuilderBase &B, LibFunc &Func);
- Value *optimizeExistingHotColdNew(CallInst *CI, IRBuilderBase &B);
+ Value *maybeOptimizeNoBuiltinOperatorNew(CallInst *CI, IRBuilderBase &B);
Value *optimizeWcslen(CallInst *CI, IRBuilderBase &B);
Value *optimizeBCopy(CallInst *CI, IRBuilderBase &B);
diff --git a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
index 8acebbaa5458b..4a1565977b91c 100644
--- a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
@@ -61,6 +61,9 @@ static cl::opt<bool> OptimizeExistingHotColdNew(
"optimize-existing-hot-cold-new", cl::Hidden, cl::init(false),
cl::desc(
"Enable optimization of existing hot/cold operator new library calls"));
+static cl::opt<bool> OptimizeNoBuiltinHotColdNew(
+ "optimize-nobuiltin-hot-cold-new-new", cl::Hidden, cl::init(false),
+ cl::desc("Enable transformation of nobuiltin operator new library calls"));
namespace {
@@ -1723,13 +1726,11 @@ Value *LibCallSimplifier::optimizeRealloc(CallInst *CI, IRBuilderBase &B) {
return nullptr;
}
-// Allow existing calls to operator new() that takes a __hot_cold_t parameter to
-// be updated with a compiler-determined hot cold hint value. This is used in
-// cases where the call is marked nobuiltin (because operator new called
-// explicitly) and therefore cannot be replaced with a different callee.
-Value *LibCallSimplifier::optimizeExistingHotColdNew(CallInst *CI,
- IRBuilderBase &B) {
- if (!OptimizeHotColdNew || !OptimizeExistingHotColdNew)
+// Optionally allow optimization of nobuiltin calls to operator new and its
+// variants.
+Value *LibCallSimplifier::maybeOptimizeNoBuiltinOperatorNew(CallInst *CI,
+ IRBuilderBase &B) {
+ if (!OptimizeHotColdNew)
return nullptr;
Function *Callee = CI->getCalledFunction();
if (!Callee)
@@ -1738,6 +1739,22 @@ Value *LibCallSimplifier::optimizeExistingHotColdNew(CallInst *CI,
if (!TLI->getLibFunc(*Callee, Func))
return nullptr;
switch (Func) {
+ case LibFunc_Znwm:
+ case LibFunc_ZnwmRKSt9nothrow_t:
+ case LibFunc_ZnwmSt11align_val_t:
+ case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t:
+ case LibFunc_Znam:
+ case LibFunc_ZnamRKSt9nothrow_t:
+ case LibFunc_ZnamSt11align_val_t:
+ case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t:
+ case LibFunc_size_returning_new:
+ case LibFunc_size_returning_new_aligned:
+ // By default normal operator new calls (not already passing a hot_cold_t
+ // parameter) are not mutated if the call is not marked builtin. Optionally
+ // enable that in cases where it is known to be safe.
+ if (!OptimizeNoBuiltinHotColdNew)
+ return nullptr;
+ break;
case LibFunc_Znwm12__hot_cold_t:
case LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t:
case LibFunc_ZnwmSt11align_val_t12__hot_cold_t:
@@ -1748,10 +1765,15 @@ Value *LibCallSimplifier::optimizeExistingHotColdNew(CallInst *CI,
case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t:
case LibFunc_size_returning_new_hot_cold:
case LibFunc_size_returning_new_aligned_hot_cold:
- return optimizeNew(CI, B, Func);
+ // If the nobuiltin call already passes a hot_cold_t parameter, allow update
+ // of that parameter when enabled.
+ if (!OptimizeExistingHotColdNew)
+ return nullptr;
+ break;
default:
return nullptr;
}
+ return optimizeNew(CI, B, Func);
}
// When enabled, replace operator new() calls marked with a hot or cold memprof
@@ -4121,9 +4143,8 @@ Value *LibCallSimplifier::optimizeCall(CallInst *CI, IRBuilderBase &Builder) {
// we can all non-FP calls with the StrictFP attribute to be
// optimized.
if (CI->isNoBuiltin()) {
- // If this is an existing call to a hot cold operator new, we can update the
- // hint parameter value, which doesn't change the callee.
- return optimizeExistingHotColdNew(CI, Builder);
+ // Optionally update operator new calls.
+ return maybeOptimizeNoBuiltinOperatorNew(CI, Builder);
}
LibFunc Func;
diff --git a/llvm/test/Transforms/InstCombine/simplify-libcalls-new.ll b/llvm/test/Transforms/InstCombine/simplify-libcalls-new.ll
index 41db7f929dfdf..5a4fb04f5f2c0 100644
--- a/llvm/test/Transforms/InstCombine/simplify-libcalls-new.ll
+++ b/llvm/test/Transforms/InstCombine/simplify-libcalls-new.ll
@@ -6,14 +6,18 @@
; OFF-LABEL: @new_hot_cold()
;; First check with the default hint values (254 = -2, 128 = -128, 222 = -34).
-; RUN: opt < %s -passes=instcombine -optimize-hot-cold-new -S | FileCheck %s --check-prefix=HOTCOLD -DCOLD=1 -DHOT=-2 -DNOTCOLD=-128 -DAMBIG=-34 -DPREVHINTCOLD=7 -DPREVHINTNOTCOLD=7 -DPREVHINTHOT=7 -DPREVHINTAMBIG=7
+; RUN: opt < %s -passes=instcombine -optimize-hot-cold-new -S | FileCheck %s --check-prefixes=HOTCOLD,NOBUILTIN-OFF -DCOLD=1 -DHOT=-2 -DNOTCOLD=-128 -DAMBIG=-34 -DPREVHINTCOLD=7 -DPREVHINTNOTCOLD=7 -DPREVHINTHOT=7 -DPREVHINTAMBIG=7
;; Next check with the non-default cold and hot hint values (200 =-56).
-; RUN: opt < %s -passes=instcombine -optimize-hot-cold-new -cold-new-hint-value=5 -hot-new-hint-value=200 -notcold-new-hint-value=99 -ambiguous-new-hint-value=44 -S | FileCheck %s --check-prefix=HOTCOLD -DCOLD=5 -DHOT=-56 -DAMBIG=44 -DNOTCOLD=99 -DPREVHINTCOLD=7 -DPREVHINTNOTCOLD=7 -DPREVHINTHOT=7 -DPREVHINTAMBIG=7
+; RUN: opt < %s -passes=instcombine -optimize-hot-cold-new -cold-new-hint-value=5 -hot-new-hint-value=200 -notcold-new-hint-value=99 -ambiguous-new-hint-value=44 -S | FileCheck %s --check-prefixes=HOTCOLD,NOBUILTIN-OFF -DCOLD=5 -DHOT=-56 -DAMBIG=44 -DNOTCOLD=99 -DPREVHINTCOLD=7 -DPREVHINTNOTCOLD=7 -DPREVHINTHOT=7 -DPREVHINTAMBIG=7
+
+;; Next check with the same non-default cold and hot hint values (200 =-56),
+;; but with transformation of nobuiltin calls enabled.
+; RUN: opt < %s -passes=instcombine -optimize-hot-cold-new -optimize-nobuiltin-hot-cold-new-new -cold-new-hint-value=5 -hot-new-hint-value=200 -notcold-new-hint-value=99 -ambiguous-new-hint-value=44 -S | FileCheck %s --check-prefixes=HOTCOLD,NOBUILTIN-ON -DCOLD=5 -DHOT=-56 -DAMBIG=44 -DNOTCOLD=99 -DPREVHINTCOLD=7 -DPREVHINTNOTCOLD=7 -DPREVHINTHOT=7 -DPREVHINTAMBIG=7
;; Try again with the non-default cold and hot hint values (200 =-56), and this
;; time specify that existing hints should be updated.
-; RUN: opt < %s -passes=instcombine -optimize-hot-cold-new -cold-new-hint-value=5 -notcold-new-hint-value=100 -hot-new-hint-value=200 -ambiguous-new-hint-value=44 -optimize-existing-hot-cold-new -S | FileCheck %s --check-prefix=HOTCOLD -DCOLD=5 -DHOT=-56 -DNOTCOLD=100 -DAMBIG=44 -DPREVHINTCOLD=5 -DPREVHINTNOTCOLD=100 -DPREVHINTHOT=-56 -DPREVHINTAMBIG=44
+; RUN: opt < %s -passes=instcombine -optimize-hot-cold-new -cold-new-hint-value=5 -notcold-new-hint-value=100 -hot-new-hint-value=200 -ambiguous-new-hint-value=44 -optimize-existing-hot-cold-new -S | FileCheck %s --check-prefixes=HOTCOLD,NOBUILTIN-OFF -DCOLD=5 -DHOT=-56 -DNOTCOLD=100 -DAMBIG=44 -DPREVHINTCOLD=5 -DPREVHINTNOTCOLD=100 -DPREVHINTHOT=-56 -DPREVHINTAMBIG=44
;; Make sure that values not in 0..255 are flagged with an error
; RUN: not opt < %s -passes=instcombine -optimize-hot-cold-new -cold-new-hint-value=256 -S 2>&1 | FileCheck %s --check-prefix=ERROR
@@ -40,8 +44,9 @@ define void @new() {
; HOTCOLD: @_Znwm12__hot_cold_t(i64 10, i8 [[AMBIG]])
%call4 = call ptr @_Znwm(i64 10) #7
call void @dummy(ptr %call4)
- ;; Attribute cold on a nobuiltin call has no effect.
- ; HOTCOLD: @_Znwm(i64 10)
+ ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled.
+ ; NOBUILTIN-OFF: @_Znwm(i64 10)
+ ; NOBUILTIN-ON: @_Znwm12__hot_cold_t(i64 10, i8 [[COLD]])
%call3 = call ptr @_Znwm(i64 10) #6
call void @dummy(ptr %call3)
ret void
@@ -68,8 +73,9 @@ define void @new_align() {
; HOTCOLD: @_ZnwmSt11align_val_t12__hot_cold_t(i64 10, i64 8, i8 [[AMBIG]])
%call4 = call ptr @_ZnwmSt11align_val_t(i64 10, i64 8) #7
call void @dummy(ptr %call4)
- ;; Attribute cold on a nobuiltin call has no effect.
- ; HOTCOLD: @_ZnwmSt11align_val_t(i64 10, i64 8)
+ ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled.
+ ; NOBUILTIN-OFF: @_ZnwmSt11align_val_t(i64 10, i64 8)
+ ; NOBUILTIN-ON: @_ZnwmSt11align_val_t12__hot_cold_t(i64 10, i64 8, i8 [[COLD]])
%call3 = call ptr @_ZnwmSt11align_val_t(i64 10, i64 8) #6
call void @dummy(ptr %call3)
ret void
@@ -97,8 +103,9 @@ define void @new_nothrow() {
; HOTCOLD: @_ZnwmRKSt9nothrow_t12__hot_cold_t(i64 10, ptr nonnull %nt, i8 [[AMBIG]])
%call4 = call ptr @_ZnwmRKSt9nothrow_t(i64 10, ptr %nt) #7
call void @dummy(ptr %call4)
- ;; Attribute cold on a nobuiltin call has no effect.
- ; HOTCOLD: @_ZnwmRKSt9nothrow_t(i64 10, ptr nonnull %nt)
+ ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled.
+ ; NOBUILTIN-OFF: @_ZnwmRKSt9nothrow_t(i64 10, ptr nonnull %nt)
+ ; NOBUILTIN-ON: @_ZnwmRKSt9nothrow_t12__hot_cold_t(i64 10, ptr nonnull %nt, i8 [[COLD]])
%call3 = call ptr @_ZnwmRKSt9nothrow_t(i64 10, ptr %nt) #6
call void @dummy(ptr %call3)
ret void
@@ -127,8 +134,9 @@ define void @new_align_nothrow() {
; HOTCOLD: @_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t(i64 10, i64 8, ptr nonnull %nt, i8 [[AMBIG]])
%call4 = call ptr @_ZnwmSt11align_val_tRKSt9nothrow_t(i64 10, i64 8, ptr %nt) #7
call void @dummy(ptr %call4)
- ;; Attribute cold on a nobuiltin call has no effect.
- ; HOTCOLD: @_ZnwmSt11align_val_tRKSt9nothrow_t(i64 10, i64 8, ptr nonnull %nt)
+ ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled.
+ ; NOBUILTIN-OFF: @_ZnwmSt11align_val_tRKSt9nothrow_t(i64 10, i64 8, ptr nonnull %nt)
+ ; NOBUILTIN-ON: @_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t(i64 10, i64 8, ptr nonnull %nt, i8 [[COLD]])
%call3 = call ptr @_ZnwmSt11align_val_tRKSt9nothrow_t(i64 10, i64 8, ptr %nt) #6
call void @dummy(ptr %call3)
ret void
@@ -154,8 +162,9 @@ define void @array_new() {
; HOTCOLD: @_Znam12__hot_cold_t(i64 10, i8 [[AMBIG]])
%call4 = call ptr @_Znam(i64 10) #7
call void @dummy(ptr %call4)
- ;; Attribute cold on a nobuiltin call has no effect.
- ; HOTCOLD: @_Znam(i64 10)
+ ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled.
+ ; NOBUILTIN-OFF: @_Znam(i64 10)
+ ; NOBUILTIN-ON: @_Znam12__hot_cold_t(i64 10, i8 [[COLD]])
%call3 = call ptr @_Znam(i64 10) #6
call void @dummy(ptr %call3)
ret void
@@ -182,8 +191,9 @@ define void @array_new_align() {
; HOTCOLD: @_ZnamSt11align_val_t12__hot_cold_t(i64 10, i64 8, i8 [[AMBIG]])
%call4 = call ptr @_ZnamSt11align_val_t(i64 10, i64 8) #7
call void @dummy(ptr %call4)
- ;; Attribute cold on a nobuiltin call has no effect.
- ; HOTCOLD: @_ZnamSt11align_val_t(i64 10, i64 8)
+ ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled.
+ ; NOBUILTIN-OFF: @_ZnamSt11align_val_t(i64 10, i64 8)
+ ; NOBUILTIN-ON: @_ZnamSt11align_val_t12__hot_cold_t(i64 10, i64 8, i8 [[COLD]])
%call3 = call ptr @_ZnamSt11align_val_t(i64 10, i64 8) #6
call void @dummy(ptr %call3)
ret void
@@ -211,8 +221,9 @@ define void @array_new_nothrow() {
; HOTCOLD: @_ZnamRKSt9nothrow_t12__hot_cold_t(i64 10, ptr nonnull %nt, i8 [[AMBIG]])
%call4 = call ptr @_ZnamRKSt9nothrow_t(i64 10, ptr %nt) #7
call void @dummy(ptr %call4)
- ;; Attribute cold on a nobuiltin call has no effect.
- ; HOTCOLD: @_ZnamRKSt9nothrow_t(i64 10, ptr nonnull %nt)
+ ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled.
+ ; NOBUILTIN-OFF: @_ZnamRKSt9nothrow_t(i64 10, ptr nonnull %nt)
+ ; NOBUILTIN-ON: @_ZnamRKSt9nothrow_t12__hot_cold_t(i64 10, ptr nonnull %nt, i8 [[COLD]])
%call3 = call ptr @_ZnamRKSt9nothrow_t(i64 10, ptr %nt) #6
call void @dummy(ptr %call3)
ret void
@@ -241,8 +252,9 @@ define void @array_new_align_nothrow() {
; HOTCOLD: @_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t(i64 10, i64 8, ptr nonnull %nt, i8 [[AMBIG]])
%call4 = call ptr @_ZnamSt11align_val_tRKSt9nothrow_t(i64 10, i64 8, ptr %nt) #7
call void @dummy(ptr %call4)
- ;; Attribute cold on a nobuiltin call has no effect.
- ; HOTCOLD: @_ZnamSt11align_val_tRKSt9nothrow_t(i64 10, i64 8, ptr nonnull %nt)
+ ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled.
+ ; NOBUILTIN-OFF: @_ZnamSt11align_val_tRKSt9nothrow_t(i64 10, i64 8, ptr nonnull %nt)
+ ; NOBUILTIN-ON: @_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t(i64 10, i64 8, ptr nonnull %nt, i8 [[COLD]])
%call3 = call ptr @_ZnamSt11align_val_tRKSt9nothrow_t(i64 10, i64 8, ptr %nt) #6
call void @dummy(ptr %call3)
ret void
@@ -492,8 +504,9 @@ define void @size_returning_test() {
%call4 = call {ptr, i64} @__size_returning_new(i64 10) #8
%p4 = extractvalue {ptr, i64} %call4, 0
call void @dummy(ptr %p4)
- ;; Attribute cold on a nobuiltin call has no effect.
- ; HOTCOLD: @__size_returning_new(i64 10)
+ ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled.
+ ; NOBUILTIN-OFF: @__size_returning_new(i64 10)
+ ; NOBUILTIN-ON: @__size_returning_new_hot_cold(i64 10, i8 [[COLD]])
%call3 = call {ptr, i64} @__size_returning_new(i64 10) #6
%p3 = extractvalue {ptr, i64} %call3, 0
call void @dummy(ptr %p3)
@@ -524,8 +537,9 @@ define void @size_returning_aligned_test() {
%call4 = call {ptr, i64} @__size_returning_new_aligned(i64 10, i64 8) #8
%p4 = extractvalue {ptr, i64} %call4, 0
call void @dummy(ptr %p4)
- ;; Attribute cold on a nobuiltin call has no effect.
- ; HOTCOLD: @__size_returning_new_aligned(i64 10, i64 8)
+ ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled.
+ ; NOBUILTIN-OFF: @__size_returning_new_aligned(i64 10, i64 8)
+ ; NOBUILTIN-ON: @__size_returning_new_aligned_hot_cold(i64 10, i64 8, i8 [[COLD]])
%call3 = call {ptr, i64} @__size_returning_new_aligned(i64 10, i64 8) #6
%p3 = extractvalue {ptr, i64} %call3, 0
call void @dummy(ptr %p3)
|
snehasish
reviewed
Sep 13, 2025
| ; HOTCOLD: @__size_returning_new_aligned(i64 10, i64 8) | ||
| ;; Attribute cold on a nobuiltin call has no effect, unless optionally enabled. | ||
| ; NOBUILTIN-OFF: @__size_returning_new_aligned(i64 10, i64 8) | ||
| ; NOBUILTIN-ON: @__size_returning_new_aligned_hot_cold(i64 10, i64 8, i8 [[COLD]]) | ||
| %call3 = call {ptr, i64} @__size_returning_new_aligned(i64 10, i64 8) #6 |
There was a problem hiding this comment.
Is __size_returning_new detected as a builtin by LLVM? I think I only added it to TLI in #101564.
There was a problem hiding this comment.
That's a good question. Here I have applied the attributes manually, but I need to check what is happening for them normally through the clang. That's orthogonal to this change though so I will commit this one and follow up on that.
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.
For cases where we can guarantee the application does not override
operator new.