[llvm] [Inliner] Propagate callee argument memory access attributes before inlining (PR #89024)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Apr 22 18:49:03 PDT 2024
https://github.com/goldsteinn updated https://github.com/llvm/llvm-project/pull/89024
>From c3e8c683172d39a272ca8a7f236249acf923ea1d Mon Sep 17 00:00:00 2001
From: Noah Goldstein <goldstein.w.n at gmail.com>
Date: Wed, 17 Apr 2024 01:30:48 -0500
Subject: [PATCH] [Inliner] Propagate callee argument memory access attributes
before inlining
To avoid losing information, we can propagate some access attribute
from the to-be-inlined callee to its callsites.
We can propagate argument memory access attributes to callsite
parameters if they are from the same underlying object.
---
llvm/lib/Transforms/Utils/InlineFunction.cpp | 74 +++++++++++++++++++
.../Inline/access-attributes-prop.ll | 18 ++---
.../Transforms/Inline/noalias-calls-always.ll | 20 ++---
llvm/test/Transforms/Inline/noalias-calls.ll | 20 ++---
4 files changed, 103 insertions(+), 29 deletions(-)
diff --git a/llvm/lib/Transforms/Utils/InlineFunction.cpp b/llvm/lib/Transforms/Utils/InlineFunction.cpp
index 833dcbec228b88..cad1f53d19835d 100644
--- a/llvm/lib/Transforms/Utils/InlineFunction.cpp
+++ b/llvm/lib/Transforms/Utils/InlineFunction.cpp
@@ -1344,6 +1344,76 @@ static bool MayContainThrowingOrExitingCallAfterCB(CallBase *Begin,
++BeginIt, End->getIterator(), InlinerAttributeWindow + 1);
}
+// Add attributes from CB params and Fn attributes that can always be propagated
+// to the corresponding argument / inner callbases.
+static void AddParamAndFnBasicAttributes(const CallBase &CB,
+ ValueToValueMapTy &VMap) {
+ auto *CalledFunction = CB.getCalledFunction();
+ auto &Context = CalledFunction->getContext();
+
+ // Collect valid attributes for all params.
+ SmallVector<AttrBuilder> ValidParamAttrs;
+ bool HasAttrToPropagate = false;
+
+ for (unsigned I = 0, E = CB.arg_size(); I < E; ++I) {
+ ValidParamAttrs.emplace_back(AttrBuilder{CB.getContext()});
+ // Access attributes can be propagated to any param with the same underlying
+ // object as the argument.
+ if (CB.paramHasAttr(I, Attribute::ReadNone))
+ ValidParamAttrs.back().addAttribute(Attribute::ReadNone);
+ if (CB.paramHasAttr(I, Attribute::ReadOnly))
+ ValidParamAttrs.back().addAttribute(Attribute::ReadOnly);
+ if (CB.paramHasAttr(I, Attribute::WriteOnly))
+ ValidParamAttrs.back().addAttribute(Attribute::WriteOnly);
+ HasAttrToPropagate |= ValidParamAttrs.back().hasAttributes();
+ }
+
+ // Won't be able to propagate anything.
+ if (!HasAttrToPropagate)
+ return;
+
+ for (BasicBlock &BB : *CalledFunction) {
+ for (Instruction &Ins : BB) {
+ if (const auto *InnerCB = dyn_cast<CallBase>(&Ins)) {
+ if (auto *NewInnerCB =
+ dyn_cast_or_null<CallBase>(VMap.lookup(InnerCB))) {
+ AttributeList AL = NewInnerCB->getAttributes();
+ for (unsigned I = 0, E = InnerCB->arg_size(); I < E; ++I) {
+ // Check if the underlying value for the parameter is an argument.
+ const Value *UnderlyingV =
+ getUnderlyingObject(InnerCB->getArgOperand(I));
+ if (const Argument *Arg = dyn_cast<Argument>(UnderlyingV)) {
+ unsigned ArgNo = Arg->getArgNo();
+ // If so, propagate its access attributes.
+ AL = AL.addParamAttributes(Context, I, ValidParamAttrs[ArgNo]);
+ // We can have conflicting attributes from the inner callsite and
+ // to-be-inlined callsite. In that case, choose the most
+ // restrictive.
+
+ // readonly + writeonly means we can never deref so make readnone.
+ if (AL.hasParamAttr(I, Attribute::ReadOnly) &&
+ AL.hasParamAttr(I, Attribute::WriteOnly))
+ AL = AL.addParamAttribute(Context, I, Attribute::ReadNone);
+
+ // If have readnone, need to clear readonly/writeonly
+ if (AL.hasParamAttr(I, Attribute::ReadNone)) {
+ AL = AL.removeParamAttribute(Context, I, Attribute::ReadOnly);
+ AL = AL.removeParamAttribute(Context, I, Attribute::WriteOnly);
+ }
+
+ // Writable cannot exist in conjunction w/ readonly/readnone
+ if (AL.hasParamAttr(I, Attribute::ReadOnly) ||
+ AL.hasParamAttr(I, Attribute::ReadNone))
+ AL = AL.removeParamAttribute(Context, I, Attribute::Writable);
+ }
+ }
+ NewInnerCB->setAttributes(AL);
+ }
+ }
+ }
+ }
+}
+
// Only allow these white listed attributes to be propagated back to the
// callee. This is because other attributes may only be valid on the call
// itself, i.e. attributes such as signext and zeroext.
@@ -2363,6 +2433,10 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
// function which feed into its return value.
AddReturnAttributes(CB, VMap);
+ // Clone attributes on the params of the callsite to calls within the
+ // inlined function which use the same param.
+ AddParamAndFnBasicAttributes(CB, VMap);
+
propagateMemProfMetadata(CalledFunc, CB,
InlinedFunctionInfo.ContainsMemProfMetadata, VMap);
diff --git a/llvm/test/Transforms/Inline/access-attributes-prop.ll b/llvm/test/Transforms/Inline/access-attributes-prop.ll
index 3b4a59897c5694..f91f3550f338f5 100644
--- a/llvm/test/Transforms/Inline/access-attributes-prop.ll
+++ b/llvm/test/Transforms/Inline/access-attributes-prop.ll
@@ -168,7 +168,7 @@ define dso_local void @foo2_through_obj(ptr %p, ptr %p2) {
define void @prop_param_func_decl(ptr %p) {
; CHECK-LABEL: define {{[^@]+}}@prop_param_func_decl
; CHECK-SAME: (ptr [[P:%.*]]) {
-; CHECK-NEXT: call void @bar1(ptr [[P]])
+; CHECK-NEXT: call void @bar1(ptr readonly [[P]])
; CHECK-NEXT: ret void
;
call void @foo1_rdonly(ptr %p)
@@ -178,7 +178,7 @@ define void @prop_param_func_decl(ptr %p) {
define void @prop_param_callbase_def(ptr %p) {
; CHECK-LABEL: define {{[^@]+}}@prop_param_callbase_def
; CHECK-SAME: (ptr [[P:%.*]]) {
-; CHECK-NEXT: call void @bar1(ptr [[P]])
+; CHECK-NEXT: call void @bar1(ptr readonly [[P]])
; CHECK-NEXT: call void @bar1(ptr [[P]])
; CHECK-NEXT: ret void
;
@@ -190,7 +190,7 @@ define void @prop_param_callbase_def(ptr %p) {
define void @prop_param_callbase_def_2x(ptr %p, ptr %p2) {
; CHECK-LABEL: define {{[^@]+}}@prop_param_callbase_def_2x
; CHECK-SAME: (ptr [[P:%.*]], ptr [[P2:%.*]]) {
-; CHECK-NEXT: call void @bar2(ptr [[P]], ptr [[P]])
+; CHECK-NEXT: call void @bar2(ptr readonly [[P]], ptr readonly [[P]])
; CHECK-NEXT: ret void
;
call void @foo2(ptr readonly %p, ptr %p)
@@ -202,7 +202,7 @@ define void @prop_param_callbase_def_2x_2(ptr %p, ptr %p2) {
; CHECK-SAME: (ptr [[P:%.*]], ptr [[P2:%.*]]) {
; CHECK-NEXT: [[PP_I:%.*]] = getelementptr i8, ptr [[P]], i64 9
; CHECK-NEXT: [[P2P_I:%.*]] = getelementptr i8, ptr [[P2]], i64 123
-; CHECK-NEXT: call void @bar2(ptr [[P2P_I]], ptr [[PP_I]])
+; CHECK-NEXT: call void @bar2(ptr writeonly [[P2P_I]], ptr readonly [[PP_I]])
; CHECK-NEXT: ret void
;
call void @foo2_through_obj(ptr readonly %p, ptr writeonly %p2)
@@ -214,7 +214,7 @@ define void @prop_param_callbase_def_2x_incompat(ptr %p, ptr %p2) {
; CHECK-SAME: (ptr [[P:%.*]], ptr [[P2:%.*]]) {
; CHECK-NEXT: [[PP_I:%.*]] = getelementptr i8, ptr [[P]], i64 9
; CHECK-NEXT: [[P2P_I:%.*]] = getelementptr i8, ptr [[P]], i64 123
-; CHECK-NEXT: call void @bar2(ptr [[P2P_I]], ptr [[PP_I]])
+; CHECK-NEXT: call void @bar2(ptr readonly [[P2P_I]], ptr readnone [[PP_I]])
; CHECK-NEXT: ret void
;
call void @foo2_through_obj(ptr readnone %p, ptr readonly %p)
@@ -224,7 +224,7 @@ define void @prop_param_callbase_def_2x_incompat(ptr %p, ptr %p2) {
define void @prop_param_callbase_def_2x_incompat_2(ptr %p, ptr %p2) {
; CHECK-LABEL: define {{[^@]+}}@prop_param_callbase_def_2x_incompat_2
; CHECK-SAME: (ptr [[P:%.*]], ptr [[P2:%.*]]) {
-; CHECK-NEXT: call void @bar2(ptr [[P]], ptr [[P]])
+; CHECK-NEXT: call void @bar2(ptr readonly [[P]], ptr readonly [[P]])
; CHECK-NEXT: ret void
;
call void @foo2(ptr readonly %p, ptr readnone %p)
@@ -234,7 +234,7 @@ define void @prop_param_callbase_def_2x_incompat_2(ptr %p, ptr %p2) {
define void @prop_param_callbase_def_2x_incompat_3(ptr %p, ptr %p2) {
; CHECK-LABEL: define {{[^@]+}}@prop_param_callbase_def_2x_incompat_3
; CHECK-SAME: (ptr [[P:%.*]], ptr [[P2:%.*]]) {
-; CHECK-NEXT: call void @bar2(ptr [[P]], ptr [[P]])
+; CHECK-NEXT: call void @bar2(ptr readnone [[P]], ptr readnone [[P]])
; CHECK-NEXT: ret void
;
call void @foo2_2(ptr readonly %p, ptr readnone %p)
@@ -244,7 +244,7 @@ define void @prop_param_callbase_def_2x_incompat_3(ptr %p, ptr %p2) {
define void @prop_param_callbase_def_1x_partial(ptr %p, ptr %p2) {
; CHECK-LABEL: define {{[^@]+}}@prop_param_callbase_def_1x_partial
; CHECK-SAME: (ptr [[P:%.*]], ptr [[P2:%.*]]) {
-; CHECK-NEXT: call void @bar2(ptr [[P]], ptr [[P]])
+; CHECK-NEXT: call void @bar2(ptr readonly [[P]], ptr readonly [[P]])
; CHECK-NEXT: ret void
;
call void @foo2(ptr readonly %p, ptr %p)
@@ -264,7 +264,7 @@ define void @prop_param_callbase_def_1x_partial_2(ptr %p, ptr %p2) {
define void @prop_param_callbase_def_1x_partial_3(ptr %p, ptr %p2) {
; CHECK-LABEL: define {{[^@]+}}@prop_param_callbase_def_1x_partial_3
; CHECK-SAME: (ptr [[P:%.*]], ptr [[P2:%.*]]) {
-; CHECK-NEXT: call void @bar2(ptr [[P]], ptr [[P]])
+; CHECK-NEXT: call void @bar2(ptr readonly [[P]], ptr readnone [[P]])
; CHECK-NEXT: ret void
;
call void @foo2_3(ptr readonly %p, ptr %p)
diff --git a/llvm/test/Transforms/Inline/noalias-calls-always.ll b/llvm/test/Transforms/Inline/noalias-calls-always.ll
index 9c851b93278392..a80cd12b26b655 100644
--- a/llvm/test/Transforms/Inline/noalias-calls-always.ll
+++ b/llvm/test/Transforms/Inline/noalias-calls-always.ll
@@ -34,11 +34,11 @@ define void @foo(ptr nocapture %a, ptr nocapture readonly %c, ptr nocapture %b)
; CHECK-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META0:![0-9]+]])
; CHECK-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META3:![0-9]+]])
; CHECK-NEXT: call void @llvm.lifetime.start.p0(i64 512, ptr [[L_I]])
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A:%.*]], ptr align 16 [[B:%.*]], i64 16, i1 false), !noalias !3
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[B]], ptr align 16 [[C:%.*]], i64 16, i1 false), !noalias !0
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr align 16 [[C]], i64 16, i1 false), !alias.scope !5
-; CHECK-NEXT: call void @hey(), !noalias !5
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[L_I]], ptr align 16 [[C]], i64 16, i1 false), !noalias !0
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A:%.*]], ptr align 16 [[B:%.*]], i64 16, i1 false), !noalias [[META3]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[B]], ptr readonly align 16 [[C:%.*]], i64 16, i1 false), !noalias [[META0]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr readonly align 16 [[C]], i64 16, i1 false), !alias.scope [[META5:![0-9]+]]
+; CHECK-NEXT: call void @hey(), !noalias [[META5]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[L_I]], ptr readonly align 16 [[C]], i64 16, i1 false), !noalias [[META0]]
; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 512, ptr [[L_I]])
; CHECK-NEXT: ret void
;
@@ -75,11 +75,11 @@ define void @foo_cs(ptr nocapture %a, ptr nocapture readonly %c, ptr nocapture %
; CHECK-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META6:![0-9]+]])
; CHECK-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META9:![0-9]+]])
; CHECK-NEXT: call void @llvm.lifetime.start.p0(i64 512, ptr [[L_I]])
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A:%.*]], ptr align 16 [[B:%.*]], i64 16, i1 false), !noalias !9
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[B]], ptr align 16 [[C:%.*]], i64 16, i1 false), !noalias !6
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr align 16 [[C]], i64 16, i1 false), !alias.scope !11
-; CHECK-NEXT: call void @hey(), !noalias !11
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[L_I]], ptr align 16 [[C]], i64 16, i1 false), !noalias !6
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A:%.*]], ptr align 16 [[B:%.*]], i64 16, i1 false), !noalias [[META9]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[B]], ptr readonly align 16 [[C:%.*]], i64 16, i1 false), !noalias [[META6]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr readonly align 16 [[C]], i64 16, i1 false), !alias.scope [[META11:![0-9]+]]
+; CHECK-NEXT: call void @hey(), !noalias [[META11]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[L_I]], ptr readonly align 16 [[C]], i64 16, i1 false), !noalias [[META6]]
; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 512, ptr [[L_I]])
; CHECK-NEXT: ret void
;
diff --git a/llvm/test/Transforms/Inline/noalias-calls.ll b/llvm/test/Transforms/Inline/noalias-calls.ll
index e3791da54b2326..0dd9ec3498a93a 100644
--- a/llvm/test/Transforms/Inline/noalias-calls.ll
+++ b/llvm/test/Transforms/Inline/noalias-calls.ll
@@ -37,11 +37,11 @@ define void @foo(ptr nocapture %a, ptr nocapture readonly %c, ptr nocapture %b)
; CHECK-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META0:![0-9]+]])
; CHECK-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META3:![0-9]+]])
; CHECK-NEXT: call void @llvm.lifetime.start.p0(i64 512, ptr [[L_I]])
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr align 16 [[B]], i64 16, i1 false), !noalias !3
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[B]], ptr align 16 [[C]], i64 16, i1 false), !noalias !0
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr align 16 [[C]], i64 16, i1 false), !alias.scope !5
-; CHECK-NEXT: call void @hey(), !noalias !5
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[L_I]], ptr align 16 [[C]], i64 16, i1 false), !noalias !0
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr align 16 [[B]], i64 16, i1 false), !noalias [[META3]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[B]], ptr readonly align 16 [[C]], i64 16, i1 false), !noalias [[META0]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr readonly align 16 [[C]], i64 16, i1 false), !alias.scope [[META5:![0-9]+]]
+; CHECK-NEXT: call void @hey(), !noalias [[META5]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[L_I]], ptr readonly align 16 [[C]], i64 16, i1 false), !noalias [[META0]]
; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 512, ptr [[L_I]])
; CHECK-NEXT: ret void
;
@@ -80,11 +80,11 @@ define void @foo_cs(ptr nocapture %a, ptr nocapture readonly %c, ptr nocapture %
; CHECK-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META6:![0-9]+]])
; CHECK-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META9:![0-9]+]])
; CHECK-NEXT: call void @llvm.lifetime.start.p0(i64 512, ptr [[L_I]])
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr align 16 [[B]], i64 16, i1 false), !noalias !9
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[B]], ptr align 16 [[C]], i64 16, i1 false), !noalias !6
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr align 16 [[C]], i64 16, i1 false), !alias.scope !11
-; CHECK-NEXT: call void @hey(), !noalias !11
-; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[L_I]], ptr align 16 [[C]], i64 16, i1 false), !noalias !6
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr align 16 [[B]], i64 16, i1 false), !noalias [[META9]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[B]], ptr readonly align 16 [[C]], i64 16, i1 false), !noalias [[META6]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[A]], ptr readonly align 16 [[C]], i64 16, i1 false), !alias.scope [[META11:![0-9]+]]
+; CHECK-NEXT: call void @hey(), !noalias [[META11]]
+; CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[L_I]], ptr readonly align 16 [[C]], i64 16, i1 false), !noalias [[META6]]
; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 512, ptr [[L_I]])
; CHECK-NEXT: ret void
;
More information about the llvm-commits
mailing list