[CIR] Upstream support for calling functions via member expressions

[andykaylor]

This adds support for calling functions via class member access expressions.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Andy Kaylor (andykaylor)

Changes

This adds support for calling functions via class member access expressions.

---

Full diff: https://github.com/llvm/llvm-project/pull/164518.diff

2 Files Affected:

• (modified) clang/lib/CIR/CodeGen/CIRGenExpr.cpp (+4-4)
• (added) clang/test/CIR/CodeGen/call-via-class-member-funcptr.cpp (+36)

diff --git a/clang/lib/CIR/CodeGen/CIRGenExpr.cpp b/clang/lib/CIR/CodeGen/CIRGenExpr.cpp index 52021fce1c675..3f504e5132684 100644 --- a/clang/lib/CIR/CodeGen/CIRGenExpr.cpp +++ b/clang/lib/CIR/CodeGen/CIRGenExpr.cpp @@ -1820,10 +1820,10 @@ CIRGenCallee CIRGenFunction::emitCallee(const clang::Expr *e) { // Resolve direct calls. const auto *funcDecl = cast<FunctionDecl>(declRef->getDecl()); return emitDirectCallee(funcDecl); - } else if (isa<MemberExpr>(e)) { - cgm.errorNYI(e->getSourceRange(), - "emitCallee: call to member function is NYI"); - return {}; + } else if (auto me = dyn_cast<MemberExpr>(e)) { + const auto *fd = cast<FunctionDecl>(me->getMemberDecl()); + emitIgnoredExpr(me->getBase()); + return emitDirectCallee(fd); } else if (auto *pde = dyn_cast<CXXPseudoDestructorExpr>(e)) { return CIRGenCallee::forPseudoDestructor(pde); } diff --git a/clang/test/CIR/CodeGen/call-via-class-member-funcptr.cpp b/clang/test/CIR/CodeGen/call-via-class-member-funcptr.cpp new file mode 100644 index 0000000000000..28f6a2020fd99 --- /dev/null +++ b/clang/test/CIR/CodeGen/call-via-class-member-funcptr.cpp @@ -0,0 +1,36 @@ +// RUN: %clang_cc1 -std=c++20 -triple aarch64-none-linux-android21 -fclangir -emit-cir %s -o %t.cir +// RUN: FileCheck --input-file=%t.cir %s --check-prefix=CIR +// RUN: %clang_cc1 -std=c++20 -triple aarch64-none-linux-android21 -fclangir -emit-llvm %s -o %t-cir.ll +// RUN: FileCheck --input-file=%t-cir.ll %s --check-prefix=LLVM +// RUN: %clang_cc1 -std=c++20 -triple aarch64-none-linux-android21 -emit-llvm %s -o %t.ll +// RUN: FileCheck --input-file=%t.ll %s --check-prefix=OGCG + +class A { +public: + static char *b(int); +}; + +int h=0; + +class F { +public: + const char *b(); + A g; +}; + +const char *F::b() { return g.b(h); } + +void fn1() { F f1; } + +// CIR: cir.func {{.*}} @_ZN1F1bEv +// CIR: %[[H_PTR:.*]] = cir.get_global @h : !cir.ptr<!s32i> +// CIR: %[[H_VAL:.*]] = cir.load{{.*}} %[[H_PTR]] : !cir.ptr<!s32i>, !s32i +// CIR: %[[RET:.*]] = cir.call @_ZN1A1bEi(%[[H_VAL]]) : (!s32i) -> !cir.ptr<!s8i> + +// LLVM: define {{.*}} ptr @_ZN1F1bEv +// LLVM: %[[VAR_H:.*]] = load i32, ptr @h +// LLVM: %[[RET:.*]] = call ptr @_ZN1A1bEi(i32 %[[VAR_H]]) + +// OGCG: define {{.*}} ptr @_ZN1F1bEv +// OGCG: %[[VAR_H:.*]] = load i32, ptr @h +// OGCG: %[[RET:.*]] = call noundef ptr @_ZN1A1bEi(i32 noundef %[[VAR_H]]) 

[llvmbot]

@llvm/pr-subscribers-clangir

Author: Andy Kaylor (andykaylor)

Changes

This adds support for calling functions via class member access expressions.

---

Full diff: https://github.com/llvm/llvm-project/pull/164518.diff

2 Files Affected:

• (modified) clang/lib/CIR/CodeGen/CIRGenExpr.cpp (+4-4)
• (added) clang/test/CIR/CodeGen/call-via-class-member-funcptr.cpp (+36)

diff --git a/clang/lib/CIR/CodeGen/CIRGenExpr.cpp b/clang/lib/CIR/CodeGen/CIRGenExpr.cpp index 52021fce1c675..3f504e5132684 100644 --- a/clang/lib/CIR/CodeGen/CIRGenExpr.cpp +++ b/clang/lib/CIR/CodeGen/CIRGenExpr.cpp @@ -1820,10 +1820,10 @@ CIRGenCallee CIRGenFunction::emitCallee(const clang::Expr *e) { // Resolve direct calls. const auto *funcDecl = cast<FunctionDecl>(declRef->getDecl()); return emitDirectCallee(funcDecl); - } else if (isa<MemberExpr>(e)) { - cgm.errorNYI(e->getSourceRange(), - "emitCallee: call to member function is NYI"); - return {}; + } else if (auto me = dyn_cast<MemberExpr>(e)) { + const auto *fd = cast<FunctionDecl>(me->getMemberDecl()); + emitIgnoredExpr(me->getBase()); + return emitDirectCallee(fd); } else if (auto *pde = dyn_cast<CXXPseudoDestructorExpr>(e)) { return CIRGenCallee::forPseudoDestructor(pde); } diff --git a/clang/test/CIR/CodeGen/call-via-class-member-funcptr.cpp b/clang/test/CIR/CodeGen/call-via-class-member-funcptr.cpp new file mode 100644 index 0000000000000..28f6a2020fd99 --- /dev/null +++ b/clang/test/CIR/CodeGen/call-via-class-member-funcptr.cpp @@ -0,0 +1,36 @@ +// RUN: %clang_cc1 -std=c++20 -triple aarch64-none-linux-android21 -fclangir -emit-cir %s -o %t.cir +// RUN: FileCheck --input-file=%t.cir %s --check-prefix=CIR +// RUN: %clang_cc1 -std=c++20 -triple aarch64-none-linux-android21 -fclangir -emit-llvm %s -o %t-cir.ll +// RUN: FileCheck --input-file=%t-cir.ll %s --check-prefix=LLVM +// RUN: %clang_cc1 -std=c++20 -triple aarch64-none-linux-android21 -emit-llvm %s -o %t.ll +// RUN: FileCheck --input-file=%t.ll %s --check-prefix=OGCG + +class A { +public: + static char *b(int); +}; + +int h=0; + +class F { +public: + const char *b(); + A g; +}; + +const char *F::b() { return g.b(h); } + +void fn1() { F f1; } + +// CIR: cir.func {{.*}} @_ZN1F1bEv +// CIR: %[[H_PTR:.*]] = cir.get_global @h : !cir.ptr<!s32i> +// CIR: %[[H_VAL:.*]] = cir.load{{.*}} %[[H_PTR]] : !cir.ptr<!s32i>, !s32i +// CIR: %[[RET:.*]] = cir.call @_ZN1A1bEi(%[[H_VAL]]) : (!s32i) -> !cir.ptr<!s8i> + +// LLVM: define {{.*}} ptr @_ZN1F1bEv +// LLVM: %[[VAR_H:.*]] = load i32, ptr @h +// LLVM: %[[RET:.*]] = call ptr @_ZN1A1bEi(i32 %[[VAR_H]]) + +// OGCG: define {{.*}} ptr @_ZN1F1bEv +// OGCG: %[[VAR_H:.*]] = load i32, ptr @h +// OGCG: %[[RET:.*]] = call noundef ptr @_ZN1A1bEi(i32 noundef %[[VAR_H]]) 

[erichkeane]

Classic-codegen only does this (I think? CGExpr.cpp~6289) IF the cast succeeds (and its a dyn_cast). Do we have reason to think that this is in error?

[andykaylor]

I noticed that, and the incubator does the same thing, but if the dyn_cast returns null, both just fall through to the indirect reference handling, and I couldn't find a case where that actually happens so I wasn't confident that it was the right thing to do. We made the same change (if (dyn_cast...) -> cast) just above this in the DeclRefExpr handling.

[erichkeane]

Hrmph.... that is concerning. And yeah, the dyn_cast behavior doesn't look like it does anything sensible too.

Can we have a comment here explaining this, so that next-guy through realizes we diverged (in case this becomes a problem)?

[andykaylor]

I was wrong. I found a way to make the cast fail.