Skip to content

Commit a96c14e

Browse files
kiran-isaacostannard
kiran-isaac
authored and
ostannard
committed
[Clang] Always forward sret parameters to musttail calls
If a call using the musttail attribute returns it's value through an sret argument pointer, we must forward an incoming sret pointer to it, instead of creating a new alloca. This is always possible because the musttail attribute requires the caller and callee to have the same argument and return types.
1 parent 78ec2e2 commit a96c14e

File tree

2 files changed

+89
-1
lines changed

2 files changed

+89
-1
lines changed

‎clang/lib/CodeGen/CGCall.cpp‎

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -5112,7 +5112,11 @@ RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
51125112
RawAddress SRetAlloca = RawAddress::invalid();
51135113
llvm::Value *UnusedReturnSizePtr = nullptr;
51145114
if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) {
5115-
if (IsVirtualFunctionPointerThunk && RetAI.isIndirect()) {
5115+
// For virtual function pointer thunks and musttail calls, we must always
5116+
// forward an incoming SRet pointer to the callee, because a local alloca
5117+
// would be de-allocated before the call. These cases both guarantee that
5118+
// there will be an incoming SRet argument of the correct type.
5119+
if ((IsVirtualFunctionPointerThunk || IsMustTail) && RetAI.isIndirect()) {
51165120
SRetPtr = makeNaturalAddressForPointer(CurFn->arg_begin() +
51175121
IRFunctionArgs.getSRetArgNo(),
51185122
RetTy, CharUnits::fromQuantity(1));

‎clang/test/CodeGen/musttail-sret.cpp‎

Lines changed: 84 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,84 @@
1+
// RUN: %clang_cc1 -triple=arm %s -emit-llvm -O3 -o - | FileCheck %s --check-prefix=CHECK-ARM
2+
// RUN: %clang_cc1 -triple=arm64 %s -emit-llvm -O3 -o - | FileCheck %s --check-prefix=CHECK-ARM64
3+
// RUN: %clang_cc1 -triple=i686 %s -emit-llvm -O3 -o - | FileCheck %s --check-prefix=CHECK-X86
4+
// RUN: %clang_cc1 -triple=x86_64 %s -emit-llvm -O3 -o - | FileCheck %s --check-prefix=CHECK-X64
5+
6+
// Sret tests
7+
struct Big {
8+
int a, b, c, d, e, f, g, h;
9+
};
10+
11+
struct Big F1(signed short P0);
12+
13+
struct Big F2(signed short P0) {
14+
signed short P1 = 20391;
15+
[[clang::musttail]] return F1(P1);
16+
}
17+
18+
// CHECK-NOT: alloca
19+
// CHECK-ARM: musttail call arm_aapcscc void @_Z2F1s(ptr dead_on_unwind writable sret(%struct.Big) align 4 %agg.result, i16 noundef signext 20391)
20+
// CHECK-ARM64: musttail call void @_Z2F1s(ptr dead_on_unwind writable sret(%struct.Big) align 4 %agg.result, i16 noundef 20391)
21+
// CHECK-X86: musttail call void @_Z2F1s(ptr dead_on_unwind writable sret(%struct.Big) align 4 %agg.result, i16 noundef signext 20391)
22+
// CHECK-X64: musttail call void @_Z2F1s(ptr dead_on_unwind writable sret(%struct.Big) align 4 %agg.result, i16 noundef signext 20391)
23+
24+
struct ReallyBig {
25+
int a[100];
26+
};
27+
28+
// Indirect sret tests
29+
// Function pointer for testing indirect musttail call.
30+
struct FunctionPointers {
31+
ReallyBig (*F3)(int, int, int, int, float, double);
32+
ReallyBig (*F4)(int, int, int, char, float, double);
33+
};
34+
35+
struct ReallyBig F3(int P0, int P1, int P2, int P3, float P4, double P5);
36+
struct ReallyBig F4(int P0, int P1, int P2, char P3, float P4, double P5);
37+
38+
static struct FunctionPointers FP = {F3, F4};
39+
40+
struct ReallyBig F5 (int P0, int P1, int P2, int P3, float P4, double P5) {
41+
[[clang::musttail]] return FP.F3(P0, P1, P2, P3, P4, P5);
42+
}
43+
44+
// CHECK-NOT: alloca
45+
// CHECK-ARM: musttail call arm_aapcscc void @_Z2F3iiiifd(ptr dead_on_unwind writable sret(%struct.ReallyBig) align 4 %agg.result, i32 noundef %P0, i32 noundef %P1, i32 noundef %P2, i32 noundef %P3, float noundef %P4, double noundef %P5)
46+
// CHECK-ARM64: musttail call void @_Z2F3iiiifd(ptr dead_on_unwind writable sret(%struct.ReallyBig) align 4 %agg.result, i32 noundef %P0, i32 noundef %P1, i32 noundef %P2, i32 noundef %P3, float noundef %P4, double noundef %P5)
47+
// CHECK-X86: musttail call void @_Z2F3iiiifd(ptr dead_on_unwind writable sret(%struct.ReallyBig) align 4 %agg.result, i32 noundef %P0, i32 noundef %P1, i32 noundef %P2, i32 noundef %P3, float noundef %P4, double noundef %P5)
48+
// CHECK-X64: musttail call void @_Z2F3iiiifd(ptr dead_on_unwind writable sret(%struct.ReallyBig) align 4 %agg.result, i32 noundef %P0, i32 noundef %P1, i32 noundef %P2, i32 noundef %P3, float noundef %P4, double noundef %P5)
49+
50+
struct ReallyBig F6 (int P0, int P1, int P2, char P3, float P4, double P5) {
51+
[[clang::musttail]] return FP.F4(P0, P1, P2, P3, P4, P5);
52+
}
53+
54+
// Complex and BitInt. Special cases for sret.
55+
// CHECK-NOT: alloca
56+
// CHECK-ARM: musttail call arm_aapcscc void @_Z2F4iiicfd(ptr dead_on_unwind writable sret(%struct.ReallyBig) align 4 %agg.result, i32 noundef %P0, i32 noundef %P1, i32 noundef %P2, i8 noundef signext %P3, float noundef %P4, double noundef %P5)
57+
// CHECK-ARM64: musttail call void @_Z2F4iiicfd(ptr dead_on_unwind writable sret(%struct.ReallyBig) align 4 %agg.result, i32 noundef %P0, i32 noundef %P1, i32 noundef %P2, i8 noundef %P3, float noundef %P4, double noundef %P5)
58+
// CHECK-X86: musttail call void @_Z2F4iiicfd(ptr dead_on_unwind writable sret(%struct.ReallyBig) align 4 %agg.result, i32 noundef %P0, i32 noundef %P1, i32 noundef %P2, i8 noundef signext %P3, float noundef %P4, double noundef %P5)
59+
// CHECK-X64: musttail call void @_Z2F4iiicfd(ptr dead_on_unwind writable sret(%struct.ReallyBig) align 4 %agg.result, i32 noundef %P0, i32 noundef %P1, i32 noundef %P2, i8 noundef signext %P3, float noundef %P4, double noundef %P5)
60+
61+
double _Complex F7(signed short P0);
62+
63+
double _Complex F8(signed short P0) {
64+
signed short P1 = 20391;
65+
[[clang::musttail]] return F7(P1);
66+
}
67+
68+
// CHECK-NOT: alloca
69+
// CHECK-ARM: musttail call arm_aapcscc void @_Z2F7s(ptr dead_on_unwind writable sret({ double, double }) align 8 %agg.result, i16 noundef signext 20391)
70+
// CHECK-ARM64: musttail call noundef { double, double } @_Z2F7s(i16 noundef 20391)
71+
// CHECK-X86: musttail call void @_Z2F7s(ptr dead_on_unwind writable sret({ double, double }) align 4 %agg.result, i16 noundef signext 20391)
72+
// CHECK-X64: musttail call noundef { double, double } @_Z2F7s(i16 noundef signext 20391)
73+
74+
signed _BitInt(100) F9(float P0, float P1, double P2, char P3);
75+
76+
signed _BitInt(100) F10(float P0, float P1, double P2, char P3) {
77+
[[clang::musttail]] return F9(P0, P1, P2, P3);
78+
}
79+
80+
// CHECK-NOT: alloca
81+
// CHECK-ARM: musttail call arm_aapcscc void @_Z2F9ffdc(ptr dead_on_unwind writable sret(i128) align 8 %agg.result, float noundef %P0, float noundef %P1, double noundef %P2, i8 noundef signext %P3)
82+
// CHECK-ARM64: musttail call noundef i100 @_Z2F9ffdc(float noundef %P0, float noundef %P1, double noundef %P2, i8 noundef %P3)
83+
// CHECK-X86: musttail call void @_Z2F9ffdc(ptr dead_on_unwind writable sret(i128) align 4 %agg.result, float noundef %P0, float noundef %P1, double noundef %P2, i8 noundef signext %P3)
84+
// CHECK-X64: musttail call noundef { i64, i64 } @_Z2F9ffdc(float noundef %P0, float noundef %P1, double noundef %P2, i8 noundef signext %P3)

0 commit comments

Comments
 (0)