[clang] [Sparc] don't pass aggregate varargs in float registers (PR #216524)
via cfe-commits
cfe-commits at lists.llvm.org
Sun Aug 16 06:47:40 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-clang-codegen
Author: Folkert de Vries (folkertdev)
<details>
<summary>Changes</summary>
https://godbolt.org/z/o1nGYEM33
```c
extern void sink(int, ...);
void variadic_cld(_Complex float z) { sink(0, z); }
```
Current clang just uses float registers
```asm
variadic_cld:
mov %g0, %o0
fmovs %f0, %f2
fmovs %f1, %f3
mov %o7, %g1
call sink
mov %g1, %o7
```
GCC packs a complex float into a single GPR
```asm
variadic_cld:
save %sp, -192, %sp
st %f0, [%fp+2043]
lduw [%fp+2043], %g1
sllx %g1, 32, %g1
srl %o1, 0, %o1
or %o1, %g1, %o1
st %f1, [%fp+2043]
lduw [%fp+2043], %g2
mov -1, %g1
sllx %g1, 32, %g1
and %o1, %g1, %o1
or %o1, %g2, %o1
call sink, 0
mov 0, %o0
return %i7+8
nop
```
---
Full diff: https://github.com/llvm/llvm-project/pull/216524.diff
2 Files Affected:
- (modified) clang/lib/CodeGen/Targets/Sparc.cpp (+17-8)
- (modified) clang/test/CodeGen/Sparc/variadic-aggregate.c (+24-36)
``````````diff
diff --git a/clang/lib/CodeGen/Targets/Sparc.cpp b/clang/lib/CodeGen/Targets/Sparc.cpp
index 25f0009fe7a09..9d8280ad25dd6 100644
--- a/clang/lib/CodeGen/Targets/Sparc.cpp
+++ b/clang/lib/CodeGen/Targets/Sparc.cpp
@@ -163,7 +163,7 @@ class SparcV9ABIInfo : public ABIInfo {
bool IsComplexGnuABI;
ABIArgInfo classifyType(QualType RetTy, unsigned SizeLimit,
- unsigned &RegOffset) const;
+ unsigned &RegOffset, bool IsVarArg) const;
void computeInfo(CGFunctionInfo &FI) const override;
RValue EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty,
AggValueSlot Slot) const override;
@@ -185,9 +185,11 @@ class SparcV9ABIInfo : public ABIInfo {
SmallVector<llvm::Type*, 8> Elems;
uint64_t Size;
bool InReg;
+ bool IsVarArg;
- CoerceBuilder(llvm::LLVMContext &c, const llvm::DataLayout &dl)
- : Context(c), DL(dl), Size(0), InReg(false) {}
+ CoerceBuilder(llvm::LLVMContext &c, const llvm::DataLayout &dl,
+ bool IsVarArg)
+ : Context(c), DL(dl), Size(0), InReg(false), IsVarArg(IsVarArg) {}
// Pad Elems with integers until Size is ToSize.
void pad(uint64_t ToSize) {
@@ -217,6 +219,9 @@ class SparcV9ABIInfo : public ABIInfo {
// Add a floating point element at Offset.
void addFloat(uint64_t Offset, llvm::Type *Ty, unsigned Bits) {
+ // Varargs are treated as integers.
+ if (IsVarArg)
+ return;
// Unaligned floats are treated as integers.
if (Offset % Bits)
return;
@@ -277,7 +282,8 @@ class SparcV9ABIInfo : public ABIInfo {
} // end anonymous namespace
ABIArgInfo SparcV9ABIInfo::classifyType(QualType Ty, unsigned SizeLimit,
- unsigned &RegOffset) const {
+ unsigned &RegOffset,
+ bool IsVarArg) const {
if (Ty->isVoidType())
return ABIArgInfo::getIgnore();
@@ -369,7 +375,7 @@ ABIArgInfo SparcV9ABIInfo::classifyType(QualType Ty, unsigned SizeLimit,
return ABIArgInfo::getDirect(/*T=*/nullptr, /*Offset=*/0, Padding);
}
- CoerceBuilder CB(VMContext, getDataLayout());
+ CoerceBuilder CB(VMContext, getDataLayout(), IsVarArg);
CB.addStruct(0, StrTy);
// All structs, even empty ones, should take up a register argument slot,
// so pin the minimum struct size to one bit.
@@ -406,13 +412,16 @@ RValue SparcV9ABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
void SparcV9ABIInfo::computeInfo(CGFunctionInfo &FI) const {
unsigned RetOffset = 0;
- ABIArgInfo RetType = classifyType(FI.getReturnType(), 32 * 8, RetOffset);
+ ABIArgInfo RetType =
+ classifyType(FI.getReturnType(), 32 * 8, RetOffset, /*IsVarArg=*/false);
FI.getReturnInfo() = RetType;
// Indirect returns will have its pointer passed as an argument.
unsigned ArgOffset = RetType.isIndirect() ? RetOffset : 0;
- for (auto &I : FI.arguments())
- I.info = classifyType(I.type, 16 * 8, ArgOffset);
+ for (auto [ArgNo, I] : llvm::enumerate(FI.arguments())) {
+ bool IsVarArg = ArgNo >= FI.getNumRequiredArgs();
+ I.info = classifyType(I.type, 16 * 8, ArgOffset, IsVarArg);
+ }
}
namespace {
diff --git a/clang/test/CodeGen/Sparc/variadic-aggregate.c b/clang/test/CodeGen/Sparc/variadic-aggregate.c
index 92093e1133b8a..fb7524b887980 100644
--- a/clang/test/CodeGen/Sparc/variadic-aggregate.c
+++ b/clang/test/CodeGen/Sparc/variadic-aggregate.c
@@ -615,11 +615,8 @@ void test_long_long_long_long(va_list *ap) {
// SPARC64-NEXT: [[COERCE_IMAGP:%.*]] = getelementptr inbounds nuw { float, float }, ptr [[COERCE]], i32 0, i32 1
// SPARC64-NEXT: store float [[X_REAL]], ptr [[COERCE_REALP]], align 4
// SPARC64-NEXT: store float [[X_IMAG]], ptr [[COERCE_IMAGP]], align 4
-// SPARC64-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw { float, float }, ptr [[COERCE]], i32 0, i32 0
-// SPARC64-NEXT: [[TMP2:%.*]] = load float, ptr [[TMP1]], align 4
-// SPARC64-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw { float, float }, ptr [[COERCE]], i32 0, i32 1
-// SPARC64-NEXT: [[TMP4:%.*]] = load float, ptr [[TMP3]], align 4
-// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, float inreg noundef [[TMP2]], float inreg noundef [[TMP4]])
+// SPARC64-NEXT: [[TMP1:%.*]] = load i64, ptr [[COERCE]], align 4
+// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, i64 noundef [[TMP1]])
// SPARC64-NEXT: ret void
//
void test_complex_float(va_list *ap) {
@@ -650,11 +647,8 @@ void test_complex_float(va_list *ap) {
// SPARC64-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i64 8
// SPARC64-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 8
// SPARC64-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 4 [[X]], ptr align 8 [[ARGP_CUR]], i64 8, i1 false)
-// SPARC64-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOAT_FLOAT]], ptr [[X]], i32 0, i32 0
-// SPARC64-NEXT: [[TMP2:%.*]] = load float, ptr [[TMP1]], align 4
-// SPARC64-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOAT_FLOAT]], ptr [[X]], i32 0, i32 1
-// SPARC64-NEXT: [[TMP4:%.*]] = load float, ptr [[TMP3]], align 4
-// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, float inreg [[TMP2]], float inreg [[TMP4]])
+// SPARC64-NEXT: [[TMP1:%.*]] = load i64, ptr [[X]], align 4
+// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, i64 [[TMP1]])
// SPARC64-NEXT: ret void
//
void test_float_float(va_list *ap) {
@@ -717,11 +711,11 @@ void test_float_float(va_list *ap) {
// SPARC64-NEXT: [[COERCE_IMAGP:%.*]] = getelementptr inbounds nuw { double, double }, ptr [[COERCE]], i32 0, i32 1
// SPARC64-NEXT: store double [[X_REAL]], ptr [[COERCE_REALP]], align 8
// SPARC64-NEXT: store double [[X_IMAG]], ptr [[COERCE_IMAGP]], align 8
-// SPARC64-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw { double, double }, ptr [[COERCE]], i32 0, i32 0
-// SPARC64-NEXT: [[TMP2:%.*]] = load double, ptr [[TMP1]], align 8
-// SPARC64-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw { double, double }, ptr [[COERCE]], i32 0, i32 1
-// SPARC64-NEXT: [[TMP4:%.*]] = load double, ptr [[TMP3]], align 8
-// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, double noundef [[TMP2]], double noundef [[TMP4]])
+// SPARC64-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw { i64, i64 }, ptr [[COERCE]], i32 0, i32 0
+// SPARC64-NEXT: [[TMP2:%.*]] = load i64, ptr [[TMP1]], align 8
+// SPARC64-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw { i64, i64 }, ptr [[COERCE]], i32 0, i32 1
+// SPARC64-NEXT: [[TMP4:%.*]] = load i64, ptr [[TMP3]], align 8
+// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, i64 noundef [[TMP2]], i64 noundef [[TMP4]])
// SPARC64-NEXT: ret void
//
void test_complex_double(va_list *ap) {
@@ -752,11 +746,11 @@ void test_complex_double(va_list *ap) {
// SPARC64-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i64 16
// SPARC64-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 8
// SPARC64-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[X]], ptr align 8 [[ARGP_CUR]], i64 16, i1 false)
-// SPARC64-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw [[STRUCT_DOUBLE_DOUBLE]], ptr [[X]], i32 0, i32 0
-// SPARC64-NEXT: [[TMP2:%.*]] = load double, ptr [[TMP1]], align 8
-// SPARC64-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw [[STRUCT_DOUBLE_DOUBLE]], ptr [[X]], i32 0, i32 1
-// SPARC64-NEXT: [[TMP4:%.*]] = load double, ptr [[TMP3]], align 8
-// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, double [[TMP2]], double [[TMP4]])
+// SPARC64-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw { i64, i64 }, ptr [[X]], i32 0, i32 0
+// SPARC64-NEXT: [[TMP2:%.*]] = load i64, ptr [[TMP1]], align 8
+// SPARC64-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw { i64, i64 }, ptr [[X]], i32 0, i32 1
+// SPARC64-NEXT: [[TMP4:%.*]] = load i64, ptr [[TMP3]], align 8
+// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, i64 [[TMP2]], i64 [[TMP4]])
// SPARC64-NEXT: ret void
//
void test_double_double(va_list *ap) {
@@ -925,13 +919,11 @@ void test_aligned_int(va_list *ap) {
// SPARC64-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR_ALIGNED]], i64 16
// SPARC64-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 8
// SPARC64-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 16 [[X]], ptr align 16 [[ARGP_CUR_ALIGNED]], i64 16, i1 false)
-// SPARC64-NEXT: [[TMP2:%.*]] = getelementptr inbounds nuw { float, i32, i64 }, ptr [[X]], i32 0, i32 0
-// SPARC64-NEXT: [[TMP3:%.*]] = load float, ptr [[TMP2]], align 16
-// SPARC64-NEXT: [[TMP4:%.*]] = getelementptr inbounds nuw { float, i32, i64 }, ptr [[X]], i32 0, i32 1
-// SPARC64-NEXT: [[TMP5:%.*]] = load i32, ptr [[TMP4]], align 4
-// SPARC64-NEXT: [[TMP6:%.*]] = getelementptr inbounds nuw { float, i32, i64 }, ptr [[X]], i32 0, i32 2
-// SPARC64-NEXT: [[TMP7:%.*]] = load i64, ptr [[TMP6]], align 8
-// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, i64 undef, float inreg [[TMP3]], i32 inreg [[TMP5]], i64 inreg [[TMP7]])
+// SPARC64-NEXT: [[TMP2:%.*]] = getelementptr inbounds nuw { i64, i64 }, ptr [[X]], i32 0, i32 0
+// SPARC64-NEXT: [[TMP3:%.*]] = load i64, ptr [[TMP2]], align 16
+// SPARC64-NEXT: [[TMP4:%.*]] = getelementptr inbounds nuw { i64, i64 }, ptr [[X]], i32 0, i32 1
+// SPARC64-NEXT: [[TMP5:%.*]] = load i64, ptr [[TMP4]], align 8
+// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, i64 undef, i64 [[TMP3]], i64 [[TMP5]])
// SPARC64-NEXT: ret void
//
void test_aligned_float(va_list *ap) {
@@ -997,15 +989,11 @@ void test_int_int_int_int(va_list *ap) {
// SPARC64-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i64 16
// SPARC64-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 8
// SPARC64-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 4 [[X]], ptr align 8 [[ARGP_CUR]], i64 16, i1 false)
-// SPARC64-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOAT_FLOAT_FLOAT_FLOAT]], ptr [[X]], i32 0, i32 0
-// SPARC64-NEXT: [[TMP2:%.*]] = load float, ptr [[TMP1]], align 4
-// SPARC64-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOAT_FLOAT_FLOAT_FLOAT]], ptr [[X]], i32 0, i32 1
-// SPARC64-NEXT: [[TMP4:%.*]] = load float, ptr [[TMP3]], align 4
-// SPARC64-NEXT: [[TMP5:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOAT_FLOAT_FLOAT_FLOAT]], ptr [[X]], i32 0, i32 2
-// SPARC64-NEXT: [[TMP6:%.*]] = load float, ptr [[TMP5]], align 4
-// SPARC64-NEXT: [[TMP7:%.*]] = getelementptr inbounds nuw [[STRUCT_FLOAT_FLOAT_FLOAT_FLOAT]], ptr [[X]], i32 0, i32 3
-// SPARC64-NEXT: [[TMP8:%.*]] = load float, ptr [[TMP7]], align 4
-// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, float inreg [[TMP2]], float inreg [[TMP4]], float inreg [[TMP6]], float inreg [[TMP8]])
+// SPARC64-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw { i64, i64 }, ptr [[X]], i32 0, i32 0
+// SPARC64-NEXT: [[TMP2:%.*]] = load i64, ptr [[TMP1]], align 4
+// SPARC64-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw { i64, i64 }, ptr [[X]], i32 0, i32 1
+// SPARC64-NEXT: [[TMP4:%.*]] = load i64, ptr [[TMP3]], align 4
+// SPARC64-NEXT: call void (i32, ...) @sink(i32 noundef signext 0, i64 [[TMP2]], i64 [[TMP4]])
// SPARC64-NEXT: ret void
//
void test_float_float_float_float(va_list *ap) {
``````````
</details>
https://github.com/llvm/llvm-project/pull/216524
More information about the cfe-commits
mailing list