[flang] AArch64 support for BIND(C) derived return types #114051
Merged
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
This patch adds support for BIND(C) derived types as return values matching the AArch64 Procedure Call Standard for C. Support for BIND(C) derived types as value parameters will be in a separate patch.
llvmbot
added
flang
|
@llvm/pr-subscribers-flang-codegen @llvm/pr-subscribers-flang-fir-hlfir Author: David Truby (DavidTruby) ChangesThis patch adds support for BIND(C) derived types as return values Support for BIND(C) derived types as value parameters will be in a Full diff: https://github.com/llvm/llvm-project/pull/114051.diff 2 Files Affected:
diff --git a/flang/lib/Optimizer/CodeGen/Target.cpp b/flang/lib/Optimizer/CodeGen/Target.cpp
index 6c148dffb0e55a..15ffdb74ef51d6 100644
--- a/flang/lib/Optimizer/CodeGen/Target.cpp
+++ b/flang/lib/Optimizer/CodeGen/Target.cpp
@@ -825,6 +825,48 @@ struct TargetAArch64 : public GenericTarget<TargetAArch64> {
}
return marshal;
}
+
+ static bool isHFA(fir::RecordType ty) {
+ auto types = ty.getTypeList();
+ if (types.empty() || types.size() > 4) {
+ return false;
+ }
+
+ if (!isa_real(types.front().second)) {
+ types.front().second.dump();
+ return false;
+ }
+
+ return llvm::all_equal(llvm::make_second_range(types));
+ }
+
+ CodeGenSpecifics::Marshalling
+ structReturnType(mlir::Location loc, fir::RecordType ty) const override {
+ CodeGenSpecifics::Marshalling marshal;
+
+ if (isHFA(ty)) {
+ auto newTy = fir::SequenceType::get({ty.getNumFields()}, ty.getType(0));
+ marshal.emplace_back(newTy, AT{});
+ return marshal;
+ }
+
+ auto [size, align] =
+ fir::getTypeSizeAndAlignmentOrCrash(loc, ty, getDataLayout(), kindMap);
+
+ // return in registers if size <= 16 bytes
+ if (size <= 16) {
+ auto dwordSize = (size + 7) / 8;
+ auto newTy = fir::SequenceType::get(
+ dwordSize, mlir::IntegerType::get(ty.getContext(), 64));
+ marshal.emplace_back(newTy, AT{});
+ return marshal;
+ }
+
+ unsigned short stackAlign = std::max<unsigned short>(align, 8u);
+ marshal.emplace_back(fir::ReferenceType::get(ty),
+ AT{stackAlign, false, true});
+ return marshal;
+ }
};
} // namespace
diff --git a/flang/test/Fir/struct-return-aarch64.fir b/flang/test/Fir/struct-return-aarch64.fir
new file mode 100644
index 00000000000000..96f2f9999b3435
--- /dev/null
+++ b/flang/test/Fir/struct-return-aarch64.fir
@@ -0,0 +1,156 @@
+// Test AArch64 ABI rewrite of struct returned by value (BIND(C), VALUE derived types).
+// RUN: fir-opt --target-rewrite="target=aarch64-unknown-linux-gnu" %s | FileCheck %s
+
+!composite = !fir.type<t1{i:f32,j:i32,k:f32}>
+// CHECK-LABEL: func.func private @test_composite() -> !fir.array<2xi64>
+func.func private @test_composite() -> !composite
+// CHECK-LABEL: func.func @test_call_composite(
+// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.type<t1{i:f32,j:i32,k:f32}>>)
+func.func @test_call_composite(%arg0 : !fir.ref<!composite>) {
+ // CHECK: %[[OUT:.*]] = fir.call @test_composite() : () -> !fir.array<2xi64>
+ // CHECK: %[[STACK:.*]] = llvm.intr.stacksave : !llvm.ptr
+ // CHECK: %[[ARR:.*]] = fir.alloca !fir.array<2xi64>
+ // CHECK: fir.store %[[OUT]] to %[[ARR]] : !fir.ref<!fir.array<2xi64>>
+ // CHECK: %[[CVT:.*]] = fir.convert %[[ARR]] : (!fir.ref<!fir.array<2xi64>>) -> !fir.ref<!fir.type<t1{i:f32,j:i32,k:f32}>>
+ // CHECK: %[[LD:.*]] = fir.load %[[CVT]] : !fir.ref<!fir.type<t1{i:f32,j:i32,k:f32}>>
+ // CHECK: llvm.intr.stackrestore %[[STACK]] : !llvm.ptr
+ %out = fir.call @test_composite() : () -> !composite
+ // CHECK: fir.store %[[LD]] to %[[ARG0]] : !fir.ref<!fir.type<t1{i:f32,j:i32,k:f32}>>
+ fir.store %out to %arg0 : !fir.ref<!composite>
+ // CHECK: return
+ return
+}
+
+!hfa_f16 = !fir.type<t2{x:f16, y:f16}>
+// CHECK-LABEL: func.func private @test_hfa_f16() -> !fir.array<2xf16>
+func.func private @test_hfa_f16() -> !hfa_f16
+// CHECK-LABEL: func.func @test_call_hfa_f16(
+// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.type<t2{x:f16,y:f16}>>) {
+func.func @test_call_hfa_f16(%arg0 : !fir.ref<!hfa_f16>) {
+ // CHECK: %[[OUT:.*]] = fir.call @test_hfa_f16() : () -> !fir.array<2xf16>
+ // CHECK: %[[STACK:.*]] = llvm.intr.stacksave : !llvm.ptr
+ // CHECK: %[[ARR:.*]] = fir.alloca !fir.array<2xf16>
+ // CHECK: fir.store %[[OUT]] to %[[ARR]] : !fir.ref<!fir.array<2xf16>>
+ // CHECK: %[[CVT:.*]] = fir.convert %[[ARR]] : (!fir.ref<!fir.array<2xf16>>) -> !fir.ref<!fir.type<t2{x:f16,y:f16}>>
+ // CHECK: %[[LD:.*]] = fir.load %[[CVT]] : !fir.ref<!fir.type<t2{x:f16,y:f16}>>
+ // CHECK: llvm.intr.stackrestore %[[STACK]] : !llvm.ptr
+ %out = fir.call @test_hfa_f16() : () -> !hfa_f16
+ // CHECK: fir.store %[[LD]] to %[[ARG0]] : !fir.ref<!fir.type<t2{x:f16,y:f16}>>
+ fir.store %out to %arg0 : !fir.ref<!hfa_f16>
+ return
+}
+
+!hfa_f32 = !fir.type<t3{w:f32, x:f32, y:f32, z:f32}>
+// CHECK-LABEL: func.func private @test_hfa_f32() -> !fir.array<4xf32>
+func.func private @test_hfa_f32() -> !hfa_f32
+// CHECK-LABEL: func.func @test_call_hfa_f32(
+// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.type<t3{w:f32,x:f32,y:f32,z:f32}>>) {
+func.func @test_call_hfa_f32(%arg0 : !fir.ref<!hfa_f32>) {
+ // CHECK: %[[OUT:.*]] = fir.call @test_hfa_f32() : () -> !fir.array<4xf32>
+ // CHECK: %[[STACK:.*]] = llvm.intr.stacksave : !llvm.ptr
+ // CHECK: %[[ARR:.*]] = fir.alloca !fir.array<4xf32>
+ // CHECK: fir.store %[[OUT]] to %[[ARR]] : !fir.ref<!fir.array<4xf32>>
+ // CHECK: %[[CVT:.*]] = fir.convert %[[ARR]] : (!fir.ref<!fir.array<4xf32>>) -> !fir.ref<!fir.type<t3{w:f32,x:f32,y:f32,z:f32}>>
+ // CHECK: %[[LD:.*]] = fir.load %[[CVT]] : !fir.ref<!fir.type<t3{w:f32,x:f32,y:f32,z:f32}>>
+ // CHECK: llvm.intr.stackrestore %[[STACK]] : !llvm.ptr
+ %out = fir.call @test_hfa_f32() : () -> !hfa_f32
+ // CHECK: fir.store %[[LD]] to %[[ARG0]] : !fir.ref<!fir.type<t3{w:f32,x:f32,y:f32,z:f32}>>
+ fir.store %out to %arg0 : !fir.ref<!hfa_f32>
+ return
+}
+
+!hfa_f64 = !fir.type<t4{x:f64, y:f64, z:f64}>
+// CHECK-LABEL: func.func private @test_hfa_f64() -> !fir.array<3xf64>
+func.func private @test_hfa_f64() -> !hfa_f64
+// CHECK-LABEL: func.func @test_call_hfa_f64(
+// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.type<t4{x:f64,y:f64,z:f64}>>)
+func.func @test_call_hfa_f64(%arg0 : !fir.ref<!hfa_f64>) {
+ // CHECK: %[[OUT:.*]] = fir.call @test_hfa_f64() : () -> !fir.array<3xf64>
+ // CHECK: %[[STACK:.*]] = llvm.intr.stacksave : !llvm.ptr
+ // CHECK: %[[ARR:.*]] = fir.alloca !fir.array<3xf64>
+ // CHECK: fir.store %[[OUT]] to %[[ARR]] : !fir.ref<!fir.array<3xf64>>
+ // CHECK: %[[CVT:.*]] = fir.convert %[[ARR]] : (!fir.ref<!fir.array<3xf64>>) -> !fir.ref<!fir.type<t4{x:f64,y:f64,z:f64}>>
+ // CHECK: %[[LD:.*]] = fir.load %[[CVT]] : !fir.ref<!fir.type<t4{x:f64,y:f64,z:f64}>>
+ // CHECK: llvm.intr.stackrestore %[[STACK]] : !llvm.ptr
+ %out = fir.call @test_hfa_f64() : () -> !hfa_f64
+ // CHECK: fir.store %[[LD]] to %[[ARG0]] : !fir.ref<!fir.type<t4{x:f64,y:f64,z:f64}>>
+ fir.store %out to %arg0 : !fir.ref<!hfa_f64>
+ return
+}
+
+!hfa_f128 = !fir.type<t5{w:f128, x:f128, y:f128, z:f128}>
+// CHECK-LABEL: func.func private @test_hfa_f128() -> !fir.array<4xf128>
+func.func private @test_hfa_f128() -> !hfa_f128
+// CHECK-LABEL: func.func @test_call_hfa_f128(
+// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.type<t5{w:f128,x:f128,y:f128,z:f128}>>) {
+func.func @test_call_hfa_f128(%arg0 : !fir.ref<!hfa_f128>) {
+ // CHECK: %[[OUT:.*]] = fir.call @test_hfa_f128() : () -> !fir.array<4xf128>
+ // CHECK: %[[STACK:.*]] = llvm.intr.stacksave : !llvm.ptr
+ // CHECK: %[[ARR:.*]] = fir.alloca !fir.array<4xf128>
+ // CHECK: fir.store %[[OUT]] to %[[ARR]] : !fir.ref<!fir.array<4xf128>>
+ // CHECK: %[[CVT:.*]] = fir.convert %[[ARR]] : (!fir.ref<!fir.array<4xf128>>) -> !fir.ref<!fir.type<t5{w:f128,x:f128,y:f128,z:f128}>>
+ // CHECK: %[[LD:.*]] = fir.load %[[CVT]] : !fir.ref<!fir.type<t5{w:f128,x:f128,y:f128,z:f128}>>
+ // CHECK: llvm.intr.stackrestore %[[STACK]] : !llvm.ptr
+ %out = fir.call @test_hfa_f128() : () -> !hfa_f128
+ // CHECK: fir.store %[[LD]] to %[[ARG0]] : !fir.ref<!fir.type<t5{w:f128,x:f128,y:f128,z:f128}>>
+ fir.store %out to %arg0 : !fir.ref<!hfa_f128>
+ return
+}
+
+!hfa_bf16 = !fir.type<t6{w:bf16, x:bf16, y:bf16, z:bf16}>
+// CHECK-LABEL: func.func private @test_hfa_bf16() -> !fir.array<4xbf16>
+func.func private @test_hfa_bf16() -> !hfa_bf16
+// CHECK-LABEL: func.func @test_call_hfa_bf16(
+// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.type<t6{w:bf16,x:bf16,y:bf16,z:bf16}>>) {
+func.func @test_call_hfa_bf16(%arg0 : !fir.ref<!hfa_bf16>) {
+ // CHECK: %[[OUT:.*]] = fir.call @test_hfa_bf16() : () -> !fir.array<4xbf16>
+ // CHECK: %[[STACK:.*]] = llvm.intr.stacksave : !llvm.ptr
+ // CHECK: %[[ARR:.*]] = fir.alloca !fir.array<4xbf16>
+ // CHECK: fir.store %[[OUT]] to %[[ARR]] : !fir.ref<!fir.array<4xbf16>>
+ // CHECK: %[[CVT:.*]] = fir.convert %[[ARR]] : (!fir.ref<!fir.array<4xbf16>>) -> !fir.ref<!fir.type<t6{w:bf16,x:bf16,y:bf16,z:bf16}>>
+ // CHECK: %[[LD:.*]] = fir.load %[[CVT]] : !fir.ref<!fir.type<t6{w:bf16,x:bf16,y:bf16,z:bf16}>>
+ // CHECK: llvm.intr.stackrestore %[[STACK]] : !llvm.ptr
+ %out = fir.call @test_hfa_bf16() : () -> !hfa_bf16
+ // CHECK: fir.store %[[LD]] to %[[ARG0]] : !fir.ref<!fir.type<t6{w:bf16,x:bf16,y:bf16,z:bf16}>>
+ fir.store %out to %arg0 : !fir.ref<!hfa_bf16>
+ return
+}
+
+!too_big = !fir.type<t7{x:i64, y:i64, z:i64}>
+// CHECK-LABEL: func.func private @test_too_big(!fir.ref<!fir.type<t7{x:i64,y:i64,z:i64}>>
+// CHECK-SAME: {llvm.align = 8 : i32, llvm.sret = !fir.type<t7{x:i64,y:i64,z:i64}>})
+func.func private @test_too_big() -> !too_big
+// CHECK-LABEL: func.func @test_call_too_big(
+// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.type<t7{x:i64,y:i64,z:i64}>>) {
+func.func @test_call_too_big(%arg0 : !fir.ref<!too_big>) {
+ // CHECK: %[[STACK:.*]] = llvm.intr.stacksave : !llvm.ptr
+ // CHECK: %[[ARG:.*]] = fir.alloca !fir.type<t7{x:i64,y:i64,z:i64}>
+ // CHECK: fir.call @test_too_big(%[[ARG]]) : (!fir.ref<!fir.type<t7{x:i64,y:i64,z:i64}>>) -> ()
+ // CHECK: %[[CVT:.*]] = fir.convert %[[ARG]] : (!fir.ref<!fir.type<t7{x:i64,y:i64,z:i64}>>) -> !fir.ref<!fir.type<t7{x:i64,y:i64,z:i64}>>
+ // CHECK: %[[LD:.*]] = fir.load %[[CVT]] : !fir.ref<!fir.type<t7{x:i64,y:i64,z:i64}>>
+ // CHECK: llvm.intr.stackrestore %[[STACK]] : !llvm.ptr
+ %out = fir.call @test_too_big() : () -> !too_big
+ // CHECK: fir.store %[[LD]] to %[[ARG0]] : !fir.ref<!fir.type<t7{x:i64,y:i64,z:i64}>>
+ fir.store %out to %arg0 : !fir.ref<!too_big>
+ return
+}
+
+
+!too_big_hfa = !fir.type<t8{i:!fir.array<5xf32>}>
+// CHECK-LABEL: func.func private @test_too_big_hfa(!fir.ref<!fir.type<t8{i:!fir.array<5xf32>}>>
+// CHECK-SAME: {llvm.align = 8 : i32, llvm.sret = !fir.type<t8{i:!fir.array<5xf32>}>})
+func.func private @test_too_big_hfa() -> !too_big_hfa
+// CHECK-LABEL: func.func @test_call_too_big_hfa(
+// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.type<t8{i:!fir.array<5xf32>}>>) {
+func.func @test_call_too_big_hfa(%arg0 : !fir.ref<!too_big_hfa>) {
+ // CHECK: %[[STACK:.*]] = llvm.intr.stacksave : !llvm.ptr
+ // CHECK: %[[ARG:.*]] = fir.alloca !fir.type<t8{i:!fir.array<5xf32>}>
+ // CHECK: fir.call @test_too_big_hfa(%[[ARG]]) : (!fir.ref<!fir.type<t8{i:!fir.array<5xf32>}>>) -> ()
+ // CHECK: %[[CVT:.*]] = fir.convert %[[ARG]] : (!fir.ref<!fir.type<t8{i:!fir.array<5xf32>}>>) -> !fir.ref<!fir.type<t8{i:!fir.array<5xf32>}>>
+ // CHECK: %[[LD:.*]] = fir.load %[[CVT]] : !fir.ref<!fir.type<t8{i:!fir.array<5xf32>}>>
+ // CHECK: llvm.intr.stackrestore %[[STACK]] : !llvm.ptr
+ %out = fir.call @test_too_big_hfa() : () -> !too_big_hfa
+ // CHECK: fir.store %[[LD]] to %[[ARG0]] : !fir.ref<!fir.type<t8{i:!fir.array<5xf32>}>>
+ fir.store %out to %arg0 : !fir.ref<!too_big_hfa>
+ return
+}
|
DavidTruby
requested review from
tblah,
jeanPerier,
pawosm-arm,
vzakhari and
kiranchandramohan
|
The ABI for return values here is also the same on Windows so I have not separated out Windows from non-Windows as is done on x86. |
|
Please could you link to some public documentation for the ABI? |
| @@ -825,6 +825,48 @@ struct TargetAArch64 : public GenericTarget<TargetAArch64> { | |||
| } | |||
| return marshal; | |||
| } | |||
|
|
|||
| static bool isHFA(fir::RecordType ty) { | |||
There was a problem hiding this comment.
What is HFA? Please document.
There was a problem hiding this comment.
@DavidTruby I think you missed this when responding to feedback
There was a problem hiding this comment.
Sorry I commented on the wrong one here. I meant the "expand auto" nit below.
| @@ -825,6 +825,48 @@ struct TargetAArch64 : public GenericTarget<TargetAArch64> { | |||
| } | |||
| return marshal; | |||
| } | |||
|
|
|||
| static bool isHFA(fir::RecordType ty) { | |||
| auto types = ty.getTypeList(); | |||
There was a problem hiding this comment.
Nit:expand auto.
| } | ||
|
|
||
| if (!isa_real(types.front().second)) { | ||
| types.front().second.dump(); |
There was a problem hiding this comment.
Is this leftover debug?
|
|
||
| // return in registers if size <= 16 bytes | ||
| if (size <= 16) { | ||
| auto dwordSize = (size + 7) / 8; |
There was a problem hiding this comment.
Nit:Spell the type.
tblah
approved these changes
Oct 30, 2024
|
|
||
| unsigned short stackAlign = std::max<unsigned short>(align, 8u); | ||
| marshal.emplace_back(fir::ReferenceType::get(ty), | ||
| AT{stackAlign, false, true}); |
There was a problem hiding this comment.
Suggested change
| AT{stackAlign, false, true}); | |
| AT{/*alignment=*/stackAlign, /*byval=*/false, /*sret=*/true}); |
jeanPerier
reviewed
Nov 4, 2024
|
|
||
| if (!isa_real(types.front().second)) { | ||
| return false; | ||
| } |
There was a problem hiding this comment.
Derived types containing small fp arrays will be rejected here, is the following C struct an HFA? :
typedef struct {
float x[2];
float y;
} S;
There was a problem hiding this comment.
Ah. It is, yes. As is:
typedef struct {
float x;
float y;
} S1;
typedef struct {
S1 s;
float z;
} S2;
Good spot.
|
I'm not all that satisfied with how I did this, I feel like I have over complicated it, although it does work. If anyone has any suggestions I'd happily take them! |
tblah
reviewed
Nov 13, 2024
|
@jeanPerier does this look ok to you now? |
jeanPerier
approved these changes
Nov 19, 2024
There was a problem hiding this comment.
Yes, that for addressing the comment.
Building the flatten type list just for the HFA check is indeed a bit heavy, if this was expected to be exercised for types with many components, I would advise some kind of type visitor without copy and with early exit, but given this is not common and you only call it for types with less than 4 components, it is just fine as it is.
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.
This patch adds support for BIND(C) derived types as return values
matching the AArch64 Procedure Call Standard for C.
Support for BIND(C) derived types as value parameters will be in a
separate patch.