Skip to content

[CIR] Fix Complex emit promotion for Div op #156963

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Sep 5, 2025
Merged

[CIR] Fix Complex emit promotion for Div op #156963

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions clang/lib/CIR/CodeGen/CIRGenExprComplex.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -749,6 +749,7 @@ mlir::Value ComplexExprEmitter::emitPromoted(const Expr *e,
HANDLE_BINOP(Add)
HANDLE_BINOP(Sub)
HANDLE_BINOP(Mul)
HANDLE_BINOP(Div)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't understand what this is doing. The test you added compiles without this change, but produces different results. Can you explain what this changes?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For this test case, when the promotion type is not null

The old code path.

  • in emitPromoted will go to the else path
mlir::Value result = Visit(const_cast<Expr *>(e));
  if (!promotionTy.isNull())
    return cgf.emitPromotedValue(result, promotionTy);

Which means visitBinDivOp, which will call emitBinOps and will promote the type f16 -> f32, then will un-promote it

if (!promotionTy.isNull())
    result = cgf.emitUnPromotedValue(result, e->getType());

And when it goes back to emitPromoted after the visit call it will promote it again (The first code snippet)

With the current solution, it will promote the type and go directly to emitBinDiv, then early return from emitPromoted that eliminates unnecessary casts

When I print with passes disabled, I found that the old approach emits 2 more casts to promote and unpromote the type

#undef HANDLE_BINOP
default:
break;
Expand Down
125 changes: 122 additions & 3 deletions clang/test/CIR/CodeGen/complex-compound-assignment.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,9 +198,9 @@ void foo3() {
// LLVM: %[[RESULT_REAL:.*]] = extractvalue { float, float } %[[RESULT]], 0
// LLVM: %[[RESULT_IMAG:.*]] = extractvalue { float, float } %[[RESULT]], 1
// LLVM: %[[RESULT_REAL_F16:.*]] = fptrunc float %[[RESULT_REAL]] to half
// LLVM: %[[RESULT_IMAG_F26:.*]] = fptrunc float %[[RESULT_IMAG]] to half
// LLVM: %[[RESULT_IMAG_F16:.*]] = fptrunc float %[[RESULT_IMAG]] to half
// LLVM: %[[TMP_RESULT_F16:.*]] = insertvalue { half, half } undef, half %[[RESULT_REAL_F16]], 0
// LLVM: %[[RESULT_F16:.*]] = insertvalue { half, half } %29, half %[[RESULT_IMAG_F26]], 1
// LLVM: %[[RESULT_F16:.*]] = insertvalue { half, half } %29, half %[[RESULT_IMAG_F16]], 1
// LLVM: store { half, half } %[[RESULT_F16]], ptr %[[B_ADDR]], align 2

// OGCG: %[[A_ADDR:.*]] = alloca { half, half }, align 2
Expand Down Expand Up @@ -701,6 +701,125 @@ void foo12() {
// OGCG: store i32 %[[RESULT_REAL]], ptr %[[A_REAL_PTR]], align 4
// OGCG: store i32 %[[RESULT_IMAG]], ptr %[[A_IMAG_PTR]], align 4

void foo13() {
_Float16 _Complex a;
_Float16 _Complex b;
b /= (a / b);
}

// CIR: %[[A_ADDR:.*]] = cir.alloca !cir.complex<!cir.f16>, !cir.ptr<!cir.complex<!cir.f16>>, ["a"]
// CIR: %[[B_ADDR:.*]] = cir.alloca !cir.complex<!cir.f16>, !cir.ptr<!cir.complex<!cir.f16>>, ["b"]
// CIR: %[[TMP_A:.*]] = cir.load{{.*}} %[[A_ADDR]] : !cir.ptr<!cir.complex<!cir.f16>>, !cir.complex<!cir.f16>
// CIR: %[[A_REAL:.*]] = cir.complex.real %[[TMP_A]] : !cir.complex<!cir.f16> -> !cir.f16
// CIR: %[[A_IMAG:.*]] = cir.complex.imag %[[TMP_A]] : !cir.complex<!cir.f16> -> !cir.f16
// CIR: %[[A_REAL_F32:.*]] = cir.cast(floating, %[[A_REAL]] : !cir.f16), !cir.float
// CIR: %[[A_IMAG_F32:.*]] = cir.cast(floating, %[[A_IMAG]] : !cir.f16), !cir.float
// CIR: %[[A_COMPLEX_F32:.*]] = cir.complex.create %[[A_REAL_F32]], %[[A_IMAG_F32]] : !cir.float -> !cir.complex<!cir.float>
// CIR: %[[TMP_B:.*]] = cir.load{{.*}} %[[B_ADDR]] : !cir.ptr<!cir.complex<!cir.f16>>, !cir.complex<!cir.f16>
// CIR: %[[B_REAL:.*]] = cir.complex.real %[[TMP_B]] : !cir.complex<!cir.f16> -> !cir.f16
// CIR: %[[B_IMAG:.*]] = cir.complex.imag %[[TMP_B]] : !cir.complex<!cir.f16> -> !cir.f16
// CIR: %[[B_REAL_F32:.*]] = cir.cast(floating, %[[B_REAL]] : !cir.f16), !cir.float
// CIR: %[[B_IMAG_F32:.*]] = cir.cast(floating, %[[B_IMAG]] : !cir.f16), !cir.float
// CIR: %[[B_COMPLEX_F32:.*]] = cir.complex.create %[[B_REAL_F32]], %[[B_IMAG_F32]] : !cir.float -> !cir.complex<!cir.float>
// CIR: %[[A_REAL_F32:.*]] = cir.complex.real %[[A_COMPLEX_F32]] : !cir.complex<!cir.float> -> !cir.float
// CIR: %[[A_IMAG_F32:.*]] = cir.complex.imag %[[A_COMPLEX_F32]] : !cir.complex<!cir.float> -> !cir.float
// CIR: %[[B_REAL_F32:.*]] = cir.complex.real %[[B_COMPLEX_F32]] : !cir.complex<!cir.float> -> !cir.float
// CIR: %[[B_IMAG_F32:.*]] = cir.complex.imag %[[B_COMPLEX_F32]] : !cir.complex<!cir.float> -> !cir.float
// CIR: %[[DIV_A_B:.*]] = cir.call @__divsc3(%[[A_REAL_F32]], %[[A_IMAG_F32]], %[[B_REAL_F32]], %[[B_IMAG_F32]]) : (!cir.float, !cir.float, !cir.float, !cir.float) -> !cir.complex<!cir.float>
// CIR: %[[TMP_B:.*]] = cir.load{{.*}} %[[B_ADDR]] : !cir.ptr<!cir.complex<!cir.f16>>, !cir.complex<!cir.f16>
// CIR: %[[B_REAL:.*]] = cir.complex.real %[[TMP_B]] : !cir.complex<!cir.f16> -> !cir.f16
// CIR: %[[B_IMAG:.*]] = cir.complex.imag %[[TMP_B]] : !cir.complex<!cir.f16> -> !cir.f16
// CIR: %[[B_REAL_F32:.*]] = cir.cast(floating, %[[B_REAL]] : !cir.f16), !cir.float
// CIR: %[[B_IMAG_F32:.*]] = cir.cast(floating, %[[B_IMAG]] : !cir.f16), !cir.float
// CIR: %[[B_COMPLEX_F32:.*]] = cir.complex.create %[[B_REAL_F32]], %[[B_IMAG_F32]] : !cir.float -> !cir.complex<!cir.float>
// CIR: %[[B_REAL_F32:.*]] = cir.complex.real %[[B_COMPLEX_F32]] : !cir.complex<!cir.float> -> !cir.float
// CIR: %[[B_IMAG_F32:.*]] = cir.complex.imag %[[B_COMPLEX_F32]] : !cir.complex<!cir.float> -> !cir.float
// CIR: %[[DIV_AB_REAL:.*]] = cir.complex.real %[[DIV_A_B]] : !cir.complex<!cir.float> -> !cir.float
// CIR: %[[DIV_AB_IMAG:.*]] = cir.complex.imag %[[DIV_A_B]] : !cir.complex<!cir.float> -> !cir.float
// CIR: %[[RESULT:.*]] = cir.call @__divsc3(%[[B_REAL_F32]], %[[B_IMAG_F32]], %[[DIV_AB_REAL]], %[[DIV_AB_IMAG]]) : (!cir.float, !cir.float, !cir.float, !cir.float) -> !cir.complex<!cir.float>
// CIR: %[[RESULT_REAL_F32:.*]] = cir.complex.real %[[RESULT]] : !cir.complex<!cir.float> -> !cir.float
// CIR: %[[RESULT_IMAG_F32:.*]] = cir.complex.imag %[[RESULT]] : !cir.complex<!cir.float> -> !cir.float
// CIR: %[[RESULT_REAL_F16:.*]] = cir.cast(floating, %[[RESULT_REAL_F32]] : !cir.float), !cir.f16
// CIR: %[[RESULT_IMAG_F16:.*]] = cir.cast(floating, %[[RESULT_IMAG_F32]] : !cir.float), !cir.f16
// CIR: %[[RESULT_COMPLEX_F16:.*]] = cir.complex.create %[[RESULT_REAL_F16]], %[[RESULT_IMAG_F16]] : !cir.f16 -> !cir.complex<!cir.f16>
// CIR: cir.store{{.*}} %[[RESULT_COMPLEX_F16]], %[[B_ADDR]] : !cir.complex<!cir.f16>, !cir.ptr<!cir.complex<!cir.f16>>

// LLVM: %[[A_ADDR:.*]] = alloca { half, half }, i64 1, align 2
// LLVM: %[[B_ADDR:.*]] = alloca { half, half }, i64 1, align 2
// LLVM: %[[TMP_A:.*]] = load { half, half }, ptr %[[A_ADDR]], align 2
// LLVM: %[[A_REAL:.*]] = extractvalue { half, half } %[[TMP_A]], 0
// LLVM: %[[A_IMAG:.*]] = extractvalue { half, half } %[[TMP_A]], 1
// LLVM: %[[A_REAL_F32:.*]] = fpext half %[[A_REAL]] to float
// LLVM: %[[A_IMAG_F32:.*]] = fpext half %[[A_IMAG]] to float
// LLVM: %[[TMP_A_COMPLEX_F32:.*]] = insertvalue { float, float } {{.*}}, float %[[A_REAL_F32]], 0
// LLVM: %[[A_COMPLEX_F32:.*]] = insertvalue { float, float } %[[TMP_A_COMPLEX_F32]], float %[[A_IMAG_F32]], 1
// LLVM: %[[TMP_B:.*]] = load { half, half }, ptr %[[B_ADDR]], align 2
// LLVM: %[[B_REAL:.*]] = extractvalue { half, half } %[[TMP_B]], 0
// LLVM: %[[B_IMAG:.*]] = extractvalue { half, half } %[[TMP_B]], 1
// LLVM: %[[B_REAL_F32:.*]] = fpext half %[[B_REAL]] to float
// LLVM: %[[B_IMAG_F32:.*]] = fpext half %[[B_IMAG]] to float
// LLVM: %[[TMP_B_COMPLEX_F32:.*]] = insertvalue { float, float } {{.*}}, float %[[B_REAL_F32]], 0
// LLVM: %[[B_COMPLEX_F32:.*]] = insertvalue { float, float } %[[TMP_B_COMPLEX_F32]], float %[[B_IMAG_F32]], 1
// LLVM: %[[DIV_A_B:.*]] = call { float, float } @__divsc3(float %[[A_REAL_F32]], float %[[A_IMAG_F32]], float %[[B_REAL_F32]], float %[[B_IMAG_F32]])
// LLVM: %[[TMP_B:.*]] = load { half, half }, ptr %[[B_ADDR]], align 2
// LLVM: %[[B_REAL:.*]] = extractvalue { half, half } %[[TMP_B]], 0
// LLVM: %[[B_IMAG:.*]] = extractvalue { half, half } %[[TMP_B]], 1
// LLVM: %[[B_REAL_F32:.*]] = fpext half %[[B_REAL]] to float
// LLVM: %[[B_IMAG_F32:.*]] = fpext half %[[B_IMAG]] to float
// LLVM: %[[TMP_B_COMPLEX_F32:.*]] = insertvalue { float, float } {{.*}}, float %[[B_REAL_F32]], 0
// LLVM: %[[B_COMPLEX_F32:.*]] = insertvalue { float, float } %[[TMP_B_COMPLEX_F32]], float %[[B_IMAG_F32]], 1
// LLVM: %[[DIV_AB_REAL:.*]] = extractvalue { float, float } %[[DIV_A_B]], 0
// LLVM: %[[DIV_AB_IMAG:.*]] = extractvalue { float, float } %[[DIV_A_B]], 1
// LLVM: %[[RESULT:.*]] = call { float, float } @__divsc3(float %[[B_REAL_F32]], float %[[B_IMAG_F32]], float %[[DIV_AB_REAL]], float %[[DIV_AB_IMAG]])
// LLVM: %[[RESULT_REAL_F32:.*]] = extractvalue { float, float } %[[RESULT]], 0
// LLVM: %[[RESULT_IMAG_F32:.*]] = extractvalue { float, float } %[[RESULT]], 1
// LLVM: %[[RESULT_REAL_F16:.*]] = fptrunc float %[[RESULT_REAL_F32]] to half
// LLVM: %[[RESULT_IMAG_F16:.*]] = fptrunc float %[[RESULT_IMAG_F32]] to half
// LLVM: %[[TMP_RESULT_F16:.*]] = insertvalue { half, half } {{.*}}, half %[[RESULT_REAL_F16]], 0
// LLVM: %[[RESULT_COMPLEX_F16:.*]] = insertvalue { half, half } %[[TMP_RESULT_F16]], half %[[RESULT_IMAG_F16]], 1
// LLVM: store { half, half } %[[RESULT_COMPLEX_F16]], ptr %[[B_ADDR]], align 2

// OGCG: %[[A_ADDR:.*]] = alloca { half, half }, align 2
// OGCG: %[[B_ADDR:.*]] = alloca { half, half }, align 2
// OGCG: %[[DIV_AB_ADDR:.*]] = alloca { float, float }, align 4
// OGCG: %[[DIV_B_AB_ADDR:.*]] = alloca { float, float }, align 4
// OGCG: %[[A_REAL_PTR:.*]] = getelementptr inbounds nuw { half, half }, ptr %[[A_ADDR]], i32 0, i32 0
// OGCG: %[[A_REAL:.*]] = load half, ptr %[[A_REAL_PTR]], align 2
// OGCG: %[[A_IMAG_PTR:.*]] = getelementptr inbounds nuw { half, half }, ptr %[[A_ADDR]], i32 0, i32 1
// OGCG: %[[A_IMAG:.*]] = load half, ptr %[[A_IMAG_PTR]], align 2
// OGCG: %[[A_REAL_F32:.*]] = fpext half %[[A_REAL]] to float
// OGCG: %[[A_IMAG_F32:.*]] = fpext half %[[A_IMAG]] to float
// OGCG: %[[B_REAL_PTR:.*]] = getelementptr inbounds nuw { half, half }, ptr %[[B_ADDR]], i32 0, i32 0
// OGCG: %[[B_REAL:.*]] = load half, ptr %[[B_REAL_PTR]], align 2
// OGCG: %[[B_IMAG_PTR:.*]] = getelementptr inbounds nuw { half, half }, ptr %[[B_ADDR]], i32 0, i32 1
// OGCG: %[[B_IMAG:.*]] = load half, ptr %[[B_IMAG_PTR]], align 2
// OGCG: %[[B_REAL_F32:.*]] = fpext half %[[B_REAL]] to float
// OGCG: %[[B_IMAG_F32:.*]] = fpext half %[[B_IMAG]] to float
// OGCG: %[[DIV_A_B:.*]] = call{{.*}} <2 x float> @__divsc3(float noundef %[[A_REAL_F32]], float noundef %[[A_IMAG_F32]], float noundef %[[B_REAL_F32]], float noundef %[[B_IMAG_F32]])
// OGCG: store <2 x float> %[[DIV_A_B]], ptr %[[DIV_AB_ADDR]], align 4
// OGCG: %[[DIV_AB_REAL_PTR:.*]] = getelementptr inbounds nuw { float, float }, ptr %[[DIV_AB_ADDR]], i32 0, i32 0
// OGCG: %[[DIV_AB_REAL:.*]] = load float, ptr %[[DIV_AB_REAL_PTR]], align 4
// OGCG: %[[DIV_AB_IMAG_PTR:.*]] = getelementptr inbounds nuw { float, float }, ptr %[[DIV_AB_ADDR]], i32 0, i32 1
// OGCG: %[[DIV_AB_IMAG:.*]] = load float, ptr %[[DIV_AB_IMAG_PTR]], align 4
// OGCG: %[[B_REAL_PTR:.*]] = getelementptr inbounds nuw { half, half }, ptr %[[B_ADDR]], i32 0, i32 0
// OGCG: %[[B_REAL:.*]] = load half, ptr %[[B_REAL_PTR]], align 2
// OGCG: %[[B_IMAG_PTR:.*]] = getelementptr inbounds nuw { half, half }, ptr %[[B_ADDR]], i32 0, i32 1
// OGCG: %[[B_IMAG:.*]] = load half, ptr %[[B_IMAG_PTR]], align 2
// OGCG: %[[B_REAL_F32:.*]] = fpext half %[[B_REAL]] to float
// OGCG: %[[B_IMAG_F32:.*]] = fpext half %[[B_IMAG]] to float
// OGCG: %[[RESULT:.*]] = call{{.*}} <2 x float> @__divsc3(float noundef %[[B_REAL_F32]], float noundef %[[B_IMAG_F32]], float noundef %[[DIV_AB_REAL]], float noundef %[[DIV_AB_IMAG]])
// OGCG: store <2 x float> %[[RESULT]], ptr %[[DIV_B_AB_ADDR]], align 4
// OGCG: %[[RESULT_REAL_PTR:.*]] = getelementptr inbounds nuw { float, float }, ptr %[[DIV_B_AB_ADDR]], i32 0, i32 0
// OGCG: %[[RESULT_REAL:.*]] = load float, ptr %[[RESULT_REAL_PTR]], align 4
// OGCG: %[[RESULT_IMAG_PTR:.*]] = getelementptr inbounds nuw { float, float }, ptr %[[DIV_B_AB_ADDR]], i32 0, i32 1
// OGCG: %[[RESULT_IMAG:.*]] = load float, ptr %[[RESULT_IMAG_PTR]], align 4
// OGCG: %[[RESULT_REAL_F16:.*]] = fptrunc float %[[RESULT_REAL]] to half
// OGCG: %[[RESULT_IMAG_F16:.*]] = fptrunc float %[[RESULT_IMAG]] to half
// OGCG: %[[B_REAL_PTR:.*]] = getelementptr inbounds nuw { half, half }, ptr %[[B_ADDR]], i32 0, i32 0
// OGCG: %[[B_IMAG_PTR:.*]] = getelementptr inbounds nuw { half, half }, ptr %[[B_ADDR]], i32 0, i32 1
// OGCG: store half %[[RESULT_REAL_F16]], ptr %[[B_REAL_PTR]], align 2
// OGCG: store half %[[RESULT_IMAG_F16]], ptr %[[B_IMAG_PTR]], align 2

#ifndef __cplusplus
void foo9() {
float _Complex a;
Expand Down Expand Up @@ -738,4 +857,4 @@ void foo9() {
// C_OGCG: %[[A_IMAG:.*]] = load float, ptr %[[A_IMAG_PTR]], align 4
// C_OGCG: %[[TMP_B:.*]] = load float, ptr %[[B_ADDR]], align 4
// C_OGCG: %[[ADD_REAL:.*]] = fadd float %[[TMP_B]], %[[A_REAL]]
// C_OGCG: store float %[[ADD_REAL]], ptr %[[B_ADDR]], align 4
// C_OGCG: store float %[[ADD_REAL]], ptr %[[B_ADDR]], align 4