[clang] [CIR] Lowering for __builtin_reduce_assoc_fadd (PR #226095)
Kunal Dubey via cfe-commits
cfe-commits at lists.llvm.org
Wed Sep 30 23:15:47 PDT 2026
https://github.com/xakep8 updated https://github.com/llvm/llvm-project/pull/226095
>From 9dbe4008c5592d3371e4948fe27ccccc6a0c2044 Mon Sep 17 00:00:00 2001
From: Kunal Dubey <xakep8 at protonmail.com>
Date: Thu, 24 Sep 2026 15:08:58 +0530
Subject: [PATCH] [CIR] Lowering for __builtin_reduce_assoc_fadd
Added lowering for __builtin_reduce_assoc_fadd with the use of the new
implementation of CIR FastMathFlags, following same lowering path as
Classic Codegen.
Added tests for the same.
---
clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp | 39 ++++++++++++++-----
.../builtin-reduce-arithmetic-sve.c | 22 +++++++++++
.../builtin-reduce-arithmetic.c | 34 ++++++++++++++++
3 files changed, 85 insertions(+), 10 deletions(-)
diff --git a/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp b/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
index 245708691b7d9..3a37966854380 100644
--- a/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenBuiltin.cpp
@@ -2240,22 +2240,41 @@ RValue CIRGenFunction::emitBuiltinExpr(const GlobalDecl &gd, unsigned builtinID,
cast<cir::VectorType>(convertType(e->getArg(0)->getType()))
.getElementType());
case Builtin::BI__builtin_reduce_assoc_fadd:
- return errorBuiltinNYI(*this, e, builtinID);
case Builtin::BI__builtin_reduce_in_order_fadd: {
- assert(e->getNumArgs() == 2 &&
- "__builtin_reduce_in_order_fadd requires a start value");
+ bool isAssociative =
+ builtinIDIfNoAsmLabel == Builtin::BI__builtin_reduce_assoc_fadd;
+
+ assert((isAssociative ? e->getNumArgs() == 1 || e->getNumArgs() == 2
+ : e->getNumArgs() == 2) &&
+ "invalid argument count for floating-point reduction");
mlir::Value vector = emitScalarExpr(e->getArg(0));
auto vectorTy = cast<cir::VectorType>(vector.getType());
mlir::Type scalarTy = vectorTy.getElementType();
mlir::Location loc = getLoc(e->getExprLoc());
- mlir::Value startValue = emitScalarExpr(e->getArg(1));
- if (startValue.getType() != scalarTy)
- startValue =
- builder.createCast(getLoc(e->getArg(1)->getExprLoc()),
- cir::CastKind::floating, startValue, scalarTy);
+ mlir::Value startValue;
+ if (e->getNumArgs() == 2) {
+ startValue = emitScalarExpr(e->getArg(1));
+ if (startValue.getType() != scalarTy)
+ startValue =
+ builder.createCast(getLoc(e->getArg(1)->getExprLoc()),
+ cir::CastKind::floating, startValue, scalarTy);
+ } else {
+ auto fpTy = cast<cir::FPTypeInterface>(scalarTy);
+ startValue = cir::ConstantOp::create(
+ builder, loc,
+ cir::FPAttr::get(scalarTy,
+ llvm::APFloat::getZero(fpTy.getFloatSemantics(),
+ /*Negative=*/true)));
+ }
+
SmallVector<mlir::Value, 2> args = {startValue, vector};
- mlir::Value result =
- builder.emitIntrinsicCallOp(loc, "vector.reduce.fadd", scalarTy, args);
+ cir::FastMathFlagsAttr fastMath;
+ if (isAssociative)
+ fastMath = cir::FastMathFlagsAttr::get(&getMLIRContext(),
+ cir::FastMathFlags::reassoc);
+
+ mlir::Value result = builder.emitIntrinsicCallOp(loc, "vector.reduce.fadd",
+ scalarTy, fastMath, args);
return RValue::get(result);
}
case Builtin::BI__builtin_reduce_maximum:
diff --git a/clang/test/CIR/CodeGenBuiltins/builtin-reduce-arithmetic-sve.c b/clang/test/CIR/CodeGenBuiltins/builtin-reduce-arithmetic-sve.c
index fbd9752b255c3..0bb3c00edf6e9 100644
--- a/clang/test/CIR/CodeGenBuiltins/builtin-reduce-arithmetic-sve.c
+++ b/clang/test/CIR/CodeGenBuiltins/builtin-reduce-arithmetic-sve.c
@@ -92,9 +92,31 @@ float test_sve_reduce_min_float(svfloat32_t x) {
return __builtin_reduce_min(x);
}
+float test_sve_reduce_assoc_fadd(svfloat32_t x, float start) {
+ // CIR-LABEL: @test_sve_reduce_assoc_fadd
+ // CIR: cir.call_llvm_intrinsic "vector.reduce.fadd" {{.*}} : (!cir.float, !cir.vector<[4] x !cir.float>) -> !cir.float {fastmath_flags = #cir.fastmath<reassoc>}
+ // CIR: cir.return
+ // LLVM-LABEL: @test_sve_reduce_assoc_fadd
+ // LLVM: call reassoc float @llvm.vector.reduce.fadd.nxv4f32(float %{{.*}}, <vscale x 4 x float>
+ // LLVM: ret float
+ return __builtin_reduce_assoc_fadd(x, start);
+}
+
+float test_sve_reduce_assoc_fadd_default_start(svfloat32_t x) {
+ // CIR-LABEL: @test_sve_reduce_assoc_fadd_default_start
+ // CIR: %[[START:.*]] = cir.const #cir.fp<-0.000000e+00> : !cir.float
+ // CIR: cir.call_llvm_intrinsic "vector.reduce.fadd" %[[START]], {{.*}} : (!cir.float, !cir.vector<[4] x !cir.float>) -> !cir.float {fastmath_flags = #cir.fastmath<reassoc>}
+ // CIR: cir.return
+ // LLVM-LABEL: @test_sve_reduce_assoc_fadd_default_start
+ // LLVM: call reassoc float @llvm.vector.reduce.fadd.nxv4f32(float -0.000000e+00, <vscale x 4 x float>
+ // LLVM: ret float
+ return __builtin_reduce_assoc_fadd(x);
+}
+
float test_sve_reduce_in_order_fadd(svfloat32_t x, float start) {
// CIR-LABEL: @test_sve_reduce_in_order_fadd
// CIR: cir.call_llvm_intrinsic "vector.reduce.fadd" {{.*}} : (!cir.float, !cir.vector<[4] x !cir.float>) -> !cir.float
+ // CIR-NOT: fastmath_flags
// CIR: cir.return
// LLVM-LABEL: @test_sve_reduce_in_order_fadd
// LLVM: call float @llvm.vector.reduce.fadd.nxv4f32(float %{{.*}}, <vscale x 4 x float>
diff --git a/clang/test/CIR/CodeGenBuiltins/builtin-reduce-arithmetic.c b/clang/test/CIR/CodeGenBuiltins/builtin-reduce-arithmetic.c
index a2a4624de1854..2bb05b6001bc0 100644
--- a/clang/test/CIR/CodeGenBuiltins/builtin-reduce-arithmetic.c
+++ b/clang/test/CIR/CodeGenBuiltins/builtin-reduce-arithmetic.c
@@ -110,9 +110,43 @@ float test_reduce_min_float(v4sf x) {
return __builtin_reduce_min(x);
}
+float test_reduce_assoc_fadd(v4sf x, float start) {
+ // CIR-LABEL: @test_reduce_assoc_fadd
+ // CIR: cir.call_llvm_intrinsic "vector.reduce.fadd" {{.*}} : (!cir.float, !cir.vector<4 x !cir.float>) -> !cir.float {fastmath_flags = #cir.fastmath<reassoc>}
+ // CIR: cir.return
+ // LLVM-LABEL: @test_reduce_assoc_fadd
+ // LLVM: call reassoc float @llvm.vector.reduce.fadd.v4f32(float %{{.*}}, <4 x float>
+ // LLVM: ret float
+ return __builtin_reduce_assoc_fadd(x, start);
+}
+
+float test_reduce_assoc_fadd_default_start(v4sf x) {
+ // CIR-LABEL: @test_reduce_assoc_fadd_default_start
+ // CIR: %[[START:.*]] = cir.const #cir.fp<-0.000000e+00> : !cir.float
+ // CIR: cir.call_llvm_intrinsic "vector.reduce.fadd" %[[START]], {{.*}} : (!cir.float, !cir.vector<4 x !cir.float>) -> !cir.float {fastmath_flags = #cir.fastmath<reassoc>}
+ // CIR: cir.return
+ // LLVM-LABEL: @test_reduce_assoc_fadd_default_start
+ // LLVM: call reassoc float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float>
+ // LLVM: ret float
+ return __builtin_reduce_assoc_fadd(x);
+}
+
+float test_reduce_assoc_fadd_cast_start(v4sf x, double start) {
+ // CIR-LABEL: @test_reduce_assoc_fadd_cast_start
+ // CIR: %[[START:.*]] = cir.cast floating {{.*}} : !cir.double -> !cir.float
+ // CIR: cir.call_llvm_intrinsic "vector.reduce.fadd" %[[START]], {{.*}} : (!cir.float, !cir.vector<4 x !cir.float>) -> !cir.float {fastmath_flags = #cir.fastmath<reassoc>}
+ // CIR: cir.return
+ // LLVM-LABEL: @test_reduce_assoc_fadd_cast_start
+ // LLVM: %[[START:.*]] = fptrunc double %{{.*}} to float
+ // LLVM: call reassoc float @llvm.vector.reduce.fadd.v4f32(float %[[START]], <4 x float>
+ // LLVM: ret float
+ return __builtin_reduce_assoc_fadd(x, start);
+}
+
float test_reduce_in_order_fadd(v4sf x, float start) {
// CIR-LABEL: @test_reduce_in_order_fadd
// CIR: cir.call_llvm_intrinsic "vector.reduce.fadd" {{.*}} : (!cir.float, !cir.vector<4 x !cir.float>) -> !cir.float
+ // CIR-NOT: fastmath_flags
// CIR: cir.return
// LLVM-LABEL: @test_reduce_in_order_fadd
// LLVM: call float @llvm.vector.reduce.fadd.v4f32(float %{{.*}}, <4 x float>
More information about the cfe-commits
mailing list