From c31ead8a15e7ddc91b4d02962164639f7ce9f4dc Mon Sep 17 00:00:00 2001
From: Chip-Kerchner <chip.kerchner@ibm.com>
Date: Wed, 24 Feb 2021 17:40:34 -0600
Subject: Having forward template function declarations in a P10 file causes
 bad code in certain situations.

---
 Eigen/src/Core/arch/AltiVec/MatrixProductCommon.h | 46 +++++++++++++++++++++++
 Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h    | 46 -----------------------
 2 files changed, 46 insertions(+), 46 deletions(-)

(limited to 'Eigen/src/Core/arch')

diff --git a/Eigen/src/Core/arch/AltiVec/MatrixProductCommon.h b/Eigen/src/Core/arch/AltiVec/MatrixProductCommon.h
index 87b60c22c..a1799c061 100644
--- a/Eigen/src/Core/arch/AltiVec/MatrixProductCommon.h
+++ b/Eigen/src/Core/arch/AltiVec/MatrixProductCommon.h
@@ -2,6 +2,52 @@ namespace Eigen {
 
 namespace internal {
 
+template<typename Scalar, typename Packet, typename DataMapper, typename Index, const Index accRows>
+EIGEN_STRONG_INLINE void gemm_extra_col(
+  const DataMapper& res,
+  const Scalar *lhs_base,
+  const Scalar *rhs_base,
+  Index depth,
+  Index strideA,
+  Index offsetA,
+  Index row,
+  Index col,
+  Index remaining_rows,
+  Index remaining_cols,
+  const Packet& pAlpha);
+
+template<typename Scalar, typename Packet, typename DataMapper, typename Index, const Index accRows>
+EIGEN_STRONG_INLINE void gemm_extra_row(
+  const DataMapper& res,
+  const Scalar *lhs_base,
+  const Scalar *rhs_base,
+  Index depth,
+  Index strideA,
+  Index offsetA,
+  Index row,
+  Index col,
+  Index cols,
+  Index remaining_rows,
+  const Packet& pAlpha,
+  const Packet& pMask);
+
+template<typename Scalar, typename Packet, typename DataMapper, typename Index, const Index accCols>
+EIGEN_STRONG_INLINE void gemm_unrolled_col(
+  const DataMapper& res,
+  const Scalar *lhs_base,
+  const Scalar *rhs_base,
+  Index depth,
+  Index strideA,
+  Index offsetA,
+  Index& row,
+  Index rows,
+  Index col,
+  Index remaining_cols,
+  const Packet& pAlpha);
+
+template<typename Packet>
+EIGEN_STRONG_INLINE Packet bmask(const int remaining_rows);
+
 const static Packet16uc p16uc_SETCOMPLEX32_FIRST = {  0,  1,  2,  3,
                                                      16, 17, 18, 19,
                                                       4,  5,  6,  7,
diff --git a/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h b/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h
index 888f7cd08..bfee9ee92 100644
--- a/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h
+++ b/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h
@@ -238,52 +238,6 @@ EIGEN_STRONG_INLINE void ploadRhsMMA<double, __vector_pair>(const double *rhs, _
   __builtin_vsx_assemble_pair(&rhsV, (__vector unsigned char)(*(((Packet2d *)rhs) + 1)), (__vector unsigned char)(*((Packet2d *)rhs)));
 }
 
-template<typename Scalar, typename Packet, typename DataMapper, typename Index, const Index accRows>
-EIGEN_STRONG_INLINE void gemm_extra_col(
-  const DataMapper& res,
-  const Scalar *lhs_base,
-  const Scalar *rhs_base,
-  Index depth,
-  Index strideA,
-  Index offsetA,
-  Index row,
-  Index col,
-  Index remaining_rows,
-  Index remaining_cols,
-  const Packet& pAlpha);
-
-template<typename Scalar, typename Packet, typename DataMapper, typename Index, const Index accRows>
-EIGEN_STRONG_INLINE void gemm_extra_row(
-  const DataMapper& res,
-  const Scalar *lhs_base,
-  const Scalar *rhs_base,
-  Index depth,
-  Index strideA,
-  Index offsetA,
-  Index row,
-  Index col,
-  Index cols,
-  Index remaining_rows,
-  const Packet& pAlpha,
-  const Packet& pMask);
-
-template<typename Scalar, typename Packet, typename DataMapper, typename Index, const Index accCols>
-EIGEN_STRONG_INLINE void gemm_unrolled_col(
-  const DataMapper& res,
-  const Scalar *lhs_base,
-  const Scalar *rhs_base,
-  Index depth,
-  Index strideA,
-  Index offsetA,
-  Index& row,
-  Index rows,
-  Index col,
-  Index remaining_cols,
-  const Packet& pAlpha);
-
-template<typename Packet>
-EIGEN_STRONG_INLINE Packet bmask(const int remaining_rows);
-
 #define MICRO_MMA_DST \
   __vector_quad *accZero0, __vector_quad *accZero1, __vector_quad *accZero2, \
   __vector_quad *accZero3, __vector_quad *accZero4, __vector_quad *accZero5, \
-- 
cgit v1.2.3