GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-07-13 00:21:49 +08:00
Code Issues Packages Projects Releases Wiki Activity

bug #1637: workaround register spilling in gebp with clang>=6.0+AVX+FMA

Browse Source 
(grafted from f233c6194db032f31c14d06b5e962197b66ea296
)
This commit is contained in:
Gael Guennebaud 2018-12-07 10:01:09 +01:00
parent 24d56f2e0e
commit c4ea9a916f
1 changed files with 7 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
Eigen/src/Core/arch/AVX/PacketMath.h
View File

@ -159,11 +159,12 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, co
#ifdef __FMA__ #ifdef __FMA__
template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) { template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) {
#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) ) #if ( (EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC<80) || (EIGEN_COMP_CLANG) )
// clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers, // Clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers,
// and gcc stupidly generates a vfmadd132ps instruction, // and even register spilling with clang>=6.0 (bug 1637).
// so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate // Gcc stupidly generates a vfmadd132ps instruction.
// the result of the product. // So let's enforce it to generate a vfmadd231ps instruction since the most common use
// case is to accumulate the result of the product.
Packet8f res = c; Packet8f res = c;
__asm__("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); __asm__("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
return res; return res;
@ -172,7 +173,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f&
#endif #endif
} }
template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) { template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) {
#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) ) #if ( (EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC<80) || (EIGEN_COMP_CLANG) )
// see above // see above
Packet4d res = c; Packet4d res = c;
__asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));