rs6000_hard_regno_nregs_internal allows SPE vectors in single registers satisfying SPE_SIMD_REGNO_P (i.e. register numbers 0 to 31). However, the corresponding test for e500 double treats all registers as being able to store a 64-bit value, rather than just those GPRs.
Logically this inconsistency is wrong; in addition, it causes problems unwinding from signal handlers. linux-unwind.h uses ARG_POINTER_REGNUM as a place to store the return address from a signal handler, but this logic in rs6000_hard_regno_nregs_internal results in that being considered an 8-byte register, resulting in assertion failures. (<https://gcc.gnu.org/ml/gcc-patches/2014-09/msg02625.html> first needs to be applied for unwinding to work in general on e500.) This patch makes rs6000_hard_regno_nregs_internal handle the e500 double case consistently with SPE vectors. Tested with no regressions with cross to powerpc-linux-gnuspe (given the aforementioned patch applied). Failures of signal handling unwinding tests such as gcc.dg/cleanup-{8,9,10,11}.c are fixed by this patch. OK to commit? 2014-10-24 Joseph Myers <[email protected]> * config/rs6000/rs6000.c (rs6000_hard_regno_nregs_internal): Do not allow e500 double in registers not satisyfing SPE_SIMD_REGNO_P. Index: gcc/config/rs6000/rs6000.c =================================================================== --- gcc/config/rs6000/rs6000.c (revision 216673) +++ gcc/config/rs6000/rs6000.c (working copy) @@ -1721,7 +1721,7 @@ rs6000_hard_regno_nregs_internal (int regno, enum SCmode so as to pass the value correctly in a pair of registers. */ else if (TARGET_E500_DOUBLE && FLOAT_MODE_P (mode) && mode != SCmode - && !DECIMAL_FLOAT_MODE_P (mode)) + && !DECIMAL_FLOAT_MODE_P (mode) && SPE_SIMD_REGNO_P (regno)) reg_size = UNITS_PER_FP_WORD; else -- Joseph S. Myers [email protected]
