This split is disabled for power7 and up, so we don't often see its bad effects. However, on a powerpc-linux compiler (which defaults to PPC750 judging from rs6000/sysv4.h) we see
long double ld1 (void) { return 1.0L; } compiled with -msoft-float -O2 -S resulting in ld1: li 3,0 li 4,0 mr 6,4 mr 5,3 li 4,0 lis 3,0x3ff0 blr Things go awry in init-regs, with the TFmode reg being initialized to zero on seeing the subreg from the split. (And that initialization itself is split by rs6000_emit_move!) Later passes apparently don't clean up the rubbish. Since the split was added for Darwin (as the comment says), let's get rid of it on other targets. Bootstrapped etc. powerpc64le-linux and powerpc64-linux. * config/rs6000/rs6000.c (rs6000_emit_move): Disable long double split for targets other than Darwin. diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c index 212e9facb3a..c339368be29 100644 --- a/gcc/config/rs6000/rs6000.c +++ b/gcc/config/rs6000/rs6000.c @@ -10010,8 +10010,8 @@ rs6000_emit_move (rtx dest, rtx source, machine_mode mode) /* 128-bit constant floating-point values on Darwin should really be loaded as two parts. However, this premature splitting is a problem when DFmode values can go into Altivec registers. */ - if (FLOAT128_IBM_P (mode) && !reg_addr[DFmode].scalar_in_vmx_p - && GET_CODE (operands[1]) == CONST_DOUBLE) + if (TARGET_MACHO && CONST_DOUBLE_P (operands[1]) && FLOAT128_IBM_P (mode) + && !reg_addr[DFmode].scalar_in_vmx_p) { rs6000_emit_move (simplify_gen_subreg (DFmode, operands[0], mode, 0), simplify_gen_subreg (DFmode, operands[1], mode, 0), -- Alan Modra Australia Development Lab, IBM