This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[PATCH, RFC] GCC 4.9, powerpc, allow TImode in VSX registers

When I did the initial power7 port, I punted on allowing TImode in the VSX
registers because I couldn't get it to work.  I am now revisiting it, and these
patches are my current effort, and I was wondering if people had comments on
them.  In terms of performance, there are two benchmarks in the Spec 2006 suite
that have minor regressions (perlbench and gamess), and 3 that have minor
improvements (hmmer, h264ref, and gromacs), so overall it looks like a wash.  I
do want to look the regressions, and see if there is something simple to

Some issues I ran into include:

I needed to set CANNOT_CHANGE_MODE so that TImode won't overlap with smaller
data types, due to the scalar portion of the register being in the upper
64-bits of the VSX register.

I limited the available address formats for TImode to be REG+REG needed for VSX

I discovered that setjmp/longjmp and exception handling needed to create TImode
values with the STACK_SAVEAREA_MODE macro.  However, the implementation of this
needs REG+OFFSET addressing.  So, I added a new type PTImode, which is only
used for STACK_SAVEAREA_MODE, and PTImode is limited to the GPRs.

If I enable logical operations in TImode mode (and, xor, etc.), the compiler
will convert DImode logical operations to TImode for 32-bit programs.  In the
future, I think I will tune this and/or provide insn splitters for DImode
logical operations.  For now, I just disallow logical operations on TImode if

I added a debug switch (-mvsx-timode) to disable putting TImode into VSX

2013-01-31  Michael Meissner  <>

	* config/rs6000/ (mul<mode>3): Use the combined macro
	VECTOR_UNIT_ALTIVEC_OR_VSX_P instead of separate calls to
	(vcond<mode><mode>): Likewise.
	(vcondu<mode><mode>): Likewise.
	(vector_gtu<mode>): Likewise.
	(vector_gte<mode>): Likewise.
	(xor<mode>3): Don't allow logical operations on TImode in 32-bit
	to prevent the compiler from converting DImode operations to
	(ior<mode>3): Likewise.
	(and<mode>3): Likewise.
	(one_cmpl<mode>2): Likewise.
	(nor<mode>3): Likewise.
	(andc<mode>3): Likewise.

	* config/rs6000/ (wt constraint): New constraint
	that returns VSX_REGS if TImode is allowed in VSX registers.

	* config/rs6000/ (easy_fp_constant): 0.0f is an easy
	constant under VSX.

	* config/rs6000/rs6000-modes.def (PTImode): Define, PTImode is
	similar to TImode, but it is restricted to being in the GPRs.

	* config/rs6000/rs6000.opt (-mvsx-timode): New switch to allow
	TImode to occupy a single VSX register.

	* config/rs6000/rs6000-cpus.def (ISA_2_6_MASKS_SERVER): Default to
	-mvsx-timode for power7/power8.
	(power7 cpu): Likewise.
	(power8 cpu): Likewise.

	* config/rs6000/rs6000.c (rs6000_hard_regno_nregs_internal): Make
	sure that TFmode/TDmode take up two registers if they are ever
	allowed in the upper VSX registers.
	(rs6000_hard_regno_mode_ok): If -mvsx-timode, allow TImode in VSX
	(rs6000_init_hard_regno_mode_ok): Likewise.
	(rs6000_debug_reg_global): Add debugging for PTImode and wt
	constraint.  Print if LRA is turned on.
	(rs6000_option_override_internal): Give an error if -mvsx-timode
	and VSX is not enabled.
	(invalid_e500_subreg): Handle PTImode, restricting it to GPRs.  If
	-mvsx-timode, restrict TImode to reg+reg addressing, and PTImode
	to reg+offset addressing.  Use PTImode when checking offset
	addresses for validity.
	(reg_offset_addressing_ok_p): Likewise.
	(rs6000_legitimate_offset_address_p): Likewise.
	(rs6000_legitimize_address): Likewise.
	(rs6000_legitimize_reload_address): Likewise.
	(rs6000_legitimate_address_p): Likewise.
	(rs6000_eliminate_indexed_memrefs): Likewise.
	(rs6000_emit_move): Likewise.
	(rs6000_secondary_reload): Likewise.
	(rs6000_secondary_reload_inner): Handle PTImode.  Allow 64-bit
	reloads to fpr registers to continue to use reg+offset addressing,
	but 64-bit reloads to altivec registers need reg+reg addressing.
	Drop test for PRE_MODIFY, since VSX loads/stores no longer support
	it.  Treat LO_SUM like a PLUS operation.
	(rs6000_secondary_reload_class): If type is 64-bit, prefer to use
	FLOAT_REGS instead of VSX_RGS to allow use of reg+offset
	(rs6000_cannot_change_mode_class): Do not allow TImode in VSX
	registers to share a register with a smaller sized type, since VSX
	puts scalars in the upper 64-bits.
	(print_operand): Add support for PTImode.
	(rs6000_register_move_cost): Use VECTOR_MEM_VSX_P instead of
	VECTOR_UNIT_VSX_P to catch types that can be loaded in VSX
	registers, but don't have arithmetic support.
	(rs6000_memory_move_cost): Add test for VSX.
	(rs6000_opt_masks): Add -mvsx-timode.

	* config/rs6000/ (VSm): Change to use 64-bit aligned moves
	for TImode.
	(VSs): Likewise.
	(VSr): Use wt constraint for TImode.
	(VSv): Drop TImode support.
	(vsx_movti): Delete, replace with versions for 32-bit and 64-bit.
	(vsx_movti_64bit): Likewise.
	(vsx_movti_32bit): Likewise.
	(vec_store_<mode>): Use VSX iterator instead of vector iterator.
	(vsx_and<mode>3): Delete use of '?' constraint on inputs, just put
	one '?' on the appropriate output constraint.  Do not allow TImode
	logical operations on 32-bit systems.
	(vsx_ior<mode>3): Likewise.
	(vsx_xor<mode>3): Likewise.
	(vsx_one_cmpl<mode>2): Likewise.
	(vsx_nor<mode>3): Likewise.
	(vsx_andc<mode>3): Likewise.
	(vsx_concat_<mode>): Likewise.
	(vsx_xxpermdi_<mode>): Fix thinko for non V2DF/V2DI modes.

	* config/rs6000/rs6000.h (MASK_VSX_TIMODE): Map from
	(enum rs6000_reg_class_enum): Add RS6000_CONSTRAINT_wt.
	(STACK_SAVEAREA_MODE): Use PTImode instead of TImode.

	* config/rs6000/ (INT mode attribute): Add PTImode.
	(TI2 iterator): New iterator for TImode, PTImode.
	(wd mode attribute): Add values for vector types.
	(movti_string): Replace TI move operations with operations for
	TImode and PTImode.  Add support for TImode being allowed in VSX
	(mov<mode>_string, TImode/PTImode): Likewise.
	(movti_ppc64): Likewise.
	(mov<mode>_ppc64, TImode/PTImode): Likewise.
	(TI mode splitters): Likewise.

	* doc/md.texi (PowerPC and IBM RS6000 constraints): Document wt

Michael Meissner, IBM
5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA	fax +1 (978) 399-6899

Attachment: gcc-power7.patch390b
Description: Text document

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]