[RFC PATCH, i386]: Pass FP arguments in i387 registers

["Uros Bizjak" <ubizjak at gmail dot com>]

Hello!

This patch implements i387 register passing ABI extension for FP
function arguments:
a) unconditionally passes XFmode arguments of local functions in i387 registers
b) passes DFmode and SFmode arguments of local functions in i387
registers for -ffast-math

c) for -m387regparm, passes XFmode arguments in i387 registers
d) for -m387regparm, passes DFmode and SFmode arguments in 387
registers if -ffast-math is selected

This patch also cares for -msseregparm setting and doesn't step on its toes.

Following is a testcase, that _illustrates_ working of this patch:

--cut here--
static double __attribute__((noinline)) test(double a, double b)
{
 return a / b;
}

double foo(double a, double b)
{
 return test(a, b);
}
--cut here--

gcc -O2 -fomit-frame-pointer -ffast-math:
(loads FP arguments passed through memory to global function "foo"
into x87 and passes them to local func "test"):
test:
       fdivp   %st, %st(1)
       ret

foo:
       fldl    4(%esp)
       fldl    12(%esp)
       fxch    %st(1)
       jmp     test

gcc -O2 -fomit-frame-pointer -ffast-math -m387regparm:
test:
       fdivp   %st, %st(1)
       ret
foo:
       jmp     test

gcc -O2 -fomit-frame-pointer -ffast-math -mfpmath=sse -msse2:
test:
       divsd   %xmm1, %xmm0
       ret
foo:
       subl    $12, %esp
       movsd   24(%esp), %xmm1
       movsd   16(%esp), %xmm0
       call    test
       movsd   %xmm0, (%esp)
       fldl    (%esp)
       addl    $12, %esp
       ret

gcc -O2 -fomit-frame-pointer -ffast-math -mfpmath=sse -msse2 -m387regparm:
test:
       divsd   %xmm1, %xmm0
       ret
foo:
       subl    $12, %esp
       fstpl   (%esp)
       movsd   (%esp), %xmm1
       fstpl   (%esp)
       movsd   (%esp), %xmm0
       call    test
       movsd   %xmm0, (%esp)
       fldl    (%esp)
       addl    $12, %esp
       ret

etc, etc, etc... It works even for the silliest cases.

The patch was bootstrapped on i386-pc-linux-gnu and regression tested
on c,c++ and fortran. Povray-3.6.1 was also compiled and its test run
produced correct results. FWIW, the patch removed 96 "fld"
instructions and 385 "fst" instructions from calls to local functions
without any ABI violation.

Actually, this patch is fully functional, the only FIXME deals with
error detection and subsequent edge stack correction if "dead"
register is to be used (I never hit this problem). What remains as a
WIP is just a couple of lines of documentation and a bunch of
testcases.

YMMV, but this patch is expected to bring ~1% speedup on 32bit x87 FP
intensive code.

2005-10-12  Uros Bizjak  <uros@kss-loka.si>

       * config/i386/i386.opt: New target option -m387regparm.

       * config/i386/i386.h (struct ix86_args): Add x87_nregs, x87_regno,
       float_in_x87: Add new variables. mmx_words, sse_words: Remove.
       (X87_REGPARM_MAX): Define.

       * config/i386/i386.c (override_options): Error out for
       -m387regparm but no 80387 support.
       (ix86_attribute_table): Add 387regparm.
       (ix86_handle_cconv_attribute): Update comments for 387regparm.
       (ix86_comp_type_attributes): Check for mismatched 387regparm types.
       (ix86_function_387regparm): New function.
       (ix86_function_arg_regno_p): Add X87_REGPARM_MAX 80387 floating
       point registers.
       (init_cumulative_args): Initialize x87_nregs and float_in_x87
       variables.
       (function_arg_advance): Process x87_nregs and x87_regno when
       floating point argument is to be passed in 80387 register.
       (function_arg): Pass SFmode and DFmode arguments in 80387
       registers when cum->float_in_x87 and flag_unsafe_math_optimizations
       are set. Pass XFmode arguments in 80387 registers when
       cum->float_in_x87 is set.

       * reg-stack.c (convert_regs_entry): =FIXME= Disable NaN load when
       function argument is passed through 80387 register.

Uros.

Attachment: i386-387regparm.c
Description: Text document