Hi all,

There was previously no backend pattern to utilise the scvtf fbits option. Where a fixed point is converted to a float, and divided by a power of 2, (or multiplied by the reciprocal of a power of 2), this can be combined into a single scvtf with fbits operation. This patch adds a pattern to combine these instructions, and adds a helper function.

For the following test case:

float f(int a) { return ((float) a) / 65536.0; }
double g(int a) { return ((double) a) / 4096.0; }

the output generated is currently:

f:
        scvtf   s1, w0  // 6    [c=8 l=4]  floatsisf2/1
        mov     w1, 931135488   // 17   [c=4 l=4]  *movsi_aarch64/3
        fmov    s0, w1  // 18   [c=4 l=4]  *movsf_aarch64/1
        fmul    s0, s1, s0      // 13   [c=8 l=4]  mulsf3
        ret             // 24   [c=0 l=4]  *do_return
g:
        scvtf   d1, w0  // 6    [c=8 l=4]  floatsidf2
        mov     x1, 4553139223271571456 // 17   [c=4 l=4] *movdi_aarch64/3
        fmov    d0, x1  // 18   [c=4 l=4]  *movdf_aarch64/1
        fmul    d0, d1, d0      // 13   [c=8 l=4]  muldf3
        ret             // 24   [c=0 l=4]  *do_return
        
The output with this patch applied is:

f:
        scvtf   s0, w0, #16     // 13   [c=24 l=4]  *combine_scvtf_SI_sf3/1
        ret             // 22   [c=0 l=4]  *do_return
g:
        scvtf   d0, w0, #12     // 13   [c=24 l=4]  *combine_scvtf_SI_df3
        ret             // 22   [c=0 l=4]  *do_return
        
gcc/ChangeLog:

2019-06-12  Joel Hutton  <Joel.Hutton@arm.com>

        * config/aarch64/aarch64-protos.h (aarch64_fpconst_pow2_recip): New prototype
        * config/aarch64/aarch64.c (aarch64_fpconst_pow2_recip): New function
        * config/aarch64/aarch64.md (*aarch64_<su_optab>cvtf_<fcvt_target>_<GPF:mode>2_mult): New pattern
        (aarch64_<su_optab>cvtf_<fcvt_iesize>_<GPF:mode>2_mult): New pattern
        * config/aarch64/constraints.md (Dt): New constraint
        * config/aarch64/predicates.md (aarch64_fpconst_pow2_recip): New predicate

gcc/testsuite/ChangeLog:

2019-06-12  Joel Hutton  <Joel.Hutton@arm.com>

        * gcc.target/aarch64/fmul_scvtf.c: New test.

Bootstrapped and regression tested on aarch64-linux-none target.