handling trigonometric function under "-ffast-math"

[k-arrows]

Consider the following code under -ffast-math.

#include <math.h>

float f1(float x)
{
    return tanf(x) * cosf(x);
}

float f2(float x)
{
    return sinf(x) / tanf(x);
}

GCC

f1:
        b       sinf
f2:
        b       cosf

Clang

f1:                                     // @f1
        stp     d9, d8, [sp, #-32]!             // 16-byte Folded Spill
        stp     x29, x30, [sp, #16]             // 16-byte Folded Spill
        add     x29, sp, #16
        fmov    s8, s0
        bl      tanf
        fmov    s9, s0
        fmov    s0, s8
        bl      cosf
        fmul    s0, s9, s0
        ldp     x29, x30, [sp, #16]             // 16-byte Folded Reload
        ldp     d9, d8, [sp], #32               // 16-byte Folded Reload
        ret
f2:                                     // @f2
        stp     d9, d8, [sp, #-32]!             // 16-byte Folded Spill
        stp     x29, x30, [sp, #16]             // 16-byte Folded Spill
        add     x29, sp, #16
        fmov    s8, s0
        bl      sinf
        fmov    s9, s0
        fmov    s0, s8
        bl      tanf
        fdiv    s0, s9, s0
        ldp     x29, x30, [sp, #16]             // 16-byte Folded Reload
        ldp     d9, d8, [sp], #32               // 16-byte Folded Reload
        ret

https://godbolt.org/z/Kf9r1aPj7

This difference is not specific to AArch64.

[RKSimon]

We do a few trig combines already (see InstCombinerImpl::visitFDiv), but we're missing a Intrinsic::tan which should be dealt with first

[RKSimon]

See: Issue #34950

[k-arrows]

Thank you for your explanation. I understand that this issue is a duplicate.