Performance: JIT is emitting multiple conversion instructions when using float-math

[hopperpl]

    public float GetComp5(uint color) {
      return color * (255 / 31f);
    }

converted to

C.GetComp5(UInt32)
    L0000: vzeroupper
    L0003: mov eax, edx
    L0005: vxorps xmm0, xmm0, xmm0
    L0009: vcvtsi2sd xmm0, xmm0, rax                 ; <-- 2x conversion, should use cvtsi2ss
    L000e: vcvtsd2ss xmm0, xmm0, xmm0
    L0012: vmulss xmm0, xmm0, [0x7ff9f52004a0]
    L001a: ret

there shouldn't be two conversions, integer directly to single-precision is possible

---

for reference, clang does this (in C++):

GetComp5(unsigned int):
        mov       edi, edi
        pxor      xmm0, xmm0
        cvtsi2ss  xmm0, rdi
        mulss     xmm0, DWORD PTR .LC2[rip]
        ret

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

    public float GetComp5(uint color) {
      return color * (255 / 31f);
    }

converted to

C.GetComp5(UInt32)
    L0000: vzeroupper
    L0003: mov eax, edx
    L0005: vxorps xmm0, xmm0, xmm0
    L0009: vcvtsi2sd xmm0, xmm0, rax                 ; <-- 2x conversion, should use cvtsi2ss
    L000e: vcvtsd2ss xmm0, xmm0, xmm0
    L0012: vmulss xmm0, xmm0, [0x7ff9f52004a0]
    L001a: ret

there shouldn't be two conversions, integer directly to single-precision is possible

---

for reference, clang does this (in C++):

GetComp5(unsigned int):
        mov       edi, edi
        pxor      xmm0, xmm0
        cvtsi2ss  xmm0, rdi
        mulss     xmm0, DWORD PTR .LC2[rip]
        ret

Author:	hopperpl
Assignees:	-
Labels:	tenet-performance, area-CodeGen-coreclr
Milestone:	-

[EgorBo]

Should be pretty straightforward to fix, marking as help-wanted

[tannergooding]

Noting this can only be done on 64-bit. On 32-bit the required REX.W bit to force a 64-bit input isn't available.

[hopperpl]

Beware, that code is not the same. The multiplication factor has to be 255.0 / 31.0 = 8.2258 and not truncated down to 8.

You have to change the code to return color * (255 / (float) 31); and make sure there is no temporary double math. The term in parenthesis avoids an expensive floating division.