Bug 17854 – Suboptimal code generated with constants and SSE

in both case with -O -release A. FP constant ============== int test(float i) { return cast(int)(2.0 * i * i); } disassembles to: 000000000045B640h push rbp 000000000045B641h mov rbp, rsp 000000000045B644h sub rsp, 10h 000000000045B648h movss xmm3, xmm0 000000000045B64Ch cvtss2sd xmm1, xmm0 000000000045B650h mov rax, 4000000000000000h 000000000045B65Ah mov qword ptr [rbp-10h], rax 000000000045B65Eh movsd xmm2, qword ptr [rbp-10h] 000000000045B663h mulsd xmm1, xmm2 000000000045B667h cvtss2sd xmm4, xmm0 000000000045B66Bh mulsd xmm1, xmm4 000000000045B66Fh cvttsd2si eax, xmm1 000000000045B673h mov rsp, rbp 000000000045B676h pop rbp 000000000045B677h ret B. Integer constant =================== int test(float i) { return cast(int)(2 * i * i); } disassembles to: 000000000045B640h push rbp 000000000045B641h mov rbp, rsp 000000000045B644h sub rsp, 10h 000000000045B648h movss xmm2, xmm0 000000000045B64Ch mov eax, 40000000h 000000000045B651h mov dword ptr [rbp-10h], eax 000000000045B654h movss xmm1, dword ptr [rbp-10h] 000000000045B659h mulss xmm0, xmm1 000000000045B65Dh mulss xmm0, xmm2 000000000045B661h cvttss2si eax, xmm0 000000000045B665h mov rsp, rbp 000000000045B668h pop rbp 000000000045B669h ret case A could clearly be compiled as B since the fractional part of the constant is empty.

The was the multiplication happens is terrible anyway. LDC generates cvtss2sd xmm0, xmm0 movapd xmm1, xmm0 addsd xmm1, xmm1 mulsd xmm1, xmm0 cvttsd2si eax, xmm1 ret

THIS ISSUE HAS BEEN MOVED TO GITHUB https://github.com/dlang/dmd/issues/19321 DO NOT COMMENT HERE ANYMORE, NOBODY WILL SEE IT, THIS ISSUE HAS BEEN MOVED TO GITHUB