[LoongArch64] Fix code generation for type conversion from float/double to uint/ulong. #105783
Conversation
…le to uint/ulong.
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
|
@shushanhf Could you please review this change？ |
|
If LA64 doesn't have native conversion instruction for fp<->unsigned, are the |
We just use some instructions embedded by the codegen to avoid calling helpers. |
|
Is there any JIT tests covering this? I can find |
|
shushanhf
left a comment
There was a problem hiding this comment.
LGTM
Thanks
@jakobbotsch could you please review this PR?
|
@jakobbotsch could you please review this PR? |
|
/azp run runtime |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/azp run runtime-dev-innerloop |
|
Azure Pipelines successfully started running 1 pipeline(s). |
The previous conversion from floating-point numbers to unsigned integers was incorrect in handling negative numbers and NaN.The previous instructions are as follows:
lu52i.d $r21, $zero, 1086movgr2fr.d $ft3, $r21fcmp.clt.d $fcc2, $fs0, $ft3fcmp.ceq.d $fcc3, $fs0, $ft3li.w $r21, 0x0bcnez $fcc2, 16fsub.d $ft3, $fs0, $ft3li.w $r21, 0x1slli.d $r21, $r21, 0x3ffsel $ft3, $ft3, $fs0, $fcc2ftintrz.l.d $ft3, $ft3movfr2gr.d $a0, $ft3bcnez $fcc3, 8beqz $a0, 8or $a0, $r21, $a0After modification, negative numbers and NaN are handled first.The modified instructions are as follows:
li.w $a0, 0x0movgr2fr.d $ft3, $zerofcmp.cult.d $fcc2, $fa0, $ft3bcnez $fcc2, 52lu52i.d $r21, $zero, 1086movgr2fr.d $ft3, $r21fcmp.clt.d $fcc2, $fa0, $ft3li.w $r21, 0x0bcnez $fcc2, 16fsub.d $ft3, $fa0, $ft3li.w $r21, 0x1slli.d $r21, $r21, 0x3ffsel $ft3, $ft3, $fa0, $fcc2ftintrz.l.d $ft3, $ft3movfr2gr.d $a0, $ft3or $a0, $r21, $a0