FMOD and FREM are broken in 80 bit non-Jit

PeterK · 21 June 2018, 12:44

Sorry for finding this bug today after the release of 4.0.0, but FMOD and FREM which correspond with FPREM and FPREM1 on the x87 FPU are broken in the 80 bit non-Jit mode only, but not in 64 bit or Jit or softfloat. I wished I did that test 1 or 2 days ago.

-----------------

Btw, on my low-end Mini-PC with just 2*1MB cache for 4 CPU cores WinUAE sometimes gets relative sluggish when the Jit cache is set to more than 8 MB. This is not new since 4.0.0 and not a problem for me at all. I just mention it for the case that other users report similar issues and can't find the reason.

Toni Wilen · 21 June 2018, 17:14

So whats the problem? I don't really care about those weird FPU instructions that no one uses anyway

PeterK · 21 June 2018, 20:01

I wrote a small program to test my Pow(x,y) function which has a double loop for x and y to check all integer operands from 0 to 31 with positive or negative values for the base. So it runs from 0^0 to 31^31 and prints out all results. For negative base values in Pow() I have to check whether the exponent in FP1 is even or odd to set the sign for (-x)^oddy. This is done with

Code:

     FREM.S #2,FP1  ; or FMOD.S
     FBEQ.W even
     BSET.L #31,D0
even RTS

To reproduce this problem it should be enough to test with a few integer values for FPx from 0 to 5 and check the result of FREM.S #2,FPx and the condition code.

This works with all configs except 80 bit non-Jit. I noticed it because it needs a few loops before the Jit starts and the wrong results suddenly changed into correct values.

Toni Wilen · 21 June 2018, 20:35

Parameters were swapped for x86 fprem and fprem1.. Fixed.

PeterK · 21 June 2018, 20:40

Is WinUAE.7z updated with that fix?

Ok, test passed.

Thanks.

Toni Wilen · 22 June 2018, 12:35

I thought I tested each fpu function with some simple tests but perhaps I used same value for both operands and assumed it worked if result was always zero

21 June 2018, 12:44	#1
PeterK Registered User Join Date: Apr 2005 Location: digital hell, Germany, after 1984, but worse Posts: 3,385	FMOD and FREM are broken in 80 bit non-Jit Sorry for finding this bug today after the release of 4.0.0, but FMOD and FREM which correspond with FPREM and FPREM1 on the x87 FPU are broken in the 80 bit non-Jit mode only, but not in 64 bit or Jit or softfloat. I wished I did that test 1 or 2 days ago. ----------------- Btw, on my low-end Mini-PC with just 2*1MB cache for 4 CPU cores WinUAE sometimes gets relative sluggish when the Jit cache is set to more than 8 MB. This is not new since 4.0.0 and not a problem for me at all. I just mention it for the case that other users report similar issues and can't find the reason.

21 June 2018, 20:40	#5
PeterK Registered User Join Date: Apr 2005 Location: digital hell, Germany, after 1984, but worse Posts: 3,385	Is WinUAE.7z updated with that fix? Ok, test passed. Thanks. Last edited by PeterK; 21 June 2018 at 21:13.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
64-bit WinUAE doesn't work on 64-bit Windows XP	mark_k	support.WinUAE	6	24 December 2017 21:17
64 bit winuae and JIT	pjhutch	support.WinUAE	1	25 September 2011 01:29
REQ: 17-Bit Artwork 2 (1988-04)(17-Bit Software)	Sea7	request.Demos	5	13 May 2011 01:07
8 bit to optimized 6 bit palette histogram improvements needed	NovaCoder	Coders. General	0	14 April 2011 02:13
Gui bug - JIT + 24 Bit (?)	rsn8887	support.WinUAE	3	29 October 2008 16:01

21 June 2018, 17:14	#2
Toni Wilen WinUAE developer Join Date: Aug 2001 Location: Hämeenlinna/Finland Age: 49 Posts: 26,571	So whats the problem? I don't really care about those weird FPU instructions that no one uses anyway

21 June 2018, 20:01	#3
PeterK Registered User Join Date: Apr 2005 Location: digital hell, Germany, after 1984, but worse Posts: 3,385	I wrote a small program to test my Pow(x,y) function which has a double loop for x and y to check all integer operands from 0 to 31 with positive or negative values for the base. So it runs from 0^0 to 31^31 and prints out all results. For negative base values in Pow() I have to check whether the exponent in FP1 is even or odd to set the sign for (-x)^oddy. This is done with Code: FREM.S #2,FP1 ; or FMOD.S FBEQ.W even BSET.L #31,D0 even RTS To reproduce this problem it should be enough to test with a few integer values for FPx from 0 to 5 and check the result of FREM.S #2,FPx and the condition code. This works with all configs except 80 bit non-Jit. I noticed it because it needs a few loops before the Jit starts and the wrong results suddenly changed into correct values.

21 June 2018, 20:35	#4
Toni Wilen WinUAE developer Join Date: Aug 2001 Location: Hämeenlinna/Finland Age: 49 Posts: 26,571	Parameters were swapped for x86 fprem and fprem1.. Fixed.

22 June 2018, 12:35	#6
Toni Wilen WinUAE developer Join Date: Aug 2001 Location: Hämeenlinna/Finland Age: 49 Posts: 26,571	I thought I tested each fpu function with some simple tests but perhaps I used same value for both operands and assumed it worked if result was always zero

Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)