fcheck-1.1, ERC32 FPU rev.B ldd/std bug checker
-----------------------------------------------

1. Introduction

Fcheck checks for the failing FPU instruction sequences identified in
the ERC32 bug list ("Current errors in ERC32 CBA", Issue 1, 04-02-1997, MHS).
It will check for all three failing sequences as described in paragraphs
3.1, 3.2, and 3.3. The failing sequences are denoted case 1, case 2 and case 3
respectively.

The fcheck program reads binaries in either a.out, srecords or tektronix hex
format. The following syntax is used:

	fcheck [-c] [-v] [-sec sec_name] files ...

The program will scan the text part of the binary and report the address of
each failing sequence found. The -v switch will also dump a dis-assembly of
the failing sequence. As default, only sections named '.text' or '.sec1' are
checked, ocrresponding to the text sections in a.out and srec formats. The
-sec switch can be used to also check sections with other names. Since
case 3 errors only occurs when waitstates are used, they can be neglected
in most systems. If the '-c' switch is given to fcheck, only case 1 and
case 2 will be flagged.

The program will report found failing sequences as follows:

> fcheck test.o
ERROR 1 at 00000010, (lddf/stdf)
ERROR 2 at 00000018, (ldfsr/stfsr)
ERROR 3 at 00000000, (ldf/FPops)

The error type correspond to the error case as defined above.
Using the -v switch, the failing instructions are also provided:

> fcheck -v test.o
loading test.o:
checking section .text at 0x00000000 (160 bytes)
ERROR 1 at 00000010, (lddf/stdf)
ldd  [ %o0 ], %f30
std  %f28, [ %fp ]
ERROR 2 at 00000018, (ldfsr/stfsr)
ld  [ %o0 ], %fsr
st  %fsr, [ %fp ]
ERROR 3 at 00000000, (ldf/FPops)
ld  [ %o0 ], %f0
fdivs  %f0, %f2, %f4



The provided program runs on sparcstaions under SunOs or Solaris.

2. Bug description

2.1 Case 1

The error occurs in the following sequence:

	ldd	[...], %fd1
	std	%fd2, [%rs1 + ...]

If %fd1 is equal to %rs1, AND %fd1 is not equal to %fd2, THEN the std
instrution will store the wrong data. Examples:

This sequence will fail:

	ldd	[%o0], %f30
	std	%f28, [%fp]

because %f30 is equal to %fp (r30).

The two following sequences will NOT fail:

	ldd	[%o0], %f30
	std	%f30, [%fp]

	ldd	[%o0], %f30
	std	%f28, [%i0]

2.2 Case 2

The error occurs in the following sequence:

 1. Single-precision load followed by single-precision store or FPop

	ld 	[..], %fN
	FPops 	%fX, %fY, %fZ
    
	ld 	[..], %fN
	st 	%fX, [..]
    
    Error is flagged if N=X or N=Y. 

 2. Single-precision load followed by double-precision store or FPop

	ld 	[..], %fN
	FPopd 	%fX, %fY, %fZ
    
	ld 	[..], %fN
	std 	%fX, [..]
    
    Error is flagged if N=X or N=Y. Note that N is truncated to an even
    value since double-precision FPop always use even %f registers.


 3. Double-precision load followed by single-precision load or FPop

	ldd 	[..], %fN
	FPops 	%fX, %fY, %fZ

	ldd 	[..], %fN
	st 	%fX, [..]
    
    
    Error is flagged if N=X or N=Y. Note that X and Y are truncated to an even
    value since double-precision load always use even %f registers.

 3. Double-precision load followed by double-precision load or FPop

	ldd 	[..], %fN
	FPop 	%fX, %fY, %fZ

	ldd 	[..], %fN
	std 	%fX, [..]
    
    Error is flagged if N=X or N=Y.

2.3 Case 3

The error occurs in the following sequence:

	ld	[..], %fsr
	st	%fsr, [..]


Report errors to jgais@wd.estec.esa.nl

Good luck!!!

Jiri.