patch breaks libgloss builds

patch breaks libgloss builds

[Eric Christopher]

Building crt0.S we fail on this instruction:

li t2,0xAAAA5555

with a mipsisa64-elf toolchain.

Interestingly enough the code that I think you added in reporting bad
numbers is failing too:

/dzur/sourceware/combined/libgloss/mips/crt0.S: Assembler messages:
/dzur/sourceware/combined/libgloss/mips/crt0.S:91: Error: Number
(0x0aaaa5555) larger than 32 bits

I'll take a look at it later, just noticed it yesterday.

-eric

Re: patch breaks libgloss builds

[Maciej W. Rozycki]

On Fri, 15 Apr 2005, Eric Christopher wrote:

> Building crt0.S we fail on this instruction:
> 
> li t2,0xAAAA5555
> 
> with a mipsisa64-elf toolchain.
> 
> Interestingly enough the code that I think you added in reporting bad
> numbers is failing too:
> 
> /dzur/sourceware/combined/libgloss/mips/crt0.S: Assembler messages:
> /dzur/sourceware/combined/libgloss/mips/crt0.S:91: Error: Number
> (0x0aaaa5555) larger than 32 bits
> 
> I'll take a look at it later, just noticed it yesterday.

 This fixes the problem and I've run the test suite for mips64el-linux-gnu 
successfully, but I don't feel that fulfils testing terribly well.  A 
proper test case for "li"/"dli" should be desirable -- I'll have a look at 
it later.

 Also it seems we call normalize_constant_expr() a little bit too 
extensively -- perhaps a few of these calls could be removed.  I think we 
have a few explicit sign-extensions elsewhere, too, and some of them could 
actually fit into one of these normalize_*() calls.

2005-04-15  Maciej W. Rozycki  <macro@linux-mips.org>

	* config/tc-mips.c (normalize_constant_expr): Don't check for
	HAVE_32BIT_GPRS.
	(check_absolute_expr): Only call normalize_constant_expr() if 
	HAVE_32BIT_GPRS.
	(mips_ip): Likewise.

	* config/tc-mips.c (check_absolute_expr): Fix formatting.

  Maciej

binutils-2.15.96-20050414-mips-gas-li.patch
diff -up --recursive --new-file binutils-2.15.96-20050414.macro/gas/config/tc-mips.c binutils-2.15.96-20050414/gas/config/tc-mips.c
--- binutils-2.15.96-20050414.macro/gas/config/tc-mips.c	2005-04-14 03:25:26.000000000 +0000
+++ binutils-2.15.96-20050414/gas/config/tc-mips.c	2005-04-15 19:25:08.000000000 +0000
@@ -3265,7 +3265,7 @@ static void
  */
 normalize_constant_expr (expressionS *ex)
 {
-  if ((ex->X_op == O_constant && HAVE_32BIT_GPRS)
+  if (ex->X_op == O_constant
       && IS_ZEXT_32BIT_NUM (ex->X_add_number))
     ex->X_add_number = (((ex->X_add_number & 0xffffffff) ^ 0x80000000)
 			- 0x80000000);
@@ -3425,9 +3425,11 @@ check_absolute_expr (struct mips_cl_insn
   if (ex->X_op == O_big)
     as_bad (_("unsupported large constant"));
   else if (ex->X_op != O_constant)
-    as_bad (_("Instruction %s requires absolute expression"), ip->insn_mo->name);
+    as_bad (_("Instruction %s requires absolute expression"),
+	    ip->insn_mo->name);
 
-  normalize_constant_expr (ex);
+  if (HAVE_32BIT_GPRS)
+    normalize_constant_expr (ex);
 }
 
 /* Count the leading zeroes by performing a binary chop. This is a
@@ -8102,7 +8104,8 @@ do_msbd:
 		  if (imm2_expr.X_op != O_big
 		      && imm2_expr.X_op != O_constant)
 		  insn_error = _("absolute expression required");
-		  normalize_constant_expr (&imm2_expr);
+		  if (HAVE_32BIT_GPRS)
+		    normalize_constant_expr (&imm2_expr);
 		  s = expr_end;
 		  continue;
 
@@ -8570,7 +8573,8 @@ do_msbd:
 	      if (imm_expr.X_op != O_big
 		  && imm_expr.X_op != O_constant)
 		insn_error = _("absolute expression required");
-	      normalize_constant_expr (&imm_expr);
+	      if (HAVE_32BIT_GPRS)
+		normalize_constant_expr (&imm_expr);
 	      s = expr_end;
 	      continue;
 

Re: patch breaks libgloss builds

[Eric Christopher]

>  This fixes the problem and I've run the test suite for mips64el-linux-gnu 
> successfully, but I don't feel that fulfils testing terribly well.  A 
> proper test case for "li"/"dli" should be desirable -- I'll have a look at 
> it later.
> 

Running a test of gcc with a combined tree for mipsisa64-elf ferrets out
a lot of problems since the unpredictable bits are set in the simulator
and libgloss uses the assembler macros for startup code.

>  Also it seems we call normalize_constant_expr() a little bit too 
> extensively -- perhaps a few of these calls could be removed.  I think we 
> have a few explicit sign-extensions elsewhere, too, and some of them could 
> actually fit into one of these normalize_*() calls.
> 
> 2005-04-15  Maciej W. Rozycki  <macro@linux-mips.org>
> 
> 	* config/tc-mips.c (normalize_constant_expr): Don't check for
> 	HAVE_32BIT_GPRS.
> 	(check_absolute_expr): Only call normalize_constant_expr() if 
> 	HAVE_32BIT_GPRS.
> 	(mips_ip): Likewise.
> 
> 	* config/tc-mips.c (check_absolute_expr): Fix formatting.

If you haven't checked it in yet OK to do so for mainline and 2.16. 

-eric