> This is a follow up of the following patch:
> which fixes an ICE bug in adddi3. A similar problem exists in subdi3
> but Richard's original patch does not fix that. The patch below
> addresses the remaining issue. it has been tested on
> arm-unknown-linux-gnueabi with no new regressions.
By visual inspection of the code there appears to be a similar problem
in negdi2, see the patch below. Unfortunately my attempts to construct
a test case triggering that pattern failed.
--- gcc-4.5-20090723/gcc/config/arm/arm.md.~1~ 2009-07-23 17:02:20.000000000
+++ gcc-4.5-20090723/gcc/config/arm/arm.md 2009-07-24 23:17:40.000000000
@@ -3507,7 +3507,7 @@ (define_expand "negdi2"
if (GET_CODE (operands) != REG)
- operands = force_reg (SImode, operands);
+ operands = force_reg (DImode, operands);