On Thu, Jan 11, 2024 at 09:53:33AM +0100, Jakub Jelinek wrote:
> On Wed, Jan 10, 2024 at 07:05:39PM +0000, Andre Vieira (lists) wrote:
> > This patch is still work in progress, but posting to show failure with
> > bitint-7 test where handle_stmt called from lower_mergeable_stmt ICE's
> > because the idx (3) is out of range for the __BitInt(135) with a limb_prec
> > of 64.
>
> I can reproduce it, will debug it momentarily.
So, the problem was that in 2 spots I was comparing TYPE_SIZE of large/huge
BITINT_TYPEs to determine if it can be handled cheaply.
On x86_64 with limb_mode == abi_limb_mode (both DImode) that works fine,
if TYPE_SIZE is equal, it means it has the same number of limbs.
But on aarch64 TYPE_SIZE of say _BitInt(135) and _BitInt(193) is the same,
both are 256-bit storage, but because DImode is used as limb_mode, the
former actually needs just 3 limbs, while the latter needs 4 limbs.
And limb_access_type was asserting that we don't try to access 4th limb
on types which actually have a precision which needs just 3 limbs.
The following patch (so far tested on x86_64 with all the bitint tests plus
on the bitint-7.c testcase in a cross to aarch64) should fix that.
Note, for the info.extended targets (currently none, but I think arm 32-bit
in the ABI is meant like that), we'll need to do something different,
because the upper bits aren't just padding and should be zero/sign extended,
so if we say have limb_mode SImode, abi_limb_mode DImode, we'll need to
treat _BitInt(135) not as 5 SImode limbs, but 6. For !info.extended targets
I think treating _BitInt(135) as 3 DImode limbs rather than 4 is fine.
2024-01-11 Jakub Jelinek <ja...@redhat.com>
* gimple-lower-bitint.cc (mergeable_op): Instead of comparing
TYPE_SIZE (t) of large/huge BITINT_TYPEs, compare
CEIL (TYPE_PRECISION (t), limb_prec).
(bitint_large_huge::handle_cast): Likewise.
--- gcc/gimple-lower-bitint.cc.jj 2024-01-08 13:58:21.448176859 +0100
+++ gcc/gimple-lower-bitint.cc 2024-01-11 11:46:49.147779946 +0100
@@ -231,7 +231,8 @@ mergeable_op (gimple *stmt)
&& TREE_CODE (rhs_type) == BITINT_TYPE
&& bitint_precision_kind (lhs_type) >= bitint_prec_large
&& bitint_precision_kind (rhs_type) >= bitint_prec_large
- && tree_int_cst_equal (TYPE_SIZE (lhs_type), TYPE_SIZE (rhs_type)))
+ && (CEIL (TYPE_PRECISION (lhs_type), limb_prec)
+ == CEIL (TYPE_PRECISION (rhs_type), limb_prec)))
{
if (TYPE_PRECISION (rhs_type) >= TYPE_PRECISION (lhs_type))
return true;
@@ -1263,8 +1264,8 @@ bitint_large_huge::handle_cast (tree lhs
if m_upwards_2limb * limb_prec is equal to
lhs precision that is not the case. */
|| (!m_var_msb
- && tree_int_cst_equal (TYPE_SIZE (rhs_type),
- TYPE_SIZE (lhs_type))
+ && (CEIL (TYPE_PRECISION (lhs_type), limb_prec)
+ == CEIL (TYPE_PRECISION (rhs_type), limb_prec))
&& (!m_upwards_2limb
|| (m_upwards_2limb * limb_prec
< TYPE_PRECISION (lhs_type)))))
Jakub
