2017-11-15 12:08:44

by Matt Redfearn

[permalink] [raw]
Subject: Re: [PATCH] MIPS: Fix exception entry when CONFIG_EVA enabled



On 13/11/17 10:47, Maciej W. Rozycki wrote:
> On Tue, 31 Oct 2017, James Hogan wrote:
>
>>> I looked this over pretty carefully and it looks correct to me.  It
>>> makes no difference
>>> in the instructions generated by the non-EVA case.  I shouldn't have
>>> missed this :(.
>>>
>>> Reviewed-by: Corey Minyard <[email protected]>
>>
>> Yeh, having stared at it for a little while it looks correct to me too.
>>
>> Reviewed-by: James Hogan <[email protected]>
>
> How about getting rid of the `noreorder' mode and the manually scheduled
> delay slot here altogether though, as I outlined?
>
> Maciej
>

Hi Maciej,

I like the change you propose, however I can't coax GAS to reorder the
instructions appropriately. With this patch on top of 4.14:

--- a/arch/mips/include/asm/stackframe.h
+++ b/arch/mips/include/asm/stackframe.h
@@ -195,14 +195,16 @@
.set push
.set noat
.set reorder
- mfc0 k0, CP0_STATUS
- sll k0, 3 /* extract cu0 bit */
- .set noreorder
- bltz k0, 8f
- move k0, sp
+ mfc0 k1, CP0_STATUS
+ sll k1, 3 /* extract cu0 bit */
+
+ move k0, sp
.if \docfi
.cfi_register sp, k0
.endif
+
+ bltz k1, 8f
+
#ifdef CONFIG_EVA
/*
* Flush interAptiv's Return Prediction Stack (RPS) by
writing
@@ -228,7 +230,6 @@
MFC0 k0, CP0_ENTRYHI
MTC0 k0, CP0_ENTRYHI
#endif
- .set reorder
/* Called from user mode, new stack. */
get_saved_sp docfi=\docfi tosp=1
8:


The generated assembly is:

80405d00 <handle_int>:
80405d00: 401b6000 mfc0 k1,c0_status
80405d04: 001bd8c0 sll k1,k1,0x3
80405d08: 03a0d025 move k0,sp
80405d0c: 07600007 bltz k1,80405d2c <handle_int+0x2c>
80405d10: 00000000 nop
80405d14: 401a2000 mfc0 k0,c0_context

Apparently GAS has not been able to reorder the move into the branch
delay slot for some reason. Any ideas?

Thanks,
Matt

From 1583947582734513723@xxx Mon Nov 13 10:48:48 +0000 2017
X-GM-THRID: 1580951206118155742
X-Gmail-Labels: Inbox,Category Forums,HistoricalUnread