[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
target/ppc: bug in optimised vsl/vsr implementation?
From: |
Mark Cave-Ayland |
Subject: |
target/ppc: bug in optimised vsl/vsr implementation? |
Date: |
Thu, 26 Sep 2019 19:04:58 +0100 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101 Thunderbird/60.9.0 |
As part of the investigation into the DFP number issue reported at
https://bugs.launchpad.net/qemu/+bug/1841990 it appears that there may also be
a bug
introduced by the new optimised vsl/vsr implementation:
commit 4e6d0920e7547e6af4bbac5ffe9adfe6ea621822
Author: Stefan Brankovic <address@hidden>
Date: Mon Jul 15 16:22:48 2019 +0200
target/ppc: Optimize emulation of vsl and vsr instructions
Optimization of altivec instructions vsl and vsr(Vector Shift Left/Rigt).
Perform shift operation (left and right respectively) on 128 bit value of
register vA by value specified in bits 125-127 of register vB. Lowest 3
bits in each byte element of register vB must be identical or result is
undefined.
For vsl instruction, the first step is bits 125-127 of register vB have
to be saved in variable sh. Then, the highest sh bits of the lower
doubleword element of register vA are saved in variable shifted,
in order not to lose those bits when shift operation is performed on
the lower doubleword element of register vA, which is the next
step. After shifting the lower doubleword element shift operation
is performed on higher doubleword element of vA, with replacement of
the lowest sh bits(that are now 0) with bits saved in shifted.
For vsr instruction, firstly, the bits 125-127 of register vB have
to be saved in variable sh. Then, the lowest sh bits of the higher
doubleword element of register vA are saved in variable shifted,
in odred not to lose those bits when the shift operation is
performed on the higher doubleword element of register vA, which is
the next step. After shifting higher doubleword element, shift operation
is performed on lower doubleword element of vA, with replacement of
highest sh bits(that are now 0) with bits saved in shifted.
Signed-off-by: Stefan Brankovic <address@hidden>
Reviewed-by: Richard Henderson <address@hidden>
Message-Id: <address@hidden>
Signed-off-by: David Gibson <address@hidden>
Reverting 4e6d0920e7547e6af4bbac5ffe9adfe6ea621822 allows the test case in the
bug
report to pass once again. Stefan, are you able to take a look at this?
ATB,
Mark.
- target/ppc: bug in optimised vsl/vsr implementation?,
Mark Cave-Ayland <=