When writing the software FMA code, I didn't realize that we can't
overwrite d if d is the same register as one of the inputs and
HandleNaNs is going to be called. This fixes that.
Typically when someone uses GetPointer, it's because they want to read
from a range of memory. GetPointer is unsafe to use for this. While it
does check that the passed-in address is valid, it doesn't know the size
of the range that will be accessed, so it can't check that the end
address is valid. The safer alternative GetPointerForRange should be
used instead.
Note that there is still the problem of many callers not checking for
nullptr.
This is part 2 of a series of changes removing the use of GetPointer
throughout the code base. After this, VideoCommon is the one major part
of Dolphin that remains.
RCOpArg::ExtractWithByteOffset is only used in one place: a special case
of rlwinmx. ExtractWithByteOffset first stores the value of the
specified register into m_ppc_state (unless it's already there), and
then returns an offset into m_ppc_state. Our use of this function has
two undesirable properties (except in the trivial case `offset == 0`):
1. ExtractWithByteOffset calls StoreFromRegister without going through
any of the usual functions. This violated an assumption I made when
working on my constant propagation PR and led to a hard-to-find bug.
2. If the specified register is in a host register and is dirty,
ExtractWithByteOffset will store its value to m_ppc_state even when
it's unnecessary. In particular, this always happens when rlwinmx
uses the same register as input and output, since rlwinmx always
allocates a host register for the output and marks it as dirty.
Since ExtractWithByteOffset is only used in one place, I figure we might
as well inline it there. This commit does that, and also alters
rlwinmx's logic so the special case code is only triggered when the
input is already in m_ppc_state.
Input in `m_ppc_state`, before (11 bytes):
mov esi, dword ptr [rbp-104]
mov dword ptr [rbp-104], esi
movzx esi, byte ptr [rbp-101]
Input in `m_ppc_state`, after (5 bytes):
movzx esi, byte ptr [rbp-101]
Input in host register, before (8 bytes):
mov dword ptr [rbp-104], esi
movzx esi, byte ptr [rbp-101]
Input in host register, after (3 bytes):
shr edi, 0x18
When the divisor is a constant value, we can emit more efficient code.
For powers of two, we can use bit shifts. For other values, we can
instead use a multiplication by magic constant method.
- Example 1 - Division by 16 (power of two)
Before:
mov w24, #0x10 ; =16
udiv w27, w25, w24
After:
lsr w27, w25, #4
- Example 2 - Division by 10 (fast)
Before:
mov w25, #0xa ; =10
udiv w27, w26, w25
After:
mov w27, #0xcccd ; =52429
movk w27, #0xcccc, lsl #16
umull x27, w26, w27
lsr x27, x27, #35
- Example 3 - Division by 127 (slow)
Before:
mov w26, #0x7f ; =127
udiv w27, w27, w26
After:
mov w26, #0x408 ; =1032
movk w26, #0x8102, lsl #16
umaddl x27, w27, w26, x26
lsr x27, x27, #38
Given we have fixed allocation orders, we can just directly return a
span instead of a pointer and a size via an out parameter.
Makes it a little more convenient, since we get both pieces of info at
once, and also have the ability to iterate directly off the span out of
the box.
