OSDN Git Service
ART: Prevent float inference back to SSA in-regs.
.method public static getInt(I)I
.registers 2
const/4 v0, 0x0
if-ne v0, v0, :after
float-to-int v0, v0
:exit
add-int/2addr v0, v1
return v0
:after
move v1, v0
goto :exit
.end method
In this code sample, v1 is the single parameter to this method. In one
of the phi-nodes inserted between :exit and add-int/2addr, v1's two
incoming SSA regs are:
- the initial def of v1 as a parameter
- the v1 def'd at move v1, v0.
During type inference, because the 2nd def is a float (because of the
earlier float-to-int v0, v0) this will change the type of the 1st def to a
float as well, which is incorrect since the first parameter is known to be
non-float.
This fix checks during phi-node type-inference if an SSA reg that is the
initial def of a parameter vreg is about to be set as float when it was
not previously, and skips the inference if so.
In this case, when using a hard-float ABI, having the in-reg v1 set as
float causes FlushIns() to read the argument to the method from an FP reg,
when the argument will be passed in a core reg by any caller.
Also included is a smali test for this bug: compare difference between
./run-test --64 800
./run-test --64 --interpreter 800
when the vreg_analysis patch has not been applied.
(Requires 64-bit because 32-bit ARM currently does not use hard-float.)
getInt(I)I should return its argument, but it returns an incorrect
value.
Change-Id: I1d4b5be6a931fe853279e89dd820820f29823da1
Signed-off-by: Stephen Kyle <stephen.kyle@arm.com>