Adam Langley wrote:
> * x87 doubles are 80-bits in registers and 64-bits in memory.
Depending on the state of the x87 floating point control word.
Can bracket significant test-impacting floating point operations with
fldcw or do something else in that code to force spills to memory?
I'm aware of the performance implication of mucking with the floating
point control register on older CPUs. I'm also aware that some GNU
libm functions assume that the x87 is set to 64/80-bit precision.
Maybe we care, maybe we don't. Maybe we can just set the register to
53/64-bit precision on renderer threads and leave it that way.
If we can't require SSE2 as a baseline on Linux, and I'd agree that we
can't, then it seems that we should find a way to do the right things
within the confines of the x87 instruction set.
Is the issue simply "test results are different due to rounding," or
are there actually cases where a spec requires that we do
floating-point operations in a certain mode? (I know that the JS and
XSL or XPath specs fall into the latter bucket.)
Mark
--~--~---------~--~----~------------~-------~--~----~
Chromium Developers mailing list: chromium-dev@googlegroups.com
View archives, change email options, or unsubscribe:
http://groups.google.com/group/chromium-dev
-~----------~----~----~----~------~----~------~--~---
