Re: [webkit-dev] jit, unordered compare
Hi fortunately, no, I don't need double branch here. cmpvs reg0, reg0 is a conditional instruction, which only executes if the v flag is set. My proposal would be to add a new DoubleCondition called DoubleEqualOrNAN, or something similar to clarify what we expect from the conditional branch. Since I feel (and by IEEE as you mentioned) a comparison to NaN should be false, not true as x86 actually does. Zoltan unfortunately cmf was used by the old fpa (floating point accelerator), which is replaced by vfp (vector floating point) for some time. Perhaps I can try a cmpvs reg0, reg0 instruction, which sets zero flag if one argument is NaN (since reg0 is always equal to reg0). Otherwise it does nothing (=nop). That should work, but it would require two branches instead of one (compare to 0, compare to self). I'm sure there must be a way to do an IEEE double comparison on ARM with only one branch. Geoff ___ webkit-dev mailing list webkit-dev@lists.webkit.org http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
Re: [webkit-dev] jit, unordered compare
Hi Zoltan.
I believe you're talking about this code for op_jfalse:
zeroDouble(fpRegT0);
emitLoadDouble(cond, fpRegT1);
addJump(branchDouble(DoubleEqual, fpRegT0, fpRegT1), target);
and this code for op_jtrue:
zeroDouble(fpRegT0);
emitLoadDouble(cond, fpRegT1);
addJump(branchDouble(DoubleNotEqual, fpRegT0, fpRegT1),
target);
The goal of this code is to perform a comparison that distinguishes
{ 0, NaN } from all other doubles. op_jfalse should branch in the case
of 0 or NaN, and, inversely, op_jtrue should branch in all other cases.
It sounds like the equal / zero condition does not indicate NaN on all
platforms. Bummer.
I'm not an expert in ARM (or other floating point processors, for that
matter), so I'm not sure what the best abstraction is for this notion
of compares to zero as equal or unordered. Since IEEE double
requires all comparisons to NaN to be implicitly false, I do expect
that there is some efficient single branch that can do this on ARM.
From http://www.heyrick.co.uk/assembler/fpops.html#cmf, it looks like
the CMF instruction should be used to test for equality (ie when a
BEQ or BNE is used afterwards) or to test for unorderedness (in the V
flag).
Gavin, do you have any ideas here?
Geoff
On Oct 12, 2009, at 4:05 AM, Zoltan Herczeg wrote:
Hi,
My stougle with USE_JSVALUE32_64 still continues on ARM. In
emit_op_jfalse, there is a comparison here (fpRegT0 contains 0.0):
addJump(branchDouble(DoubleEqual, fpRegT0, fpRegT1), target + 2);
In x86, if either operand is NaN, the zero flag is set by definition
(ucomisd instruction). I have no idea why (NaN == anything is true?).
Unfortunately, ARM does not set the zero flag. What would be a
portable
way to improve this instruction? (and emit_op_jtrue) Perhaps this
affects
the possible thumb2 implementation with USE_JSVALUE32_64 as well.
Zoltan
___
webkit-dev mailing list
webkit-dev@lists.webkit.org
http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
___
webkit-dev mailing list
webkit-dev@lists.webkit.org
http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
Re: [webkit-dev] jit, unordered compare
Hi,
unfortunately cmf was used by the old fpa (floating point accelerator),
which is replaced by vfp (vector floating point) for some time. Perhaps I
can try a cmpvs reg0, reg0 instruction, which sets zero flag if one
argument is NaN (since reg0 is always equal to reg0). Otherwise it does
nothing (=nop).
Zoltan
Hi Zoltan.
I believe you're talking about this code for op_jfalse:
zeroDouble(fpRegT0);
emitLoadDouble(cond, fpRegT1);
addJump(branchDouble(DoubleEqual, fpRegT0, fpRegT1), target);
and this code for op_jtrue:
zeroDouble(fpRegT0);
emitLoadDouble(cond, fpRegT1);
addJump(branchDouble(DoubleNotEqual, fpRegT0, fpRegT1),
target);
The goal of this code is to perform a comparison that distinguishes
{ 0, NaN } from all other doubles. op_jfalse should branch in the case
of 0 or NaN, and, inversely, op_jtrue should branch in all other cases.
It sounds like the equal / zero condition does not indicate NaN on all
platforms. Bummer.
I'm not an expert in ARM (or other floating point processors, for that
matter), so I'm not sure what the best abstraction is for this notion
of compares to zero as equal or unordered. Since IEEE double
requires all comparisons to NaN to be implicitly false, I do expect
that there is some efficient single branch that can do this on ARM.
From http://www.heyrick.co.uk/assembler/fpops.html#cmf, it looks like
the CMF instruction should be used to test for equality (ie when a
BEQ or BNE is used afterwards) or to test for unorderedness (in the V
flag).
Gavin, do you have any ideas here?
Geoff
On Oct 12, 2009, at 4:05 AM, Zoltan Herczeg wrote:
Hi,
My stougle with USE_JSVALUE32_64 still continues on ARM. In
emit_op_jfalse, there is a comparison here (fpRegT0 contains 0.0):
addJump(branchDouble(DoubleEqual, fpRegT0, fpRegT1), target + 2);
In x86, if either operand is NaN, the zero flag is set by definition
(ucomisd instruction). I have no idea why (NaN == anything is true?).
Unfortunately, ARM does not set the zero flag. What would be a
portable
way to improve this instruction? (and emit_op_jtrue) Perhaps this
affects
the possible thumb2 implementation with USE_JSVALUE32_64 as well.
Zoltan
___
webkit-dev mailing list
webkit-dev@lists.webkit.org
http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
___
webkit-dev mailing list
webkit-dev@lists.webkit.org
http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
Re: [webkit-dev] jit, unordered compare
unfortunately cmf was used by the old fpa (floating point accelerator), which is replaced by vfp (vector floating point) for some time. Perhaps I can try a cmpvs reg0, reg0 instruction, which sets zero flag if one argument is NaN (since reg0 is always equal to reg0). Otherwise it does nothing (=nop). That should work, but it would require two branches instead of one (compare to 0, compare to self). I'm sure there must be a way to do an IEEE double comparison on ARM with only one branch. Geoff ___ webkit-dev mailing list webkit-dev@lists.webkit.org http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
