[llvm-commits] [llvm-gcc-4.2] r82640 - /llvm-gcc-4.2/trunk/gcc/llvm-convert.cpp

[Chris Lattner]

On Sep 23, 2009, at 6:10 PM, Evan Cheng wrote:
On Sep 23, 2009, at 3:21 PM, Dale Johannesen wrote:
>> On Sep 23, 2009, at 3:18 PMPDT, Evan Cheng wrote:
>>
>>> Are we generating a call to sqrt now? If so, that's bad. We should
>>> be using SSE sqrts* instructions.
>>
>> Agreed.  My proposed semantic change would fix that.
>
> You mean change to llvm.sqrt and then llvm-gcc can switch bad to
> generating the intrinsic? It seems like the current fix is not what we
> want. Perhaps we should revert it first?
>
> According to Chris, the semantics of sqrt of negative value is defined
> and this is just some optimization bug.

Sorry, I'm just catching up on this now.  There are a couple of things  
that confuse the issue, but I'll just try to keep it "to the point"  
instead of rambling about history.

1. I did tell Evan that sqrt is defined on negative number, but I  
misunderstood and didn't think about llvm.sqrt.  Please disregard my  
comment Evan.  llvm.sqrt should be undefined on negative numbers as  
langref says, and llvm-gcc/clang should only transform sqrt to  
llvm.sqrt if the appropriate "I don't care about fp semantics" flag is  
set.

2. Dale's patch to llvm-gcc is ok, but it would be better to still do  
the transformation when -ffast-math is specified or whatever the more  
precise "nan's aren't generated" flag is.  We should do the same thing  
for clang as well.

3. Please make sure that llvm-gcc/clang on the mac (and other targets  
with -fno-math-errno) are producing a call to sqrt that is marked as  
readnone.  Given this, the mid-level optimizer should hoist and cse  
the calls to sqrt just as well as it did calls to llvm.sqrt.

4. The constant folding of llvm.sqrt(-123) -> 0 is ok because the  
intrinsic really is undefined on negative.  The constant folding of  
sqrt(-123) doesn't fold if the input is negative, so it will just not  
optimize the curious case.

5. Please make the X86 backend compile calls to readonly/readnone  
"sqrt" produce a sqrtsd (etc) instruction.  We really don't want to  
get a function call on the mac (or other x86 target with -fno-math- 
errno).  Like malloc, if someone cares about -fno-builtin-sqrt, they  
can solve the general problem.  We already constant fold "real sqrt"  
calls in Analysis/ConstantFolding.cpp anyway.

6. Please add handling of "real sqrt" to llvm::CannotBeNegativeZero to  
match the handling of llvm.sqrt.

Thanks!

-Chris