[LLVMdev] floor

[Reed Kotler]

Consider

include <math.h>

float y = 1.0;

int main() {
   float x = sin(y);
   printf("%e \n", x);

}


Sin works.

/local/llvmpb_config/install/bin/clang -target mipsel-linux-gnu sin1.c 
-o sin1.s -lm -mips16 -S -fPIC


....
	lw	$3, %got(y)($2)
	lw	$4, 0($3)
	lw	$3, %call16(__mips16_extendsfdf2)($2)
	move	$25, $3
	move	$gp, $2
	sw	$2, 36 ( $16 );
	sw	$3, 32 ( $16 );
	jalrc 	$3
	lw	$4, 36 ( $16 );
	lw	$5, %call16(sin)($4)
	lw	$6, %got(__mips16_call_stub_df_2)($4)
	sw	$2, 28 ( $16 );
	move	$2, $5
	lw	$4, 28 ( $16 );
	move	$5, $3
	lw	$3, 36 ( $16 );
	move	$gp, $3
	jalrc 	$6


....

On 07/26/2013 03:59 PM, Reed Kotler wrote:
>
> Here is a test case:
>
> extern double floor(double);
> extern double floor_(double);
>
> double x = 1.5;
> double y, y_;
>
> void foo() {
>
>    double y = floor(x);
>    double y_ = floor_(x);
> }
>
>
> If I compile this for Mips16, it calls the proper helper function for
> floor_ but not for floor, because the signature for floor in callee info
> is wrong. Args[0] = void RetTy = void
>
> /local/llvmpb_config/install/bin/clang -target mipsel-linux-gnu floor1.c
> -o floor1.s -mips16 -S -fPIC
>
> .....
>      lw    $3, %got(x)($2)
>      lw    $4, 0($3)
>      lw    $5, 4($3)
>      lw    $6, %call16(floor)($2)
>      move    $25, $6
>      move    $gp, $2
>      sw    $2, 20 ( $16 );
>      sw    $3, 16 ( $16 );
>      jalrc     $6
>      sw    $3, 36($16)
>      sw    $2, 32($16)
>      lw    $2, 16 ( $16 );
>      lw    $5, 4($2)
>      lw    $4, 0($2)
>      lw    $3, 20 ( $16 );
>      lw    $2, %call16(floor_)($3)
>      lw    $6, %got(__mips16_call_stub_df_2)($3)
>      move    $gp, $3
>      jalrc     $6
>
> ....
>
>
> On 07/26/2013 03:33 PM, reed kotler wrote:
>> I'm getting some problems because it seems that the compiler is treating
>> "floor" differently from other math library functions like "sin".
>>
>> The Args and RetVal have the parameter and return types marked as void.
>>
>> For mips16, it's important that I be able to know the original signature
>> for floating point functions.
>>
>> In some cases, need to create calls to helper functions based on the
>> signatures.
>>
>> In newer code I've written, I've actually moved this logica to an IR
>> pass and in that case I know for sure.
>>
>> But this part of the code is in ISelLowering code and I rely on getting
>> the proper signature information.
>>
>> I'm looking at llvm now to see how this is occurring but maybe someone
>> just knows.
>>
>> Tia.
>>
>> Reed