Question about ASTContext::getPromotedIntegerType() function

Clang Frontend
#1

Hi all,

I have a question about ASTContext::getPromotedIntegerType() function.

I have tested a code as following:

source code:
int main()
{
   volatile unsigned short x;
   volatile unsigned short y;
   volatile unsigned short result;

   x = 8;
   y = 2;

   result = x / y;

   return 0;
}

Generated IR from clang:
   6 define i32 @main() #0 {
   7 entry:
   8 %retval = alloca i32, align 4
   9 %x = alloca i16, align 2
  10 %y = alloca i16, align 2
  11 %result = alloca i16, align 2
  12 store i32 0, i32* %retval
  13 store volatile i16 8, i16* %x, align 2
  14 store volatile i16 2, i16* %y, align 2
  15 %0 = load volatile i16* %x, align 2
  16 %conv = zext i16 %0 to i32
  17 %1 = load volatile i16* %y, align 2
  18 %conv1 = zext i16 %1 to i32
  19 %div = sdiv i32 %conv, %conv1
  20 %conv2 = trunc i32 %div to i16
  21 store volatile i16 %conv2, i16* %result, align 2
  22 ret i32 0
  23 }

I expected "udiv" instruction on line 19 but there was "sdiv" instruction because of "ASTContext::getPromotedIntegerType()" function promotes "unsigned short" to "int".

Could someone explain why this function returns "IntTy" when "PromotableSize" is not same with "IntSize"? I think that the line of "ASTContext::getPromotedIntegerType()" function should be changed.
ex)
return (PromotableSize != IntSize) ? IntTy : UnsignedIntTy;
--> return UnsignedIntTy;

Thanks,
JinGu Kang

clang.patch (551 Bytes)

#2

Because that's what the standard says. C11 6.3.1.1: "If an int can
represent all values of the original type (as restricted by the width,
for a bit-field), the value is converted to an int; otherwise, it is
converted to an unsigned int."

-Eli

#3

Hi Eli,

I appreciate your comment. It's really helpful to understand the code.

How do you feel about generating "sdiv" instruction on previous example?

Thanks,
JinGu Kang

#4

How else would you divide two signed integers? Instcombine knows how
to optimize it to a udiv. I mean, we could hack an optimization into
clang, but it would be the same optimization instcombine does, and
there isn't much point to optimizing the performance of -O0 code.

-Eli

#5

2013-07-17 오전 3:00, Eli Friedman 쓴 글: