[builtins] Inline __paritysi2 into __paritydi2 and inline __paritydi2 into __parityti2.

[craig.topper]

Summary

No point in making parityti2 go through 2 calls to get to
paritysi2.

[MaskRay]

I think this is correct, but does their performance matter? llvm/IR/RuntimeLibcalls.def does not define them (they cannot be produced by llvm). Targets either emit popcount & 1 or detect the idiom and emit an optimized parity (x86 after PR46954)

[craig.topper]

In D87218#2259972, @MaskRay wrote:

I think this is correct, but does their performance matter? llvm/IR/RuntimeLibcalls.def does not define them (they cannot be produced by llvm). Targets either emit popcount & 1 or detect the idiom and emit an optimized parity (x86 after PR46954)

Performance probably doesn't matter. Just one of the things a certain someone wasn't amused by.

[MaskRay]

LG. Please wait a bit for @efriedma's comments.

[efriedma]

LGTM

compiler-rt/lib/builtins/paritydi2.c

 
 #include "int_lib.h"
 
 // Returns: 1 if number of bits is odd else returns 0
 
 COMPILER_RT_ABI int __paritydi2(di_int a) {
   dwords x;
   x.all = a;
-  return __paritysi2(x.s.high ^ x.s.low);
+  su_int x2 = x.s.high ^ x.s.low;
+  x2 ^= x2 >> 16;
+  x2 ^= x2 >> 8;
+  x2 ^= x2 >> 4;
+  return (0x6996 >> (x2 & 0xF)) & 1;
 }

compiler-rt/lib/builtins/parityti2.c

 #include "int_lib.h"
 
 #ifdef CRT_HAS_128BIT
 
 // Returns: 1 if number of bits is odd else returns 0
 
 COMPILER_RT_ABI int __parityti2(ti_int a) {
   twords x;
+  dwords x2;
   x.all = a;
-  return __paritydi2(x.s.high ^ x.s.low);
+  x2.all = x.s.high ^ x.s.low;
+  su_int x3 = x2.s.high ^ x2.s.low;
+  x3 ^= x3 >> 16;
+  x3 ^= x3 >> 8;
+  x3 ^= x3 >> 4;
+  return (0x6996 >> (x3 & 0xF)) & 1;
 }
 
 #endif // CRT_HAS_128BIT