[DenseMap] Update combineHashValue

[MaskRay]

combineHashValue is a custom bit mixer from 2008
(5fc8ab6) used for std::pair and
std::tuple. It has a long dependency chain and slow. Replace it with
a simply multiply-xorshift style hash using a constant from
splitmix641. abseil-cpp and carbon also use this style, but with
uint128 to probably get a lower avalanche bias. We don't use uint128 for
MSVC portability.

Measured time to compute [0,1000000000) values on an i7-11850H:

• old: 1.163s
• new: 0.427s

[llvmbot]

@llvm/pr-subscribers-llvm-adt

Author: Fangrui Song (MaskRay)

Changes

combineHashValue is a custom bit mixer from 2008
(5fc8ab6) used for std::pair and
std::tuple. It has a long dependency chain and slow. Replace it with
Moremur from Pelle Evensen1

Measured time to compute [0,1000000000) values on an i7-11850H:

• old: 1.152s
• new: 0.769s

---

Full diff: https://github.com/llvm/llvm-project/pull/95970.diff

1 Files Affected:

• (modified) llvm/include/llvm/ADT/DenseMapInfo.h (+8-10)

diff --git a/llvm/include/llvm/ADT/DenseMapInfo.h b/llvm/include/llvm/ADT/DenseMapInfo.h
index 5b7dce7b53c62..ba05309e4e413 100644
--- a/llvm/include/llvm/ADT/DenseMapInfo.h
+++ b/llvm/include/llvm/ADT/DenseMapInfo.h
@@ -26,17 +26,15 @@ namespace llvm {
 namespace detail {
 
 /// Simplistic combination of 32-bit hash values into 32-bit hash values.
+/// This uses a Murmur3-type mixer "Moremur" from Pelle Evensen.
 static inline unsigned combineHashValue(unsigned a, unsigned b) {
-  uint64_t key = (uint64_t)a << 32 | (uint64_t)b;
-  key += ~(key << 32);
-  key ^= (key >> 22);
-  key += ~(key << 13);
-  key ^= (key >> 8);
-  key += (key << 3);
-  key ^= (key >> 15);
-  key += ~(key << 27);
-  key ^= (key >> 31);
-  return (unsigned)key;
+  uint64_t x = (uint64_t)a << 32 | (uint64_t)b;
+  x ^= x >> 27;
+  x *= 0x3C79AC492BA7B653UL;
+  x ^= x >> 33;
+  x *= 0x1C69B3F74AC4AE35UL;
+  x ^= x >> 27;
+  return (unsigned)x;
 }
 
 } // end namespace detail

[nikic]

LGTM

[ChuanqiXu9]

LGTM.

[MaskRay]

Significant instruction:u difference for -O3 builds. This commit itself affects std::pair.

https://llvm-compile-time-tracker.com/compare.php?from=58d7a6e0e6361871442df956bb88798ce602b09d&to=fb17bbce80cf76ce1a31eff463f451f626bc36b5&stat=instructions:u

stage2-O3:

Benchmark	Old	New
kimwitu++	39227M	39106M (-0.31%)
sqlite3	35441M	35309M (-0.37%)
consumer-typeset	32091M	32021M (-0.22%)
Bullet	93517M	93231M (-0.31%)
tramp3d-v4	79022M	78706M (-0.40%)
mafft	33162M	33040M (-0.37%)
ClamAV	50706M	50535M (-0.34%)
lencod	61560M	61298M (-0.42%)
SPASS	43262M	43080M (-0.42%)
7zip	192562M	192060M (-0.26%)
geomean	55639M	55449M (-0.34%)

After #95734 and af82e63 , DenseMap<unsigned long, X> and DenseMap<unsigned long long, X> are affected as well.

[kazutakahirata]

Significant instruction:u difference for -O3 builds.

Very nice!