Emulate all 128 bit ops

I thought we could use the fallback code from compiler-builtins
but I guess not? Most things were broken.

I had AI bang out our implementations, I did not write them.

This is generating a lot of code, perhaps better to generate a fn and then
have everything call it?

I added an example and confirmed it works / passes.

Author: LegNeato

Cargo.toml

@@ -16,6 +16,8 @@ members = [
   "examples/cuda/path_tracer/kernels",
   "examples/cuda/sha2_crates_io",
   "examples/cuda/sha2_crates_io/kernels",
+  "examples/cuda/i128_demo",
+  "examples/cuda/i128_demo/kernels",
 
   "examples/optix/*",
   "tests/compiletests",

crates/rustc_codegen_nvvm/libintrinsics.bc

@@ -239,113 +239,6 @@ start:
 }
 declare {i16, i1} @llvm.umul.with.overflow.i16(i16, i16) #0
 
-; This is a bit weird, we need to use functions defined in rust crates (compiler_builtins)
-; as intrinsics in the codegen, but we can't directly use their name, otherwise we will have
-; really odd and incorrect behavior in the crate theyre defined in. So we need to make a wrapper for them that is opaque
-; to the codegen, which is what this is doing.
-
-define {<2 x i64>, i1} @__nvvm_i128_addo(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call {<2 x i64>, i1} @__rust_i128_addo(<2 x i64> %0, <2 x i64> %1)
-  ret {<2 x i64>, i1} %2
-}
-declare {<2 x i64>, i1} @__rust_i128_addo(<2 x i64>, <2 x i64>) #0
-
-define {<2 x i64>, i1} @__nvvm_u128_addo(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call {<2 x i64>, i1} @__rust_u128_addo(<2 x i64> %0, <2 x i64> %1)
-  ret {<2 x i64>, i1} %2
-}
-declare {<2 x i64>, i1} @__rust_u128_addo(<2 x i64>, <2 x i64>) #0
-
-define {<2 x i64>, i1} @__nvvm_i128_subo(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call {<2 x i64>, i1} @__rust_i128_subo(<2 x i64> %0, <2 x i64> %1)
-  ret {<2 x i64>, i1} %2
-}
-declare {<2 x i64>, i1} @__rust_i128_subo(<2 x i64>, <2 x i64>) #0
-
-define {<2 x i64>, i1} @__nvvm_u128_subo(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call {<2 x i64>, i1} @__rust_u128_subo(<2 x i64> %0, <2 x i64> %1)
-  ret {<2 x i64>, i1} %2
-}
-declare {<2 x i64>, i1} @__rust_u128_subo(<2 x i64>, <2 x i64>) #0
-
-define {<2 x i64>, i1} @__nvvm_i128_mulo(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call {<2 x i64>, i1} @__rust_i128_mulo(<2 x i64> %0, <2 x i64> %1)
-  ret {<2 x i64>, i1} %2
-}
-declare {<2 x i64>, i1} @__rust_i128_mulo(<2 x i64>, <2 x i64>) #0
-
-define {<2 x i64>, i1} @__nvvm_u128_mulo(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call {<2 x i64>, i1} @__rust_u128_mulo(<2 x i64> %0, <2 x i64> %1)
-  ret {<2 x i64>, i1} %2
-}
-declare {<2 x i64>, i1} @__rust_u128_mulo(<2 x i64>, <2 x i64>) #0
-
-; Division operations from compiler-builtins
-define <2 x i64> @__nvvm_divti3(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call <2 x i64> @__divti3(<2 x i64> %0, <2 x i64> %1)
-  ret <2 x i64> %2
-}
-declare <2 x i64> @__divti3(<2 x i64>, <2 x i64>) #0
-
-define <2 x i64> @__nvvm_udivti3(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call <2 x i64> @__udivti3(<2 x i64> %0, <2 x i64> %1)
-  ret <2 x i64> %2
-}
-declare <2 x i64> @__udivti3(<2 x i64>, <2 x i64>) #0
-
-; Remainder operations from compiler-builtins
-define <2 x i64> @__nvvm_modti3(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call <2 x i64> @__modti3(<2 x i64> %0, <2 x i64> %1)
-  ret <2 x i64> %2
-}
-declare <2 x i64> @__modti3(<2 x i64>, <2 x i64>) #0
-
-define <2 x i64> @__nvvm_umodti3(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call <2 x i64> @__umodti3(<2 x i64> %0, <2 x i64> %1)
-  ret <2 x i64> %2
-}
-declare <2 x i64> @__umodti3(<2 x i64>, <2 x i64>) #0
-
-; Multiplication from compiler-builtins
-define <2 x i64> @__nvvm_multi3(<2 x i64>, <2 x i64>) #0 {
-start:
-  %2 = call <2 x i64> @__multi3(<2 x i64> %0, <2 x i64> %1)
-  ret <2 x i64> %2
-}
-declare <2 x i64> @__multi3(<2 x i64>, <2 x i64>) #0
-
-; Shift operations from compiler-builtins
-define <2 x i64> @__nvvm_ashlti3(<2 x i64>, i32) #0 {
-start:
-  %2 = call <2 x i64> @__ashlti3(<2 x i64> %0, i32 %1)
-  ret <2 x i64> %2
-}
-declare <2 x i64> @__ashlti3(<2 x i64>, i32) #0
-
-define <2 x i64> @__nvvm_ashrti3(<2 x i64>, i32) #0 {
-start:
-  %2 = call <2 x i64> @__ashrti3(<2 x i64> %0, i32 %1)
-  ret <2 x i64> %2
-}
-declare <2 x i64> @__ashrti3(<2 x i64>, i32) #0
-
-define <2 x i64> @__nvvm_lshrti3(<2 x i64>, i32) #0 {
-start:
-  %2 = call <2 x i64> @__lshrti3(<2 x i64> %0, i32 %1)
-  ret <2 x i64> %2
-}
-declare <2 x i64> @__lshrti3(<2 x i64>, i32) #0
-
 ; Required because we need to explicitly generate { i32, i1 } for the following intrinsics
 ; except rustc will not generate them (it will make { i32, i8 }) which libnvvm rejects.