Emulate i128 ops

LLVM 7.1 (and thus nvvm) lacks native 128-bit integer intrinsics, so emulate them using 64-bit operations.

I also added an example with `sha2` to confirm it works and show
folks that crates.io crates can often be used directly unmodified.

Fixes #207.

Author: LegNeato

Cargo.toml

@@ -14,6 +14,8 @@ members = [
   "examples/cuda/gemm/kernels",
   "examples/cuda/path_tracer",
   "examples/cuda/path_tracer/kernels",
+  "examples/cuda/sha2_crates_io",
+  "examples/cuda/sha2_crates_io/kernels",
 
   "examples/optix/*",
   "tests/compiletests",

crates/optix-sys/build/main.rs

@@ -16,7 +16,7 @@ fn main() {
 
     println!("cargo::rerun-if-changed=build");
     for e in sdk.related_optix_envs() {
-        println!("cargo::rerun-if-env-changed={}", e);
+        println!("cargo::rerun-if-env-changed={e}");
     }
     // Emit metadata for the build script.
     println!("cargo::metadata=root={}", sdk.optix_root().display());
@@ -36,7 +36,7 @@ fn main() {
     let metadata_optix_include = env::join_paths(sdk.optix_include_paths())
         .map(|s| s.to_string_lossy().to_string())
         .expect("Failed to build metadata for include.");
-    println!("cargo::metadata=include_dir={}", metadata_optix_include);
+    println!("cargo::metadata=include_dir={metadata_optix_include}");
 
     // Generate optix bindings.
     create_optix_bindings(&sdk, &cuda_include_paths);

crates/optix-sys/build/optix_sdk.rs

@@ -89,7 +89,7 @@ impl OptiXSdk {
             .ok_or("Cannot find OPTIX_VERSION from OptiX header file.")?;
         let version = version
             .parse::<u32>()
-            .map_err(|_| format!("Cannot parse OPTIX_VERSION as u32: '{}'", version))?;
+            .map_err(|_| format!("Cannot parse OPTIX_VERSION as u32: '{version}'"))?;
         Ok(version)
     }
 }

crates/rustc_codegen_nvvm/libintrinsics.bc

@@ -286,6 +286,66 @@ start:
 }
 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.