[clang] [llvm] [DirectX] emmit `dx.precise` metadata when using `-Gis` flag (PR #192526)

[via cfe-commits]

================
@@ -5,13 +5,76 @@
 
 target triple = "dxil-pc-shadermodel6.6-compute"
 
-; ENABLED: call float @dx.op.unary.f32(i32 12, float %conv)
+; ENABLED: call float @dx.op.unary.f32(i32 7,
 ; ENABLED-SAME: !dx.precise ![[SM:[0-9]+]]
+; ENABLED-COUNT-26: !dx.precise ![[SM]]
+; ENABLED-NOT: !dx.precise ![[SM]]
 ; ENABLED: ![[SM]] = !{i32 1}
 
 ; DISABLED-NOT: !dx.precise ![[SM:[0-9]+]]
-define void @main(float %conv) {
+define void @unary(float %p) {
----------------
joaosaffran wrote:

I've refactored the test to make it easier to review. All checks are now per function. I've made the `@wave_rla_i32` check explicit, so it is clearly checking for the absense of dx.precise. This was the only operations that had an i32 overload for me to test it. All the others only allow float types.

https://github.com/llvm/llvm-project/pull/192526