[Triton] Clean up unused/old env vars (NFC) (#7229)

Author: Mogball

include/triton/Tools/Sys/GetEnv.hpp

@@ -15,7 +15,6 @@ inline const std::set<std::string> CACHE_INVALIDATING_ENV_VARS = {
     // clang-format off
     "AMDGCN_ENABLE_DUMP",
     "AMDGCN_USE_BUFFER_OPS",
-    "DISABLE_FAST_REDUCTION",
     "DISABLE_LLVM_OPT",
     "DISABLE_MMA_V3",
     "DISABLE_MMA_V5",
@@ -30,7 +29,6 @@ inline const std::set<std::string> CACHE_INVALIDATING_ENV_VARS = {
     "MLIR_DISABLE_MULTITHREADING",
     "TRITON_DEFAULT_FP_FUSION",
     "TRITON_DISABLE_LINE_INFO",
-    "TRITON_DISABLE_RESHAPE_ENCODING_INFERENCE",
     "TRITON_ENABLE_LLVM_DEBUG",
     "TRITON_HIP_GLOBAL_PREFETCH",
     "TRITON_HIP_LOCAL_PREFETCH",
@@ -42,7 +40,6 @@ inline const std::set<std::string> CACHE_INVALIDATING_ENV_VARS = {
     "TRITON_OVERRIDE_ARCH",
     "USE_IR_LOC",
     "NVPTX_ENABLE_DUMP",
-    "STORE_TMEM_TO_GLOBAL_BYPASS_SMEM",
     "ALLOW_LHS_TMEM_LAYOUT_CONVERSION",
     "TRITON_F32_DEFAULT",
     "TRITON_PREFER_TMEM_16x256_LAYOUT",

lib/Dialect/TritonGPU/IR/Dialect.cpp

@@ -2241,13 +2241,6 @@ struct TritonGPUInferLayoutInterface
       return success();
     }
 
-    // Feature flag to disable this routine while it's relatively new.
-    // TODO(jlebar): Remove this once we're confident in the code.
-    if (triton::tools::getBoolEnv(
-            "TRITON_DISABLE_RESHAPE_ENCODING_INFERENCE")) {
-      return failure();
-    }
-
     // Cowardly refuse to handle encodings with multiple CTAs.  CTAsPerCGA
     // should be like the other fields in blocked encoding, but I'm not sure how
     // to handle CTASplitNum.