update analyzed_proc_list to include instrs

src/exo/LoopIR_compiler.py

@@ -378,7 +378,7 @@ def from_lines(x):
     private_fwd_decls = []
     proc_bodies = []
     instrs_global = []
-    new_proc_list = []
+    analyzed_proc_list = []
 
     needed_helpers = set()
 
@@ -410,7 +410,6 @@ def from_lines(x):
             p = PrecisionAnalysis().run(p)
             p = WindowAnalysis().apply_proc(p)
             p = MemoryAnalysis().run(p)
-            new_proc_list.append(p)
 
             comp = Compiler(p, ctxt_name, is_public_decl=is_public_decl)
             d, b = comp.comp_top()
@@ -424,6 +423,8 @@ def from_lines(x):
 
             proc_bodies.append(b)
 
+        analyzed_proc_list.append(p)
+
     # Structs are just blobs of code... still sort them for output stability
     struct_defns = [x.definition for x in sorted(struct_defns, key=lambda x: x.name)]
 
@@ -454,8 +455,8 @@ def from_lines(x):
 {from_lines(public_fwd_decls)}
 """
 
-    memory_code = _compile_memories(find_all_mems(new_proc_list))
-    extern_code = _compile_externs(find_all_externs(new_proc_list))
+    memory_code = _compile_memories(find_all_mems(analyzed_proc_list))
+    extern_code = _compile_externs(find_all_externs(analyzed_proc_list))
 
     helper_code = [_static_helpers[v] for v in needed_helpers]
     body_contents = [

tests/golden/asplos25/test_gemmini_matmul_new/test_matmul.txt

@@ -99,6 +99,11 @@ void matmul_on_gemmini( c_code_str_Context *ctxt, int_fast32_t N, int_fast32_t M
 #include "gemm_acc_malloc.h"
 #include <include/gemmini.h>
 #include "gemm_malloc.h"
+int8_t _relu_int8_t(int8_t x) {
+    if (x > 0.0) return x;
+    else return 0.0;
+}
+
 
 /* relying on the following instruction..."
 config_ld_i8_id1(src_stride)

tests/golden/test_apps/test_x86_conv.txt

@@ -65,6 +65,11 @@ void conv_specialized( void *ctxt, const float* inp, float* output, const float*
 #include <stdio.h>
 #include <stdlib.h>
 
+float _relu_float(float x) {
+    if (x > 0.0) return x;
+    else return 0.0;
+}
+
 // conv_specialized(
 //     inp : f32[5, 82, 102, 128] @DRAM,
 //     output : f32[5, 80, 100, 128] @DRAM,