Add environment variable STDSIMD_ASSERT_INSTR_LIMIT to control the instruction limit

Author: gnzlbg

crates/stdsimd-test/src/lib.rs

@@ -257,7 +257,6 @@ fn parse_dumpbin(output: &str) -> HashMap<String, Vec<Function>> {
     ret
 }
 
-
 #[wasm_bindgen(module = "child_process")]
 extern "C" {
     #[wasm_bindgen(js_name = execSync)]
@@ -420,26 +419,29 @@ pub fn assert(fnptr: usize, fnname: &str, expected: &str) {
         break;
     }
 
-    let instruction_limit = match expected {
-        // cpuid returns a pretty big aggregate structure so excempt it from
-        // the slightly more restrictive 22 instructions below
-        "cpuid" => 30,
-
-        // Apparently on Windows LLVM generates a bunch of saves/restores of
-        // xmm registers around these intstructions which blows the 20
-        // limit below. As it seems dictates by Windows's abi (I
-        // guess?) we probably can't do much about it...
-        "vzeroall" | "vzeroupper" if cfg!(windows) => 30,
-
-        // Intrinsics using `cvtpi2ps` are typically "composites" and in some
-        // cases exceed the limit.
-        "cvtpi2ps" => 25,
-
-        // Original limit was 20 instructions, but ARM DSP Intrinsics are
-        // exactly 20 instructions long. So bump the limit to 22 instead of
-        // adding here a long list of expections.
-        _ => 22,
-    };
+    let instruction_limit = std::env::var("STDSIMD_ASSERT_INSTR_LIMIT")
+        .map(|v| v.parse().unwrap())
+        .unwrap_or_else(|_| match expected {
+            // cpuid returns a pretty big aggregate structure so exempt it from
+            // the slightly more restrictive 22 instructions below
+            "cpuid" => 30,
+
+            // Apparently on Windows LLVM generates a bunch of saves/restores
+            // of xmm registers around these intstructions which
+            // blows the 20 limit below. As it seems dictates by
+            // Windows's abi (I guess?) we probably can't do much
+            // about it...
+            "vzeroall" | "vzeroupper" if cfg!(windows) => 30,
+
+            // Intrinsics using `cvtpi2ps` are typically "composites" and in
+            // some cases exceed the limit.
+            "cvtpi2ps" => 25,
+
+            // Original limit was 20 instructions, but ARM DSP Intrinsics are
+            // exactly 20 instructions long. So bump the limit to 22 instead of
+            // adding here a long list of exceptions.
+            _ => 22,
+        });
     let probably_only_one_instruction = instrs.len() < instruction_limit;
 
     if found && probably_only_one_instruction && !inlining_failed {