From 2e911139688fde9138fa4ee84e999bcc78c56931 Mon Sep 17 00:00:00 2001
From: clamchowder <lamchester@gmail.com>
Date: Thu, 8 Aug 2024 12:52:37 -0700
Subject: [PATCH] save x87 state

---
 AsmGen/UarchTest.cs        |  1 +
 AsmGen/UarchTestHelpers.cs | 39 +++-----------------------------------
 AsmGen/tests/MmxRfTest.cs  |  9 +++++++--
 3 files changed, 11 insertions(+), 38 deletions(-)

diff --git a/AsmGen/UarchTest.cs b/AsmGen/UarchTest.cs
index 18c11ba..3530b34 100644
--- a/AsmGen/UarchTest.cs
+++ b/AsmGen/UarchTest.cs
@@ -93,6 +93,7 @@ public void GenerateTestBlock(StringBuilder sb, IUarchTest.ISA isa)
                 sb.AppendLine("#endif");
                 sb.AppendLine("    gettimeofday(&endTv, &endTz);");
                 sb.AppendLine("    time_diff_ms = 1000 * (endTv.tv_sec - startTv.tv_sec) + ((endTv.tv_usec - startTv.tv_usec) / 1000);");
+                //sb.AppendLine("    fprintf(stderr, \"%lu ms elapsed, %lu iter\\n\", time_diff_ms, structIterations);");
                 if (DivideTimeByCount)
                     sb.AppendLine("    latency = 1e6 * (float)time_diff_ms / (float)(iterations);");
                 else
diff --git a/AsmGen/UarchTestHelpers.cs b/AsmGen/UarchTestHelpers.cs
index 16ef70c..2dbe569 100644
--- a/AsmGen/UarchTestHelpers.cs
+++ b/AsmGen/UarchTestHelpers.cs
@@ -47,41 +47,6 @@ public static void GenerateVsExternLines(StringBuilder sb, UarchTest test)
                 sb.AppendLine("extern \"C\" uint64_t " + test.Prefix + counts[i] + $"({test.FunctionDefinitionParameters});");
         }
 
-        public static void GenerateTestBlock(StringBuilder sb, UarchTest test)
-        {
-            sb.AppendLine("  if (argc > 1 && strncmp(test_name, \"" + test.Prefix + "\", " + test.Prefix.Length + ") == 0) {");
-            sb.AppendLine("    printf(\"" + test.Description + ":\\n\");");
-
-            int[] counts = test.Counts;
-            for (int i = 0; i < counts.Length; i++)
-            {
-                // use more iterations (iterations = structIterations * 100) and divide iteration count by tested-thing count
-                // for certain tests like call stack depth
-                if (test.DivideTimeByCount)
-                {
-                    sb.AppendLine("    tmp = structIterations;");
-                    sb.AppendLine("    structIterations = iterations / " + counts[i] + ";");
-                }
-
-                sb.AppendLine("    gettimeofday(&startTv, &startTz);");
-                sb.AppendLine("    " + test.Prefix + counts[i] + $"({test.GetFunctionCallParameters});");
-                sb.AppendLine("    gettimeofday(&endTv, &endTz);");
-                sb.AppendLine("    time_diff_ms = 1000 * (endTv.tv_sec - startTv.tv_sec) + ((endTv.tv_usec - startTv.tv_usec) / 1000);");
-                if (test.DivideTimeByCount)
-                    sb.AppendLine("    latency = 1e6 * (float)time_diff_ms / (float)(iterations);");
-                else
-                    sb.AppendLine("    latency = 1e6 * (float)time_diff_ms / (float)(structIterations);");
-                sb.AppendLine("    printf(\"" + counts[i] + ",%f\\n\", latency);\n");
-
-                if (test.DivideTimeByCount)
-                {
-                    sb.AppendLine("    structIterations = tmp;");
-                }
-            }
-
-            sb.AppendLine("  }\n");
-        }
-
         /// <summary>
         /// Generates test functions in assembly, with filler instructions between two divs
         /// Args are put into rcx, rdx, r8 (in that order) to match Windows calling convention
@@ -324,7 +289,8 @@ public static void GenerateX86AsmStructureTestFuncs(StringBuilder sb,
             string initInstrs = null,
             string postLoadInstrs1 = null,
             string postLoadInstrs2 = null,
-            bool lfence = true)
+            bool lfence = true,
+            string cleanupInstrs = null)
         {
             for (int i = 0; i < counts.Length; i++)
             {
@@ -383,6 +349,7 @@ public static void GenerateX86AsmStructureTestFuncs(StringBuilder sb,
 
                 sb.AppendLine("  dec %rcx");
                 sb.AppendLine("  jne " + funcName + "start");
+                if (cleanupInstrs != null) sb.AppendLine(cleanupInstrs);
                 sb.AppendLine("  pop %rdx");
                 sb.AppendLine("  pop %rcx");
                 sb.AppendLine("  pop %r8");
diff --git a/AsmGen/tests/MmxRfTest.cs b/AsmGen/tests/MmxRfTest.cs
index 47a5995..65e4fa6 100644
--- a/AsmGen/tests/MmxRfTest.cs
+++ b/AsmGen/tests/MmxRfTest.cs
@@ -26,19 +26,24 @@ public override void GenerateAsm(StringBuilder sb, IUarchTest.ISA isa)
 
         public void GenerateX86GccAsm(StringBuilder sb)
         {
-            string initInstrs = "  movq (%rdx), %mm0\n" +
+            string initInstrs = 
+                "  fsave (%r8)\n" +
+                "  movq (%rdx), %mm0\n" +
                 "  movq 8(%rdx), %mm1\n" +
                 "  movq 16(%rdx), %mm2\n" +
                 "  movq 24(%rdx), %mm3\n" +
                 "  movq 32(%rdx), %mm4\n";
 
+            string cleanupInstrs = "  frstor (%r8)";
+
             string[] unrolledAdds = new string[4];
             unrolledAdds[0] = "  paddw %mm0, %mm1";
             unrolledAdds[1] = "  paddw %mm0, %mm2";
             unrolledAdds[2] = "  paddw %mm0, %mm3";
             unrolledAdds[3] = "  paddw %mm0, %mm4";
 
-            UarchTestHelpers.GenerateX86AsmStructureTestFuncs(sb, this.Counts, this.Prefix, unrolledAdds, unrolledAdds, initInstrs: initInstrs);
+            UarchTestHelpers.GenerateX86AsmStructureTestFuncs(
+                sb, this.Counts, this.Prefix, unrolledAdds, unrolledAdds, initInstrs: initInstrs, cleanupInstrs: cleanupInstrs);
         }
     }
 }
\ No newline at end of file