[patch] improved patch to remove ebx from asm blocks in test.c

memtest86+-5.01-test_c_noebx.patch (text/plain), 10.81 KB, created by Peter Fox on 2016-12-26 20:23:18 UTC

--- test.orig.c	2016-12-25 22:22:44.065762386 +0000
+++ test.c	2016-12-26 19:43:45.618874725 +0000
@@ -348,18 +369,26 @@
  */
 
                         asm __volatile__ (
-                                "jmp L200\n\t"
-                                ".p2align 4,,7\n\t"
-                                "L201:\n\t"
-                                "addl $4,%%edi\n\t"
-                                "L200:\n\t"
-				"pushl %%ecx\n\t" \
-                                "call rand\n\t"
-				"popl %%ecx\n\t" \
-				"movl %%eax,(%%edi)\n\t"
-                                "cmpl %%ebx,%%edi\n\t"
-                                "jb L201\n\t"
-                                : : "D" (p), "b" (pe), "c" (me)
+							// Put pe(edx) on stack for protection from rand calls
+							"pushl %%edx\n\t"
+							// Put me(ecx) on stack for repeated calls
+							"pushl %%ecx\n\t"
+							"jmp L200\n\t"
+							".p2align 4,,7\n\t"
+							"L201:\n\t"
+							"addl $4,%%edi\n\t"
+							"L200:\n\t"
+							// Destroys ecx, but we don't care as it is already on tos as the first parameter
+							"call rand\n\t"
+							"movl %%eax,(%%edi)\n\t"
+							// pe(edx) is on stack now
+							"cmpl 4(%%esp),%%edi\n\t"
+							"jb L201\n\t"
+							// Finally pop me(ecx)
+							"popl %%ecx\n\t"
+							// and pe(edx)
+							"popl %%edx\n\t"
+							: : "D" (p), "d" (pe), "c" (me)
 				: "eax"
                         );
 			p = pe + 1;
@@ -413,7 +442,11 @@
 				}
 				asm __volatile__ (
 					
-                    "pushl %%ebp\n\t"
+					// Save xorVal on stack as we use ecx later
+                    "pushl %%ecx\n\t"
+					// Push me(edx) onto tos ready for rand calls
+					"pushl %%edx\n\t" // Push function inputs onto stack
+					// edx now available as general register
 
 					// Skip first increment
 					"jmp L26\n\t"
@@ -429,16 +462,14 @@
 					// Get next random number, pass in me(edx), random value returned in num(eax)
 					// num = rand(me);
 					// cdecl call maintains all registers except eax, ecx, and edx
-					// We maintain edx with a push and pop here using it also as an input
+					// me(edx) is already on tos, and we keep it there saving pushes and pops
 					// we don't need the current eax value and want it to change to the return value
 					// we overwrite ecx shortly after this discarding its current value
-					"pushl %%edx\n\t" // Push function inputs onto stack
 					"call rand\n\t"
-					"popl %%edx\n\t" // Remove function inputs from stack
 
-					// XOR the random number with xorVal(ebx), which is either 0xffffffff or 0 depending on the outer loop
+					// XOR the random number with xorVal(ecx), which is either 0xffffffff or 0 depending on the outer loop
 					// if (i) { num = ~num; }
-					"xorl %%ebx,%%eax\n\t"
+					"xorl 4(%%esp),%%eax\n\t"
 
 					// Move the current value of the current position p(edi) into bad(ecx)
 					// (bad=*p)
@@ -453,19 +484,20 @@
 					// Set a new value or not num(eax) at the current position p(edi)
 					// *p = ~num;
 					"L25:\n\t"
-					"movl $0xffffffff,%%ebp\n\t"
-					"xorl %%ebp,%%eax\n\t"
+					"movl $0xffffffff,%%edx\n\t"
+					"xorl %%edx,%%eax\n\t"
 					"movl %%eax,(%%edi)\n\t"
 
 					// Loop until current position p(edi) equals the end position pe(esi)
-					"cmpl %%esi,%%edi\n\t"
+					"movl %%esi,%%edx\n\t"
+					"cmpl %%edx,%%edi\n\t"
 					"jb L27\n\t"
 					"jmp L24\n"
 
 					// Error case
 					"L23:\n\t"
 					// Must manually maintain eax, ecx, and edx as part of cdecl call convention
-					"pushl %%edx\n\t"
+					// but don't need to preserve edx register
 					"pushl %%ecx\n\t" // Next three pushes are functions input
 					"pushl %%eax\n\t"
 					"pushl %%edi\n\t"
@@ -473,14 +505,14 @@
 					"popl %%edi\n\t" // Remove function inputs from stack and restore register values
 					"popl %%eax\n\t"
 					"popl %%ecx\n\t"
-					"popl %%edx\n\t"
 					"jmp L25\n" 
 
 					"L24:\n\t"
-                                        "popl %%ebp\n\t"
-					:: "D" (p), "S" (pe), "b" (xorVal),
+					"popl %%edx\n\t" // Remove rand function input me(edx) from stack
+					"popl %%ecx\n\t" // xorVal
+					:: "D" (p), "S" (pe), "c" (xorVal),
 						 "d" (me)
-					: "eax", "ecx"
+					: "eax"
 				);
 				p = pe + 1;
 			} while (!done);
@@ -589,7 +621,7 @@
 					"cmpl %%eax,%%ecx\n\t" \
 					"jne L3\n\t" \
 					"L5:\n\t" \
-					"movl %%ebx,(%%edi)\n\t" \
+					"movl %%esi,(%%edi)\n\t" \
 					"cmpl %%edx,%%edi\n\t" \
 					"jb L0\n\t" \
 					"jmp L4\n" \
@@ -609,7 +641,7 @@
 					"jmp L5\n" \
 
 					"L4:\n\t" \
-					:: "a" (p1), "D" (p), "d" (pe), "b" (p2)
+					:: "a" (p1), "D" (p), "d" (pe), "S" (p2)
 					: "ecx"
 				);
 				p = pe + 1;
@@ -658,7 +690,7 @@
 					"subl $4, %%edi\n\t"
 					"L9:\n\t"
 					"movl (%%edi),%%ecx\n\t"
-					"cmpl %%ebx,%%ecx\n\t"
+					"cmpl %%esi,%%ecx\n\t"
 					"jne L6\n\t"
 					"L10:\n\t"
 					"movl %%eax,(%%edi)\n\t"
@@ -670,18 +702,18 @@
 					"pushl %%edx\n\t"
 					"pushl %%eax\n\t"
 					"pushl %%ecx\n\t"
-					"pushl %%ebx\n\t"
+					"pushl %%esi\n\t"
 					"pushl %%edi\n\t"
 					"call error\n\t"
 					"popl %%edi\n\t"
-					"popl %%ebx\n\t"
+					"popl %%esi\n\t"
 					"popl %%ecx\n\t"
 					"popl %%eax\n\t"
 					"popl %%edx\n\t"
 					"jmp L10\n"
 
 					"L7:\n\t"
-					:: "a" (p1), "D" (p), "d" (pe), "b" (p2)
+					:: "a" (p1), "D" (p), "d" (pe), "S" (p2)
 					: "ecx"
 				);
 				p = pe - 1;
@@ -745,20 +777,20 @@
                                 "addl $4,%%edi\n\t"
                                 "L20:\n\t"
                                 "movl %%ecx,(%%edi)\n\t"
-                                "addl $1,%%ebx\n\t"
-                                "cmpl $32,%%ebx\n\t"
+                                "addl $1,%%edx\n\t"
+                                "cmpl $32,%%edx\n\t"
                                 "jne L21\n\t"
                                 "movl %%esi,%%ecx\n\t"
-                                "xorl %%ebx,%%ebx\n\t"
+                                "xorl %%edx,%%edx\n\t"
                                 "jmp L22\n"
                                 "L21:\n\t"
                                 "shll $1,%%ecx\n\t"
                                 "orl %%eax,%%ecx\n\t"
                                 "L22:\n\t"
-                                "cmpl %%edx,%%edi\n\t"
+                                "cmpl %3,%%edi\n\t"
                                 "jb L923\n\t"
-                                : "=b" (k), "=c" (pat)
-                                : "D" (p),"d" (pe),"b" (k),"c" (pat),
+                                : "=d" (k), "=c" (pat)
+                                : "D" (p),"m" (pe),"d" (k),"c" (pat),
                                         "a" (sval), "S" (lb)
 			);
 			p = pe + 1;
@@ -826,17 +858,19 @@
                                         "notl %%ecx\n\t"
                                         "movl %%ecx,(%%edi)\n\t"
                                         "notl %%ecx\n\t"
-                                        "incl %%ebx\n\t"
-                                        "cmpl $32,%%ebx\n\t"
+                                        "incl %%edx\n\t"
+                                        "cmpl $32,%%edx\n\t"
                                         "jne L31\n\t"
                                         "movl %%esi,%%ecx\n\t"
-                                        "xorl %%ebx,%%ebx\n\t"
+                                        "xorl %%edx,%%edx\n\t"
                                         "jmp L32\n"
                                         "L31:\n\t"
                                         "shll $1,%%ecx\n\t"
                                         "orl %%eax,%%ecx\n\t"
 					"L32:\n\t"
-                                        "cmpl %%edx,%%edi\n\t"
+                                        "popl %%ebp\n\t"
+                                        "cmpl %3,%%edi\n\t"
+                                        "pushl %%ebp\n\t"
                                         "jb L930\n\t"
                                         "jmp L33\n\t"
 
@@ -860,8 +894,8 @@
 
                                         "L33:\n\t"
                                         "popl %%ebp\n\t"
-                                        : "=b" (k),"=c" (pat)
-                                        : "D" (p),"d" (pe),"b" (k),"c" (pat),
+                                        : "=d" (k),"=c" (pat)
+                                        : "D" (p),"m" (pe),"d" (k),"c" (pat),
                                                 "a" (sval), "S" (lb)
 				);
 				p = pe + 1;
@@ -935,17 +969,19 @@
                                         "L45:\n\t"
                                         "notl %%ecx\n\t"
                                         "movl %%ecx,(%%edi)\n\t"
-                                        "decl %%ebx\n\t"
-                                        "cmpl $0,%%ebx\n\t"
+                                        "decl %%edx\n\t"
+                                        "cmpl $0,%%edx\n\t"
                                         "jg L41\n\t"
                                         "movl %%esi,%%ecx\n\t"
-                                        "movl $32,%%ebx\n\t"
+                                        "movl $32,%%edx\n\t"
                                         "jmp L42\n"
                                         "L41:\n\t"
                                         "shrl $1,%%ecx\n\t"
                                         "orl %%eax,%%ecx\n\t"
 					"L42:\n\t"
-                                        "cmpl %%edx,%%edi\n\t"
+                                        "popl %%ebp\n\t"
+                                        "cmpl %3,%%edi\n\t"
+                                        "pushl %%ebp\n\t"
                                         "ja L49\n\t"
                                         "jmp L43\n\t"
 
@@ -969,8 +1005,8 @@
 
                                         "L43:\n\t"
                                         "popl %%ebp\n\t"
-                                        : "=b" (k), "=c" (pat)
-                                        : "D" (p),"d" (pe),"b" (k),"c" (pat),
+                                        : "=d" (k), "=c" (pat)
+                                        : "D" (p),"m" (pe),"d" (k),"c" (pat),
                                                 "a" (p3), "S" (hb)
 				);
 				p = pe - 1;
@@ -1082,20 +1118,20 @@
 					"L54:\n\t" \
 					"addl $4,%%edi\n\t" \
 					"L50:\n\t" \
-					"cmpl %%ebx,%%ecx\n\t" \
+					"cmpl %%edx,%%ecx\n\t" \
 					"je L52\n\t" \
 					  "movl %%eax,(%%edi)\n\t" \
 					"L52:\n\t" \
-					"incl %%ebx\n\t" \
-					"cmpl $19,%%ebx\n\t" \
+					"incl %%edx\n\t" \
+					"cmpl $19,%%edx\n\t" \
 					"jle L53\n\t" \
-					  "xorl %%ebx,%%ebx\n\t" \
+					  "xorl %%edx,%%edx\n\t" \
 					"L53:\n\t" \
-					"cmpl %%edx,%%edi\n\t" \
+					"cmpl %%esi,%%edi\n\t" \
 					"jb L54\n\t" \
-					: "=b" (k)
-					: "D" (p), "d" (pe), "a" (p2),
-						"b" (k), "c" (offset)
+					: "=d" (k)
+					: "D" (p), "S" (pe), "a" (p2),
+						"d" (k), "c" (offset)
 				);
 				p = pe + 1;
 			} while (!done);