1 // This program is a thorough test of the LOADVn/STOREVn shadow memory
2 // operations.
3
4 #include <assert.h>
5 #include <stdlib.h>
6 #include <stdio.h>
7 #include <string.h>
8 #include "memcheck/memcheck.h"
9
10 // All the sizes here are in *bytes*, not bits.
11
12 typedef unsigned char U1;
13 typedef unsigned short U2;
14 typedef unsigned int U4;
15 typedef unsigned long long U8;
16
17 typedef float F4;
18 typedef double F8;
19
20 #define SZB_OF_a 64
21
22 // a[] is the array in which we do our loads and stores.
23 // b[] is another one in which we do some copying.
24 U8 a [SZB_OF_a / 8]; // Type is U8 to ensure it's 8-aligned
25 U8 b [SZB_OF_a / 8]; // same size as a[]
26
27 // XXX: should check the error cases for SET/GET_VBITS also
28
29 // For the byte 'x', build a value of 'size' bytes from that byte, eg:
30 // size 1 --> x
31 // size 2 --> xx
32 // size 4 --> xxxx
33 // size 8 --> xxxxxxxx
34 // where the 0 bits are seen by Memcheck as defined, and the 1 bits are
35 // seen as undefined (ie. the value of each bit matches its V bit, ie. the
36 // resulting value is the same as its metavalue).
37 //
build(int size,U1 byte)38 U8 build(int size, U1 byte)
39 {
40 int i;
41 U8 mask = 0;
42 U8 shres;
43 U8 res = 0xffffffffffffffffULL, res2;
44 (void)VALGRIND_MAKE_MEM_UNDEFINED(&res, 8);
45 assert(1 == size || 2 == size || 4 == size || 8 == size);
46
47 for (i = 0; i < size; i++) {
48 mask <<= 8;
49 mask |= (U8)byte;
50 }
51
52 res &= mask;
53
54 // res is now considered partially defined, but we know exactly what its
55 // value is (it happens to be the same as its metavalue).
56
57 (void)VALGRIND_GET_VBITS(&res, &shres, 8);
58 res2 = res;
59 (void)VALGRIND_MAKE_MEM_DEFINED(&res2, 8); // avoid the 'undefined' warning
60 assert(res2 == shres);
61 return res;
62 }
63
64 // Check that all the bytes in a[x..y-1] have their V byte equal
65 // to either 'expected_byte' or 'expected_byte_alt'.
66 // 'str' and 'offset' are only used for printing an error message if
67 // something goes wrong.
check_all(U4 x,U4 y,U1 expected_byte,U1 expected_byte_alt,char * str,int offset)68 void check_all(U4 x, U4 y, U1 expected_byte, U1 expected_byte_alt,
69 char* str, int offset)
70 {
71 U1 sh[SZB_OF_a]; // Used for getting a[]'s V bits
72 int i;
73
74 (void)VALGRIND_GET_VBITS(a, sh, sizeof(a));
75 for (i = x; i < y; i++) {
76 if ( expected_byte != sh[i] && expected_byte_alt != sh[i] ) {
77 fprintf(stderr, "\n\nFAILURE: %s, offset %d, byte %d -- "
78 "is 0x%x, should be 0x%x or 0x%x\n\n",
79 str, offset, i, sh[i], expected_byte,
80 expected_byte_alt);
81 exit(1);
82 }
83 }
84 }
85
main(void)86 int main(void)
87 {
88 int h, i, j;
89 U1 *undefA, expected_byte, expected_byte_alt;
90
91 if (0 == RUNNING_ON_VALGRIND) {
92 fprintf(stderr,
93 "error: this program only works when run under Valgrind\n");
94 exit(1);
95 }
96
97 // Check a[] has the expected alignment, and that it's not too high in
98 // the address space (which would trigger the slow cases in
99 // LOADVn/STOREVn) on 64-bit platforms).
100 assert( 0 == (long)a % 8);
101 if (sizeof(void*) == 8) {
102 assert( ((U1*)(&a[0])) < ((U1*)(32ULL * 1024*1024*1024)/*32G*/) );
103 }
104
105 // Check basic types have the expected sizes.
106 assert(1 == sizeof(U1));
107 assert(2 == sizeof(U2));
108 assert(4 == sizeof(U4));
109 assert(8 == sizeof(U8));
110
111 // Create an array of values that has all the possible V bit metavalues.
112 // Because 0 represents a defined bit, and because undefA[] is initially
113 // zeroed, we have the nice property that:
114 //
115 // i == undefA[i] == V_bits_of(undefA[i])
116 //
117 // which is useful for testing below.
118 undefA = calloc(1, 256); // one for each possible undefinedness value
119 (void)VALGRIND_MAKE_MEM_UNDEFINED(undefA, 256);
120 for (i = 0; i < 256; i++) {
121 undefA[i] &= i;
122 }
123
124 // This code does a whole lot of reads and writes of a particular size
125 // (NNN = 1, 2, 4 or 8), with varying alignments, of values with
126 // different not/partially/fully defined metavalues, and checks that the
127 // V bits are set in a[] as expected using GET_VBITS.
128 //
129 // 'Ty' is the type of the thing we are copying. It can be an integer
130 // type or an FP type. 'ITy' is the same-sized integer type (and thus
131 // will be the same as 'Ty' if 'ITy' is an integer type). 'ITy' is used
132 // when doing shifting/masking and stuff like that.
133
134 #define DO(NNN, Ty, ITy, isF4) \
135 fprintf(stderr, "-- NNN: %d %s %s ------------------------\n", \
136 NNN, #Ty, #ITy); \
137 /* For all of the alignments from (0..NNN-1), eg. if NNN==4, we do */ \
138 /* alignments of 0, 1, 2, 3. */ \
139 for (h = 0; h < NNN; h++) { \
140 \
141 size_t n = sizeof(a); \
142 size_t nN = n / sizeof(Ty); \
143 Ty* aN = (Ty*)a; \
144 Ty* bN = (Ty*)b; \
145 Ty* aNb = (Ty*)(((U1*)aN) + h); /* set offset from a[] */ \
146 Ty* bNb = (Ty*)(((U1*)bN) + h); /* set offset from b[] */ \
147 \
148 fprintf(stderr, "h = %d (checking %d..%d) ", h, h, (int)(n-NNN+h)); \
149 \
150 /* For each of the 256 possible V byte values... */ \
151 for (j = 0; j < 256; j++) { \
152 /* build the value for i (one of: i, ii, iiii, iiiiiiii) */ \
153 U8 tmp = build(NNN, j); \
154 ITy undefN_ITy = (ITy)tmp; \
155 Ty* undefN_Ty; \
156 { /* This just checks that no overflow occurred when squeezing */ \
157 /* the output of build() into a variable of type 'Ty'. */ \
158 U8 tmpDef = tmp; \
159 ITy undefN_ITyDef = undefN_ITy; \
160 (void)VALGRIND_MAKE_MEM_DEFINED(&tmpDef, 8 ); \
161 (void)VALGRIND_MAKE_MEM_DEFINED(&undefN_ITyDef, NNN); \
162 assert(tmpDef == (U8)undefN_ITyDef); \
163 } \
164 \
165 /* We have to use an array for undefN_Ty -- because if we try to
166 * convert an integer type from build into an FP type with a
167 * straight cast -- eg "float f = (float)i" -- the value gets
168 * converted. With this pointer/array nonsense the exact bit
169 * pattern gets used as an FP value unchanged (that FP value is
170 * undoubtedly nonsense, but that's not a problem here). */ \
171 undefN_Ty = (Ty*)&undefN_ITy; \
172 if (0 == j % 32) fprintf(stderr, "%d...", j); /* progress meter */ \
173 \
174 /* A nasty exception: most machines so far (x86/PPC32/PPC64)
175 * don't have 32-bit floats. So 32-bit floats get cast to 64-bit
176 * floats. Memcheck does a PCast in this case, which means that if
177 * any V bits for the 32-bit float are undefined (ie. 0 != j), all
178 * the V bits in the 64-bit float are undefined. So account for
179 * this when checking. AMD64 typically does FP arithmetic on
180 * SSE, effectively giving it access to 32-bit FP registers. So
181 * in short, for floats, we have to allow either 'j' or 0xFF
182 * as an acceptable result. Sigh. */ \
183 if (isF4) { \
184 expected_byte = j; \
185 expected_byte_alt = 0 != j ? 0xFF : j; \
186 } else { \
187 expected_byte = j; \
188 expected_byte_alt = j; \
189 } \
190 \
191 /* STOREVn. Note that we use the first element of the undefN_Ty
192 * array, as explained above. */ \
193 for (i = 0; i < nN-1; i++) { aNb[i] = undefN_Ty[0]; } \
194 check_all(h, n-NNN+h, expected_byte, expected_byte_alt, \
195 "STOREVn", h); \
196 \
197 /* LOADVn -- by copying the values to one place and then back,
198 * we ensure that LOADVn gets exercised. */ \
199 for (i = 0; i < nN-1; i++) { bNb[i] = aNb[i]; } \
200 for (i = 0; i < nN-1; i++) { aNb[i] = bNb[i]; } \
201 check_all(h, n-NNN+h, expected_byte, expected_byte_alt, "LOADVn", h); \
202 } \
203 fprintf(stderr, "\n"); \
204 }
205
206 // For sizes 4 and 8 we do both integer and floating-point types. The
207 // reason being that on 32-bit machines just using integer types never
208 // exercises LOADV8/STOREV8 -- for integer types these loads/stores get
209 // broken into two 32-bit loads/stores.
210 DO(1, U1, U1, /*isF4*/0);
211 DO(2, U2, U2, /*isF4*/0);
212 DO(4, U4, U4, /*isF4*/0);
213 DO(4, F4, U4, /*isF4*/1);
214 DO(8, U8, U8, /*isF4*/0);
215 DO(8, F8, U8, /*isF4*/0);
216
217 return 0;
218 }
219