Add test cases for PCMPxSTRx cases 0x30 and 0x40. Pertains to #320998.

git-svn-id: svn://svn.valgrind.org/valgrind/trunk@13654

diff --git a/none/tests/amd64/pcmpstr64.c b/none/tests/amd64/pcmpstr64.c
index 9e5320216050159d03fc3854ce9887d171d0259a..d28c611708fee6598bfba5163d6b2d54297600ae 100644 (file)
--- a/none/tests/amd64/pcmpstr64.c
+++ b/none/tests/amd64/pcmpstr64.c
@@ -206,6 +206,7 @@ Bool pcmpXstrX_WRK ( /*OUT*/V128* resV,
       case 0x00:
       case 0x02: case 0x08: case 0x0C: case 0x12: case 0x1A:
       case 0x38: case 0x3A: case 0x44: case 0x46: case 0x4A:
+      case 0x30: case 0x40:
          break;
       default:
          return False;
@@ -1383,6 +1384,166 @@ void istri_46 ( void )
 }
 
 
+//////////////////////////////////////////////////////////
+//                                                      //
+//                       ISTRI_30                       //
+//                                                      //
+//////////////////////////////////////////////////////////
+
+UInt h_pcmpistri_30 ( V128* argL, V128* argR )
+{
+   V128 block[2];
+   memcpy(&block[0], argL, sizeof(V128));
+   memcpy(&block[1], argR, sizeof(V128));
+   ULong res, flags;
+   __asm__ __volatile__(
+      "subq      $1024,  %%rsp"             "\n\t"
+      "movdqu    0(%2),  %%xmm2"            "\n\t"
+      "movdqu    16(%2), %%xmm11"           "\n\t"
+      "pcmpistri $0x30,  %%xmm2, %%xmm11"   "\n\t"
+      "pushfq"                              "\n\t"
+      "popq      %%rdx"                     "\n\t"
+      "movq      %%rcx,  %0"                "\n\t"
+      "movq      %%rdx,  %1"                "\n\t"
+      "addq      $1024,  %%rsp"             "\n\t"
+      : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0])
+      : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory"
+   );
+   return ((flags & 0x8D5) << 16) | (res & 0xFFFF);
+}
+
+UInt s_pcmpistri_30 ( V128* argLU, V128* argRU )
+{
+   V128 resV;
+   UInt resOSZACP, resECX;
+   Bool ok
+      = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU,
+                       zmask_from_V128(argLU),
+                       zmask_from_V128(argRU),
+                       0x30, False/*!isSTRM*/
+        );
+   assert(ok);
+   resECX = resV.uInt[0];
+   return (resOSZACP << 16) | resECX;
+}
+
+void istri_30 ( void )
+{
+   char* wot = "30";
+   UInt(*h)(V128*,V128*) = h_pcmpistri_30;
+   UInt(*s)(V128*,V128*) = s_pcmpistri_30;
+
+   try_istri(wot,h,s, "abcdacbdabcdabcd", "000000000000000a"); 
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000000b"); 
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "00000000000000ab"); 
+   try_istri(wot,h,s, "abcdabc0abcdabcd", "000000000000abcd"); 
+
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); 
+   try_istri(wot,h,s, "0bcdabcdabcdabcd", "000000000000abcd"); 
+   try_istri(wot,h,s, "abcdabcdabcda0cd", "000000000000abcd"); 
+   try_istri(wot,h,s, "abcdabcdabcdab0d", "000000000000abcd"); 
+   try_istri(wot,h,s, "abcdabcdabcdabc0", "000000000000abcd"); 
+
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); 
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000a0cd"); 
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000ab0d"); 
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abc0"); 
+
+   try_istri(wot,h,s, "0000000000000000", "0000000000000000"); 
+   try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); 
+
+   try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000abcd"); 
+   try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000dcba"); 
+   try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000bbbb"); 
+   try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000baba"); 
+
+   try_istri(wot,h,s, "0000abcdabcdabcd", "00000000000baba0"); 
+
+   try_istri(wot,h,s, "0ddc0ffeebadf00d", "00000000cafebabe"); 
+   try_istri(wot,h,s, "0ddc0ffeebadfeed", "00000000cafebabe"); 
+}
+
+
+//////////////////////////////////////////////////////////
+//                                                      //
+//                       ISTRI_40                       //
+//                                                      //
+//////////////////////////////////////////////////////////
+
+UInt h_pcmpistri_40 ( V128* argL, V128* argR )
+{
+   V128 block[2];
+   memcpy(&block[0], argL, sizeof(V128));
+   memcpy(&block[1], argR, sizeof(V128));
+   ULong res, flags;
+   __asm__ __volatile__(
+      "subq      $1024,  %%rsp"             "\n\t"
+      "movdqu    0(%2),  %%xmm2"            "\n\t"
+      "movdqu    16(%2), %%xmm11"           "\n\t"
+      "pcmpistri $0x40,  %%xmm2, %%xmm11"   "\n\t"
+      "pushfq"                              "\n\t"
+      "popq      %%rdx"                     "\n\t"
+      "movq      %%rcx,  %0"                "\n\t"
+      "movq      %%rdx,  %1"                "\n\t"
+      "addq      $1024,  %%rsp"             "\n\t"
+      : /*out*/ "=r"(res), "=r"(flags) : "r"/*in*/(&block[0])
+      : "rcx","rdx","xmm0","xmm2","xmm11","cc","memory"
+   );
+   return ((flags & 0x8D5) << 16) | (res & 0xFFFF);
+}
+
+UInt s_pcmpistri_40 ( V128* argLU, V128* argRU )
+{
+   V128 resV;
+   UInt resOSZACP, resECX;
+   Bool ok
+      = pcmpXstrX_WRK( &resV, &resOSZACP, argLU, argRU,
+                       zmask_from_V128(argLU),
+                       zmask_from_V128(argRU),
+                       0x40, False/*!isSTRM*/
+        );
+   assert(ok);
+   resECX = resV.uInt[0];
+   return (resOSZACP << 16) | resECX;
+}
+
+void istri_40 ( void )
+{
+   char* wot = "40";
+   UInt(*h)(V128*,V128*) = h_pcmpistri_40;
+   UInt(*s)(V128*,V128*) = s_pcmpistri_40;
+
+   try_istri(wot,h,s, "abcdacbdabcdabcd", "000000000000000a"); 
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000000b"); 
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "00000000000000ab"); 
+   try_istri(wot,h,s, "abcdabc0abcdabcd", "000000000000abcd"); 
+
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); 
+   try_istri(wot,h,s, "0bcdabcdabcdabcd", "000000000000abcd"); 
+   try_istri(wot,h,s, "abcdabcdabcda0cd", "000000000000abcd"); 
+   try_istri(wot,h,s, "abcdabcdabcdab0d", "000000000000abcd"); 
+   try_istri(wot,h,s, "abcdabcdabcdabc0", "000000000000abcd"); 
+
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abcd"); 
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000a0cd"); 
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000ab0d"); 
+   try_istri(wot,h,s, "abcdabcdabcdabcd", "000000000000abc0"); 
+
+   try_istri(wot,h,s, "0000000000000000", "0000000000000000"); 
+   try_istri(wot,h,s, "aaaaaaaaaaaaaaaa", "aaaaaaaaaaaaaaaa"); 
+
+   try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000abcd"); 
+   try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000dcba"); 
+   try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000bbbb"); 
+   try_istri(wot,h,s, "0000abcdabcdabcd", "000000000000baba"); 
+
+   try_istri(wot,h,s, "0000abcdabcdabcd", "00000000000baba0"); 
+
+   try_istri(wot,h,s, "0ddc0ffeebadf00d", "00000000cafebabe"); 
+   try_istri(wot,h,s, "0ddc0ffeebadfeed", "00000000cafebabe"); 
+}
+
+
 //////////////////////////////////////////////////////////
 //                                                      //
 //                         main                         //
@@ -1402,5 +1563,7 @@ int main ( void )
    istri_00();
    istri_38();
    istri_46();
+   istri_30();
+   istri_40();
    return 0;
 }

diff --git a/none/tests/amd64/pcmpstr64.stdout.exp b/none/tests/amd64/pcmpstr64.stdout.exp
index e326d16cd4b8b869f95ef1233cd6de17f049ca7b..58acb7fc3634b68351f2996e84eddc1da1e54984 100644 (file)
--- a/none/tests/amd64/pcmpstr64.stdout.exp
+++ b/none/tests/amd64/pcmpstr64.stdout.exp
@@ -288,3 +288,47 @@ istri 46  163887ec041a9b72 fcd75adb9b3e895a -> 08410006 08410006
 istri 46  fc937cbfbf53f8e2 0d136bcb024d3fb7 -> 0881000f 0881000f 
 istri 46  2ca34182c29a82ab 302ebd646775ab54 -> 0081000f 0081000f 
 istri 46  3f2987608c11be6f a9ecb661f8e0a8cb -> 00c10007 00c10007 
+istri 30  abcdacbdabcdabcd 000000000000000a -> 08810000 08810000 
+istri 30  abcdabcdabcdabcd 000000000000000b -> 08810000 08810000 
+istri 30  abcdabcdabcdabcd 00000000000000ab -> 08810000 08810000 
+istri 30  abcdabc0abcdabcd 000000000000abcd -> 00c00010 00c00010 
+istri 30  abcdabcdabcdabcd 000000000000abcd -> 00800010 00800010 
+istri 30  0bcdabcdabcdabcd 000000000000abcd -> 00c00010 00c00010 
+istri 30  abcdabcdabcda0cd 000000000000abcd -> 00c00010 00c00010 
+istri 30  abcdabcdabcdab0d 000000000000abcd -> 00c00010 00c00010 
+istri 30  abcdabcdabcdabc0 000000000000abcd -> 00c00010 00c00010 
+istri 30  abcdabcdabcdabcd 000000000000abcd -> 00800010 00800010 
+istri 30  abcdabcdabcdabcd 000000000000a0cd -> 00810002 00810002 
+istri 30  abcdabcdabcdabcd 000000000000ab0d -> 00810001 00810001 
+istri 30  abcdabcdabcdabcd 000000000000abc0 -> 08810000 08810000 
+istri 30  0000000000000000 0000000000000000 -> 00c00010 00c00010 
+istri 30  aaaaaaaaaaaaaaaa aaaaaaaaaaaaaaaa -> 00000010 00000010 
+istri 30  0000abcdabcdabcd 000000000000abcd -> 00c00010 00c00010 
+istri 30  0000abcdabcdabcd 000000000000dcba -> 00c00010 00c00010 
+istri 30  0000abcdabcdabcd 000000000000bbbb -> 08c10000 08c10000 
+istri 30  0000abcdabcdabcd 000000000000baba -> 08c10000 08c10000 
+istri 30  0000abcdabcdabcd 00000000000baba0 -> 08c10000 08c10000 
+istri 30  0ddc0ffeebadf00d 00000000cafebabe -> 08c10000 08c10000 
+istri 30  0ddc0ffeebadfeed 00000000cafebabe -> 08c10000 08c10000 
+istri 40  abcdacbdabcdabcd 000000000000000a -> 0081000f 0081000f 
+istri 40  abcdabcdabcdabcd 000000000000000b -> 0081000e 0081000e 
+istri 40  abcdabcdabcdabcd 00000000000000ab -> 0081000f 0081000f 
+istri 40  abcdabc0abcdabcd 000000000000abcd -> 08c10007 08c10007 
+istri 40  abcdabcdabcdabcd 000000000000abcd -> 0881000f 0881000f 
+istri 40  0bcdabcdabcdabcd 000000000000abcd -> 08c1000e 08c1000e 
+istri 40  abcdabcdabcda0cd 000000000000abcd -> 08c10001 08c10001 
+istri 40  abcdabcdabcdab0d 000000000000abcd -> 08c10000 08c10000 
+istri 40  abcdabcdabcdabc0 000000000000abcd -> 00c00010 00c00010 
+istri 40  abcdabcdabcdabcd 000000000000abcd -> 0881000f 0881000f 
+istri 40  abcdabcdabcdabcd 000000000000a0cd -> 0881000d 0881000d 
+istri 40  abcdabcdabcdabcd 000000000000ab0d -> 0881000c 0881000c 
+istri 40  abcdabcdabcdabcd 000000000000abc0 -> 00800010 00800010 
+istri 40  0000000000000000 0000000000000000 -> 00c00010 00c00010 
+istri 40  aaaaaaaaaaaaaaaa aaaaaaaaaaaaaaaa -> 0801000f 0801000f 
+istri 40  0000abcdabcdabcd 000000000000abcd -> 08c1000b 08c1000b 
+istri 40  0000abcdabcdabcd 000000000000dcba -> 08c1000b 08c1000b 
+istri 40  0000abcdabcdabcd 000000000000bbbb -> 00c1000a 00c1000a 
+istri 40  0000abcdabcdabcd 000000000000baba -> 00c1000b 00c1000b 
+istri 40  0000abcdabcdabcd 00000000000baba0 -> 00c00010 00c00010 
+istri 40  0ddc0ffeebadf00d 00000000cafebabe -> 00c00010 00c00010 
+istri 40  0ddc0ffeebadfeed 00000000cafebabe -> 00c1000a 00c1000a