gallivm: Detect VSX separately from Altivec
authorVicki Pfau <vi@endrift.com>
Sun, 19 Aug 2018 21:17:01 +0000 (14:17 -0700)
committerRoland Scheidegger <sroland@vmware.com>
Thu, 30 Aug 2018 04:09:49 +0000 (06:09 +0200)
Previously gallivm would attempt to use VSX instructions on all systems
where it detected that Altivec is supported; however, VSX was added to
POWER long after Altivec, causing lots of crashes on older POWER/PPC
hardware, e.g. PPC Macs. By detecting VSX separately from Altivec we can
automatically disable it on hardware that supports Altivec but not VSX

Signed-off-by: Vicki Pfau <vi@endrift.com>
src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
src/gallium/auxiliary/util/u_cpu_detect.c
src/gallium/auxiliary/util/u_cpu_detect.h

index 79dbedbb565bc195b486d42d70afc9248f9eb9fd..fcbdd5050fe5642b0a414b67c856908656284449 100644 (file)
@@ -650,26 +650,11 @@ lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
     * which are fixed in LLVM 4.0.
     *
     * With LLVM 4.0 or higher:
-    * Make sure VSX instructions are ENABLED, unless
-    * a) the entire -mattr option is overridden via GALLIVM_MATTRS, or
-    * b) VSX instructions are explicitly enabled/disabled via GALLIVM_VSX=1 or 0.
+    * Make sure VSX instructions are ENABLED (if supported), unless
+    * VSX instructions are explicitly enabled/disabled via GALLIVM_VSX=1 or 0.
     */
    if (util_cpu_caps.has_altivec) {
-      char *env_mattrs = getenv("GALLIVM_MATTRS");
-      if (env_mattrs) {
-         MAttrs.push_back(env_mattrs);
-      }
-      else {
-         boolean enable_vsx = true;
-         char *env_vsx = getenv("GALLIVM_VSX");
-         if (env_vsx && env_vsx[0] == '0') {
-            enable_vsx = false;
-         }
-         if (enable_vsx)
-            MAttrs.push_back("+vsx");
-         else
-            MAttrs.push_back("-vsx");
-      }
+      MAttrs.push_back(util_cpu_caps.has_vsx ? "+vsx" : "-vsx");
    }
 #endif
 #endif
index 3c6ae4ea1a3cb0edd649a25feab2eddedb443b3c..14003aa76928cab01aebb5828e3953971b1ef192 100644 (file)
@@ -133,6 +133,7 @@ check_os_altivec_support(void)
       signal(SIGILL, SIG_DFL);
    } else {
       boolean enable_altivec = TRUE;    /* Default: enable  if available, and if not overridden */
+      boolean enable_vsx = TRUE;
 #ifdef DEBUG
       /* Disabling Altivec code generation is not the same as disabling VSX code generation,
        * which can be done simply by passing -mattr=-vsx to the LLVM compiler; cf.
@@ -144,6 +145,11 @@ check_os_altivec_support(void)
          enable_altivec = FALSE;
       }
 #endif
+      /* VSX instructions can be explicitly enabled/disabled via GALLIVM_VSX=1 or 0 */
+      char *env_vsx = getenv("GALLIVM_VSX");
+      if (env_vsx && env_vsx[0] == '0') {
+         enable_vsx = FALSE;
+      }
       if (enable_altivec) {
          __lv_powerpc_canjump = 1;
 
@@ -153,8 +159,13 @@ check_os_altivec_support(void)
              :
              : "r" (-1));
 
-         signal(SIGILL, SIG_DFL);
          util_cpu_caps.has_altivec = 1;
+
+         if (enable_vsx) {
+            __asm __volatile("xxland %vs0, %vs0, %vs0");
+            util_cpu_caps.has_vsx = 1;
+         }
+         signal(SIGILL, SIG_DFL);
       } else {
          util_cpu_caps.has_altivec = 0;
       }
@@ -536,6 +547,7 @@ util_cpu_detect(void)
       debug_printf("util_cpu_caps.has_3dnow_ext = %u\n", util_cpu_caps.has_3dnow_ext);
       debug_printf("util_cpu_caps.has_xop = %u\n", util_cpu_caps.has_xop);
       debug_printf("util_cpu_caps.has_altivec = %u\n", util_cpu_caps.has_altivec);
+      debug_printf("util_cpu_caps.has_vsx = %u\n", util_cpu_caps.has_vsx);
       debug_printf("util_cpu_caps.has_neon = %u\n", util_cpu_caps.has_neon);
       debug_printf("util_cpu_caps.has_daz = %u\n", util_cpu_caps.has_daz);
       debug_printf("util_cpu_caps.has_avx512f = %u\n", util_cpu_caps.has_avx512f);
index 7a63d55028945d07824b650e97acef842d8eafc7..19f5567ca7ba107460967c810244c8ccb0768bb5 100644 (file)
@@ -71,6 +71,7 @@ struct util_cpu_caps {
    unsigned has_3dnow_ext:1;
    unsigned has_xop:1;
    unsigned has_altivec:1;
+   unsigned has_vsx:1;
    unsigned has_daz:1;
    unsigned has_neon:1;