util,gallivm: Explicitly enable/disable fma attribute.

As suggested by Roland Scheidegger.

Use the same logic as f16c, since fma requires VEX encoding.

But disable FMA on LLVM 3.3 without MCJIT.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>

src/gallium/auxiliary/gallivm/lp_bld_init.c

@@ -420,6 +420,7 @@ lp_build_init(void)
       util_cpu_caps.has_avx = 0;
       util_cpu_caps.has_avx2 = 0;
       util_cpu_caps.has_f16c = 0;
+      util_cpu_caps.has_fma = 0;
    }
 #endif
 
@@ -454,6 +455,7 @@ lp_build_init(void)
       util_cpu_caps.has_avx = 0;
       util_cpu_caps.has_avx2 = 0;
       util_cpu_caps.has_f16c = 0;
+      util_cpu_caps.has_fma = 0;
    }
 
 #ifdef PIPE_ARCH_PPC_64

src/gallium/auxiliary/gallivm/lp_bld_misc.cpp

@@ -570,6 +570,15 @@ lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
     */
    MAttrs.push_back(util_cpu_caps.has_avx  ? "+avx"  : "-avx");
    MAttrs.push_back(util_cpu_caps.has_f16c ? "+f16c" : "-f16c");
+   if (HAVE_LLVM >= 0x0304 || useMCJIT) {
+      MAttrs.push_back(util_cpu_caps.has_fma  ? "+fma"  : "-fma");
+   } else {
+      /*
+       * The old JIT in LLVM 3.3 has a bug encoding llvm.fmuladd.f32 and
+       * llvm.fmuladd.v2f32 intrinsics when FMA is available.
+       */
+      MAttrs.push_back("-fma");
+   }
    MAttrs.push_back(util_cpu_caps.has_avx2 ? "+avx2" : "-avx2");
    /* disable avx512 and all subvariants */
 #if HAVE_LLVM >= 0x0304

src/gallium/auxiliary/util/u_cpu_detect.c

@@ -369,6 +369,7 @@ util_cpu_detect(void)
                                     ((regs2[2] >> 27) & 1) && // OSXSAVE
                                     ((xgetbv() & 6) == 6);    // XMM & YMM
          util_cpu_caps.has_f16c   = ((regs2[2] >> 29) & 1) && util_cpu_caps.has_avx;
+         util_cpu_caps.has_fma    = ((regs2[2] >> 12) & 1) && util_cpu_caps.has_avx;
          util_cpu_caps.has_mmx2   = util_cpu_caps.has_sse; /* SSE cpus supports mmxext too */
 #if defined(PIPE_ARCH_X86_64)
          util_cpu_caps.has_daz = 1;

src/gallium/auxiliary/util/u_cpu_detect.h

@@ -66,6 +66,7 @@ struct util_cpu_caps {
    unsigned has_avx:1;
    unsigned has_avx2:1;
    unsigned has_f16c:1;
+   unsigned has_fma:1;
    unsigned has_3dnow:1;
    unsigned has_3dnow_ext:1;
    unsigned has_xop:1;