gallivm: disable f16c when not using AVX
authorRoland Scheidegger <sroland@vmware.com>
Mon, 26 Oct 2015 15:44:47 +0000 (16:44 +0100)
committerRoland Scheidegger <sroland@vmware.com>
Mon, 26 Oct 2015 15:45:49 +0000 (16:45 +0100)
f16c intrinsic can only be emitted when AVX is used. So when we disable AVX
due to forcing 128bit vectors we must not use this intrinsic (depending on
llvm version, this worked previously because llvm used AVX even when we didn't
tell it to, however I've seen this fail with llvm 3.3 since
718249843b915decf8fccec92e466ac1a6219934 which seems to have the side effect
of disabling avx in llvm albeit it only touches sse flags really, but
with ea421e919ae6e72e1319fb205c42a6fb53ca2f82 it's now really disabled).
Albeit being able to use AVX with 128bit vectors also would have its uses, the
code as is really was meant to emulate jit code creation for less capable cpus.
v2: add some (ifdefed out) missing de-featuring options for simulating
less capable cpus.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
src/gallium/auxiliary/gallivm/lp_bld_init.c

index 017d075..96aba73 100644 (file)
@@ -427,6 +427,7 @@ lp_build_init(void)
        */
       util_cpu_caps.has_avx = 0;
       util_cpu_caps.has_avx2 = 0;
+      util_cpu_caps.has_f16c = 0;
    }
 
 #ifdef PIPE_ARCH_PPC_64
@@ -458,7 +459,9 @@ lp_build_init(void)
    util_cpu_caps.has_sse3 = 0;
    util_cpu_caps.has_ssse3 = 0;
    util_cpu_caps.has_sse4_1 = 0;
+   util_cpu_caps.has_sse4_2 = 0;
    util_cpu_caps.has_avx = 0;
+   util_cpu_caps.has_avx2 = 0;
    util_cpu_caps.has_f16c = 0;
 #endif