diff options
| author | José Fonseca <[email protected]> | 2012-10-10 19:44:49 +0100 |
|---|---|---|
| committer | José Fonseca <[email protected]> | 2012-10-12 18:52:28 +0100 |
| commit | bf2edc776b02a2a63862bf69a23adf666ecfcc57 (patch) | |
| tree | 9190851410f4d7c59ebb0902bff9e4d292332306 | |
| parent | 9ccf91f9ef72f63871b0a8c6b43f7bd452d73b94 (diff) | |
gallivm: Don't use llvm.x86.avx.max/min.ps.256 inadvertently.
Could happen when CPU supports AVX, but LLVM doesn't.
| -rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_arit.c | 8 |
1 files changed, 4 insertions, 4 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.c b/src/gallium/auxiliary/gallivm/lp_bld_arit.c index d23ff0bf996..f6310a8ef30 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_arit.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.c @@ -92,7 +92,7 @@ lp_build_min_simple(struct lp_build_context *bld, intrinsic = "llvm.x86.sse.min.ss"; intr_size = 128; } - else if (type.length <= 4 || !util_cpu_caps.has_avx) { + else if (type.length <= 4 || !util_cpu_caps.has_avx || lp_native_vector_width < 256) { intrinsic = "llvm.x86.sse.min.ps"; intr_size = 128; } @@ -106,7 +106,7 @@ lp_build_min_simple(struct lp_build_context *bld, intrinsic = "llvm.x86.sse2.min.sd"; intr_size = 128; } - else if (type.length == 2 || !util_cpu_caps.has_avx) { + else if (type.length == 2 || !util_cpu_caps.has_avx || lp_native_vector_width < 256) { intrinsic = "llvm.x86.sse2.min.pd"; intr_size = 128; } @@ -182,7 +182,7 @@ lp_build_max_simple(struct lp_build_context *bld, intrinsic = "llvm.x86.sse.max.ss"; intr_size = 128; } - else if (type.length <= 4 || !util_cpu_caps.has_avx) { + else if (type.length <= 4 || !util_cpu_caps.has_avx || lp_native_vector_width < 256) { intrinsic = "llvm.x86.sse.max.ps"; intr_size = 128; } @@ -196,7 +196,7 @@ lp_build_max_simple(struct lp_build_context *bld, intrinsic = "llvm.x86.sse2.max.sd"; intr_size = 128; } - else if (type.length == 2 || !util_cpu_caps.has_avx) { + else if (type.length == 2 || !util_cpu_caps.has_avx || lp_native_vector_width < 256) { intrinsic = "llvm.x86.sse2.max.pd"; intr_size = 128; } |