Revert "gallivm: Emit vector selects."

It caused inumerous regressions (LLVM 3.1) in blending. In particular:

 - lp_test_blend

    type=u8nx16 rgb_func=sub rgb_src_factor=zero rgb_dst_factor=inv_src_color alpha_func=rev_sub alpha_src_factor=one alpha_dst_factor=const_color ...  MISMATCH
     Src:  0  0  0 b5 49 29  0 a2  0 21 de  0 c3 1b ec  0
     Src1: 2d 85 14  0 f8  0 79 a1 99  0 d8  0 59 16  0  0
     Dst:  0 a9 97  0 c0  0 78  0  0 8b aa f0 bd  0 78 f6
     Con: 7d  0 c0  0  0 bb 77  0  0  0 50  0 40 51  0  0
     Res:  0  0  0  0  0 29  0  0  0  0 c8  0 97 1b e3  0
     Ref:  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
    type=u8nx16 rgb_func=max rgb_src_factor=one rgb_dst_factor=inv_const_color alpha_func=min alpha_src_factor=zero alpha_dst_factor=inv_src1_alpha ...  MISMATCH
     Src:  d  0  0 e9  0 37 35 f0 62  0  0 b2 e9 f7  0 5c
     Src1: 8f  0 bf  0 a8  5  0  0 c4  0 d7  7 92  a  0 17
     Dst: cb  0 1e  0  0  0 19 8e  0 4d  0  0  0  0  3 46
     Con: aa 5a 5f 8f  0  0 bc 92  0 88  0  0 b7 8a c0 88
     Res: 44  0 13  0  0  0  7 8e  0 24  0  0  0  0  1 40
     Ref: 44  0 13  0  0 37 35  0 62 24  0  0 e9 f7  1  0

This reverts commit 1e266c7ef01251ecf72347a2ba1d174b035cbe3b.

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_logic.c b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
index cdb7e0ad5445a6c668bdaccaa5dca95990377c57..168bc2629216cc5a63f9f3980da35ac5dc747998 100644 (file)
--- a/src/gallium/auxiliary/gallivm/lp_bld_logic.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_logic.c
@@ -458,10 +458,22 @@ lp_build_select(struct lp_build_context *bld,
       mask = LLVMBuildTrunc(builder, mask, LLVMInt1TypeInContext(lc), "");
       res = LLVMBuildSelect(builder, mask, a, b, "");
    }
-   else if (HAVE_LLVM >= 0x301) {
+   else if (0) {
       /* Generate a vector select.
        *
-       * Only supported on LLVM 3.1 onwards
+       * XXX: Using vector selects would avoid emitting intrinsics, but they aren't
+       * properly supported yet.
+       *
+       * LLVM 3.1 supports it, but it yields buggy code (e.g. lp_blend_test).
+       *
+       * LLVM 3.0 includes experimental support provided the -promote-elements
+       * options is passed to LLVM's command line (e.g., via
+       * llvm::cl::ParseCommandLineOptions), but resulting code quality is much
+       * worse, probably because some optimization passes don't know how to
+       * handle vector selects.
+       *
+       * See also:
+       * - http://lists.cs.uiuc.edu/pipermail/llvmdev/2011-October/043659.html
        */
 
       /* Convert the mask to a vector of booleans.