llvmpipe: (trivial) minimally simplify mask construction
authorRoland Scheidegger <sroland@vmware.com>
Wed, 21 Dec 2016 03:52:16 +0000 (04:52 +0100)
committerRoland Scheidegger <sroland@vmware.com>
Thu, 5 Jan 2017 22:59:38 +0000 (23:59 +0100)
simd instruction sets usually have comparisons for equal, not unequal.
So use a different comparison against the mask itself - which also means
we don't need a all-zero as well as a all-one (for the pxor) reg.

Also add code to avoid scalar expansion of i1 values which we definitely
shouldn't do. There's problems with this though with llvm select
interaction, so it's disabled (basically using llvm select instead of
intrinsics may still produce atrocious code, even in cases where we
figured it should not, albeit I think this could probably be fixed
with some better selection of optimization passes, but I have zero
idea there really).

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
src/gallium/auxiliary/gallivm/lp_bld_logic.c
src/gallium/drivers/llvmpipe/lp_bld_depth.c
src/gallium/drivers/llvmpipe/lp_state_fs.c

index 1a50e82c24166be634e2121f17ea01c94a108817..524917abe278662cd3eafe783201dc6b5474ec63 100644 (file)
@@ -327,6 +327,8 @@ lp_build_select(struct lp_build_context *bld,
        * supported yet for a long time, and LLVM will generate poor code when
        * the mask is not the result of a comparison.
        * Also, llvm 3.7 may miscompile them (bug 94972).
+       * XXX: Even if the instruction was an SExt, this may still produce
+       * terrible code. Try piglit stencil-twoside.
        */
 
       /* Convert the mask to a vector of booleans.
index 0c27c2f89723bd3f63ab2a3127e61ca152718117..d5d5c5a786e95bc5cc0add9272e8f629356a0444 100644 (file)
@@ -963,16 +963,48 @@ lp_build_depth_stencil_test(struct gallivm_state *gallivm,
    if (stencil[0].enabled) {
 
       if (face) {
-         LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
-
-         /* front_facing = face != 0 ? ~0 : 0 */
-         front_facing = LLVMBuildICmp(builder, LLVMIntNE, face, zero, "");
-         front_facing = LLVMBuildSExt(builder, front_facing,
-                                      LLVMIntTypeInContext(gallivm->context,
-                                             s_bld.type.length*s_bld.type.width),
-                                      "");
-         front_facing = LLVMBuildBitCast(builder, front_facing,
-                                         s_bld.int_vec_type, "");
+         if (0) {
+            /*
+             * XXX: the scalar expansion below produces atrocious code
+             * (basically producing a 64bit scalar value, then moving the 2
+             * 32bit pieces separately to simd, plus 4 shuffles, which is
+             * seriously lame). But the scalar-simd transitions are always
+             * tricky, so no big surprise there.
+             * This here would be way better, however llvm has some serious
+             * trouble later using it in the select, probably because it will
+             * recognize the expression as constant and move the simd value
+             * away (out of the loop) - and then it will suddenly try
+             * constructing i1 high-bit masks out of it later...
+             * (Try piglit stencil-twoside.)
+             * Note this is NOT due to using SExt/Trunc, it fails exactly the
+             * same even when using native compare/select.
+             * I cannot reproduce this problem when using stand-alone compiler
+             * though, suggesting some problem with optimization passes...
+             * (With stand-alone compilation, the construction of this mask
+             * value, no matter if the easy 3 instruction here or the complex
+             * 16+ one below, never gets separated from where it's used.)
+             * The scalar code still has the same problem, but the generated
+             * code looks a bit better at least for some reason, even if
+             * mostly by luck (the fundamental issue clearly is the same).
+             */
+            front_facing = lp_build_broadcast(gallivm, s_bld.vec_type, face);
+            /* front_facing = face != 0 ? ~0 : 0 */
+            front_facing = lp_build_compare(gallivm, s_bld.type,
+                                            PIPE_FUNC_NOTEQUAL,
+                                            front_facing, s_bld.zero);
+         } else {
+            LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
+
+            /* front_facing = face != 0 ? ~0 : 0 */
+            front_facing = LLVMBuildICmp(builder, LLVMIntNE, face, zero, "");
+            front_facing = LLVMBuildSExt(builder, front_facing,
+                                         LLVMIntTypeInContext(gallivm->context,
+                                                s_bld.type.length*s_bld.type.width),
+                                         "");
+            front_facing = LLVMBuildBitCast(builder, front_facing,
+                                            s_bld.int_vec_type, "");
+
+         }
       }
 
       s_pass_mask = lp_build_stencil_test(&s_bld, stencil,
index 09108159febad54ef5d38b07f73e93e24bc29ee5..a36389ccc32e706feda813032854f72acde9598c 100644 (file)
@@ -127,7 +127,7 @@ generate_quad_mask(struct gallivm_state *gallivm,
    struct lp_type mask_type;
    LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
    LLVMValueRef bits[16];
-   LLVMValueRef mask;
+   LLVMValueRef mask, bits_vec;
    int shift, i;
 
    /*
@@ -179,15 +179,15 @@ generate_quad_mask(struct gallivm_state *gallivm,
       bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
       bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (j + 5), 0);
    }
-   mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
+   bits_vec = LLVMConstVector(bits, fs_type.length);
+   mask = LLVMBuildAnd(builder, mask, bits_vec, "");
 
    /*
-    * mask = mask != 0 ? ~0 : 0
+    * mask = mask == bits ? ~0 : 0
     */
    mask = lp_build_compare(gallivm,
-                           mask_type, PIPE_FUNC_NOTEQUAL,
-                           mask,
-                           lp_build_const_int_vec(gallivm, mask_type, 0));
+                           mask_type, PIPE_FUNC_EQUAL,
+                           mask, bits_vec);
 
    return mask;
 }
@@ -2476,8 +2476,10 @@ dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
    for (i = 0; i < key->nr_cbufs; ++i) {
       debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
    }
-   if (key->depth.enabled) {
+   if (key->depth.enabled || key->stencil[0].enabled) {
       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
+   }
+   if (key->depth.enabled) {
       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
       debug_printf("depth.writemask = %u\n", key->depth.writemask);
    }