- for(attrib = 0; attrib < bld->num_attribs; ++attrib) {
- unsigned mask = bld->mask[attrib];
- unsigned mode = bld->mode[attrib];
- for(chan = 0; chan < NUM_CHANNELS; ++chan) {
- if(mask & (1 << chan)) {
- LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), attrib*NUM_CHANNELS + chan, 0);
- LLVMValueRef a0 = NULL;
- LLVMValueRef dadx = NULL;
- LLVMValueRef dady = NULL;
+ for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
+ /*
+ * always fetch all 4 values for performance/simplicity
+ * Note: we do that here because it seems to generate better
+ * code. It generates a lot of moves initially but less
+ * moves later. As far as I can tell this looks like a
+ * llvm issue, instead of simply reloading the values from
+ * the passed in pointers it if it runs out of registers
+ * it spills/reloads them. Maybe some optimization passes
+ * would help.
+ * Might want to investigate this again later.
+ */
+ const unsigned interp = bld->interp[attrib];
+ LLVMValueRef index = lp_build_const_int32(gallivm,
+ attrib * TGSI_NUM_CHANNELS);
+ LLVMValueRef ptr;
+ LLVMValueRef dadxaos = setup_bld->zero;
+ LLVMValueRef dadyaos = setup_bld->zero;
+ LLVMValueRef a0aos = setup_bld->zero;
+
+ switch (interp) {
+ case LP_INTERP_PERSPECTIVE:
+ /* fall-through */
+
+ case LP_INTERP_LINEAR:
+ ptr = LLVMBuildGEP(builder, dadx_ptr, &index, 1, "");
+ ptr = LLVMBuildBitCast(builder, ptr,
+ LLVMPointerType(setup_bld->vec_type, 0), "");
+ dadxaos = LLVMBuildLoad(builder, ptr, "");
+
+ ptr = LLVMBuildGEP(builder, dady_ptr, &index, 1, "");
+ ptr = LLVMBuildBitCast(builder, ptr,
+ LLVMPointerType(setup_bld->vec_type, 0), "");
+ dadyaos = LLVMBuildLoad(builder, ptr, "");
+
+ attrib_name(dadxaos, attrib, 0, ".dadxaos");
+ attrib_name(dadyaos, attrib, 0, ".dadyaos");
+ /* fall-through */
+
+ case LP_INTERP_CONSTANT:
+ case LP_INTERP_FACING:
+ ptr = LLVMBuildGEP(builder, a0_ptr, &index, 1, "");
+ ptr = LLVMBuildBitCast(builder, ptr,
+ LLVMPointerType(setup_bld->vec_type, 0), "");
+ a0aos = LLVMBuildLoad(builder, ptr, "");
+ attrib_name(a0aos, attrib, 0, ".a0aos");
+ break;