On 26.10.2016 01:16, Marek Olšák wrote:
From: Marek Olšák <marek.ol...@amd.com>

12.0 and older need the same fix but elsewhere.

Cc: 13.0 <mesa-sta...@lists.freedesktop.org>
---
 src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c 
b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
index 1ee9afb..123ff5d 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c
@@ -484,37 +484,46 @@ static void emit_bfi(const struct lp_build_tgsi_action 
*action,
                                LLVMBuildXor(builder, bfi_args[1], bfi_args[2],
                                             ""), ""), "");
 }

 /* this is ffs in C */
 static void emit_lsb(const struct lp_build_tgsi_action *action,
                     struct lp_build_tgsi_context *bld_base,
                     struct lp_build_emit_data *emit_data)
 {
        struct gallivm_state *gallivm = bld_base->base.gallivm;
+       LLVMBuilderRef builder = gallivm->builder;
        LLVMValueRef args[2] = {
                emit_data->args[0],

                /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
                 * add special code to check for x=0. The reason is that
                 * the LLVM behavior for x=0 is different from what we
-                * need here.
-                *
-                * The hardware already implements the correct behavior.
+                * need here. However, LLVM also assumes that ffs(x) is
+                * in [0, 31], but GLSL expects that ffs(0) = -1, so
+                * a conditional assignment to handle 0 is still required.
                 */
                LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
        };

-       emit_data->output[emit_data->chan] =
+       LLVMValueRef lsb =
                lp_build_intrinsic(gallivm->builder, "llvm.cttz.i32",
                                emit_data->dst_type, args, ARRAY_SIZE(args),
                                LLVMReadNoneAttribute);
+
+       /* TODO: We need an intrinsic to skip this conditional. */

An instruction selection pattern might do the trick as well. Anyway,

Reviewed-by: Nicolai Hähnle <nicolai.haeh...@amd.com>

+       /* Check for zero: */
+       emit_data->output[emit_data->chan] =
+               LLVMBuildSelect(builder,
+                               LLVMBuildICmp(builder, LLVMIntEQ, args[0],
+                                             bld_base->uint_bld.zero, ""),
+                               lp_build_const_int32(gallivm, -1), lsb, "");
 }

 /* Find the last bit set. */
 static void emit_umsb(const struct lp_build_tgsi_action *action,
                      struct lp_build_tgsi_context *bld_base,
                      struct lp_build_emit_data *emit_data)
 {
        struct gallivm_state *gallivm = bld_base->base.gallivm;
        LLVMBuilderRef builder = gallivm->builder;
        LLVMValueRef args[2] = {

_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev

Reply via email to