AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

aco: Don't use vcmpx with DPP.

Browse Source 
V_CMPX+DPP returns 0 with reads from disabled lanes, unlike V_CMP+DPP (RDNA3 ISA doc, 7.7)

Fixes: baab6f18c9 ("aco: Optimize branching sequence during SSA elimination.")
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20537>
This commit is contained in:
Georg Lehmann
2023-01-05 17:20:42 +01:00
committed by Marge Bot
parent 41b0407d5c
commit 4fbcd046ce
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/compiler/aco_ssa_elimination.cpp
+6 -3
View File
@@ -363,6 +363,10 @@ try_optimize_branching_sequence(ssa_elimination_ctx& ctx, Block& block, const in
exec_val->isVOPC() ? get_vcmpx(exec_val->opcode) : aco_opcode::num_opcodes;
const bool vopc = v_cmpx_op != aco_opcode::num_opcodes;
/* V_CMPX+DPP returns 0 with reads from disabled lanes, unlike V_CMP+DPP (RDNA3 ISA doc, 7.7) */
if (vopc && exec_val->isDPP())
return;
/* If s_and_saveexec is used, we'll need to insert a new instruction to save the old exec. */
bool save_original_exec = exec_copy->opcode == and_saveexec;
@@ -452,11 +456,10 @@ try_optimize_branching_sequence(ssa_elimination_ctx& ctx, Block& block, const in
if (vopc) {
/* Add one extra definition for exec and copy the VOP3-specific fields if present. */
if (!vcmpx_exec_only) {
if (exec_val->isSDWA() || exec_val->isDPP()) {
if (exec_val->isSDWA()) {
/* This might work but it needs testing and more code to copy the instruction. */
return;
}
else if (!exec_val->isVOP3()) {
} else if (!exec_val->isVOP3()) {
aco_ptr<Instruction> tmp = std::move(exec_val);
exec_val.reset(create_instruction<VOPC_instruction>(
tmp->opcode, tmp->format, tmp->operands.size(), tmp->definitions.size() + 1));