[llvm] [AMDGPU] Fix checkVALUHazardsHelper distance accounting on branchy CFG (PR #203770)
Shilei Tian via llvm-commits
llvm-commits at lists.llvm.org
Mon Jun 15 07:59:19 PDT 2026
================
@@ -932,65 +932,77 @@ int GCNHazardRecognizer::createsVALUHazard(const MachineInstr &MI) const {
return -1;
}
-int GCNHazardRecognizer::checkVALUHazardsHelper(
- const MachineOperand &Def, const MachineRegisterInfo &MRI) const {
- // Helper to check for the hazard where VMEM instructions that store more
- // than 8 bytes can have their store data overwritten by the next
- // instruction. On gfx940-family the window depends on the producer's
+int GCNHazardRecognizer::checkPreGFX940VALUHazardsHelper(Register Reg) const {
+ // Pre-gfx940 wide stores need a single wait-state bubble before a VALU that
+ // overwrites store data. createsVALUHazard already excludes MUBUF/MTBUF
+ // stores with an SGPR SOFFSET.
+ const SIRegisterInfo *TRI = ST.getRegisterInfo();
+
+ auto IsHazard = [this, TRI, Reg](const MachineInstr &MI) {
+ int DataIdx = createsVALUHazard(MI);
+ return DataIdx >= 0 &&
+ TRI->regsOverlap(MI.getOperand(DataIdx).getReg(), Reg);
+ };
+
+ return std::max(0, 1 - getWaitStatesSince(IsHazard, 1));
+}
+
+int GCNHazardRecognizer::checkGFX940FamilyVALUHazardsHelper(
+ Register Reg) const {
----------------
shiltian wrote:
The format seems off here.
https://github.com/llvm/llvm-project/pull/203770
More information about the llvm-commits
mailing list