[llvm-branch-commits] [llvm] [AMDGPU] Take the VGPR and AVGPR maximum pressure from one point (PR #236175)
Valery Pykhtin via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Oct 28 04:13:55 PDT 2026
================
@@ -186,10 +186,23 @@ struct GCNRegPressure {
unsigned DynamicVGPRBlockSize);
};
+/// \returns the maximum of \p P1 and \p P2 for each register file. The VGPR and
+/// AVGPR components are queried as one sum, so both come from the operand with
+/// the larger sum. The AGPR maximum stays separate also for a unified VGPR
+/// file: ArchVGPRs and AGPRs use separate ranges of it, so the unified pressure
+/// is the sum of both peaks even when they are at different points.
inline GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2) {
GCNRegPressure Res;
for (unsigned I = 0; I < GCNRegPressure::ValueArraySize; ++I)
Res.Value[I] = std::max(P1.Value[I], P2.Value[I]);
+ for (unsigned Base : {0u, unsigned(GCNRegPressure::TOTAL_KINDS)}) {
+ unsigned V = Base + GCNRegPressure::VGPR;
+ unsigned AV = Base + GCNRegPressure::AVGPR;
+ const GCNRegPressure &From =
+ P1.Value[V] + P1.Value[AV] >= P2.Value[V] + P2.Value[AV] ? P1 : P2;
----------------
vpykhtin wrote:
This changes the semantics of this max, we should come up with new function name that clearly states the result.
https://github.com/llvm/llvm-project/pull/236175
More information about the llvm-branch-commits
mailing list