Mesa (master): aco: ignore blocked registers when checking edges in get_reg_impl()

GitLab Mirror gitlab-mirror at kemper.freedesktop.org
Wed Jun 24 20:54:17 UTC 2020


Module: Mesa
Branch: master
Commit: 7cad27831d27b668f592a7b8fb2b09b6f443cd8f
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=7cad27831d27b668f592a7b8fb2b09b6f443cd8f

Author: Rhys Perry <pendingchaos02 at gmail.com>
Date:   Wed Jun 24 13:23:12 2020 +0100

aco: ignore blocked registers when checking edges in get_reg_impl()

If the only two registers available are consecutive and used by killed
operands, both of them will be blocked and fail the edge check.

Totals from 903 (0.66% of 135946) affected shaders:
VGPRs: 30892 -> 30884 (-0.03%)
CodeSize: 1584468 -> 1584044 (-0.03%); split: -0.05%, +0.02%
MaxWaves: 14374 -> 14378 (+0.03%)
Instrs: 306482 -> 306399 (-0.03%); split: -0.06%, +0.03%

Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
Reviewed-by: Daniel Schürmann <daniel at schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5626>

---

 src/amd/compiler/aco_register_allocation.cpp | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/src/amd/compiler/aco_register_allocation.cpp b/src/amd/compiler/aco_register_allocation.cpp
index 9af9bdfe235..a59a9af7add 100644
--- a/src/amd/compiler/aco_register_allocation.cpp
+++ b/src/amd/compiler/aco_register_allocation.cpp
@@ -185,6 +185,13 @@ public:
       return false;
    }
 
+   bool is_empty_or_blocked(PhysReg start) {
+      if (regs[start] == 0xF0000000) {
+         return subdword_regs[start][start.byte()] + 1 <= 1;
+      }
+      return regs[start] + 1 <= 1;
+   }
+
    void clear(PhysReg start, RegClass rc) {
       if (rc.is_subdword())
          fill_subdword(start, rc.bytes(), 0);
@@ -947,9 +954,11 @@ std::pair<PhysReg, bool> get_reg_impl(ra_ctx& ctx,
    unsigned reg_hi = lb + size - 1;
    for (reg_lo = lb, reg_hi = lb + size - 1; reg_hi < ub; reg_lo += stride, reg_hi += stride) {
       /* first check the edges: this is what we have to fix to allow for num_moves > size */
-      if (reg_lo > lb && reg_file[reg_lo] != 0 && reg_file.get_id(PhysReg(reg_lo)) == reg_file.get_id(PhysReg(reg_lo).advance(-1)))
+      if (reg_lo > lb && !reg_file.is_empty_or_blocked(PhysReg(reg_lo)) &&
+          reg_file.get_id(PhysReg(reg_lo)) == reg_file.get_id(PhysReg(reg_lo).advance(-1)))
          continue;
-      if (reg_hi < ub - 1 && reg_file[reg_hi] != 0 && reg_file.get_id(PhysReg(reg_hi).advance(3)) == reg_file.get_id(PhysReg(reg_hi).advance(4)))
+      if (reg_hi < ub - 1 && !reg_file.is_empty_or_blocked(PhysReg(reg_hi).advance(3)) &&
+          reg_file.get_id(PhysReg(reg_hi).advance(3)) == reg_file.get_id(PhysReg(reg_hi).advance(4)))
          continue;
 
       /* second, check that we have at most k=num_moves elements in the window



More information about the mesa-commit mailing list