search for: d26224

...at the current GVN doesn't do. NewGVN is probably less lazy than it should be, but it didn't matter for us so we didn't consider it a priority. If somebody cares and finds a case where NewGVN substantially regresses, speak up. The initial code can be found here https://reviews.llvm.org/D26224 The current patch includes only the "core" GVN algorithm, i.e. the expression framework/the logic to perform congruence finding. Other pieces (e.g. PRE will build on top of that). My rough plan is: * Try to get this reviewed/tested and in tree. * Fix miscompiles/bugs (and performance pro...

...ably less lazy than it > > should be, but it didn't matter for us so we didn't consider it a > > priority. If somebody cares and finds a case where NewGVN > > substantially regresses, speak up. > > > > The initial code can be found here https://reviews.llvm.org/D26224 > > The current patch includes only the "core" GVN algorithm, i.e. the > > expression framework/the logic to perform congruence finding. Other > > pieces (e.g. PRE will build on top of that). > > > > My rough plan is: > > * Try to get this reviewed/test...

...ere are quite a few things to notice: 1) GVN used to be the top pass in the middle-end, in some cases, and pretty much always in the top-3. This is not the case anymore, but it's still a pass where we spend a lot of time. This is being worked on by Daniel Berlin and me) https://reviews.llvm.org/D26224 so there's some hope that will be sorted out (or at least there's a plan for it). 2) For clang, we spend 35% more time inside instcombine, and for game7 instcombine seems to largely dominate the amount of time we spend optimizing IR. I tried to bisect (which is not easy considering the test...

...: >> 1) GVN used to be the top pass in the middle-end, in some cases, and >> pretty much always in the top-3. This is not the case anymore, but >> it's still a pass where we spend a lot of time. This is being worked >> on by Daniel Berlin and me) https://reviews.llvm.org/D26224 so there's >> some hope that will be sorted out (or at least there's a plan for it). >> 2) For clang, we spend 35% more time inside instcombine, and for game7 >> instcombine seems to largely dominate the amount of time we spend >> optimizing IR. I tried to bisect (wh...

...things to notice: > 1) GVN used to be the top pass in the middle-end, in some cases, and > pretty much always in the top-3. This is not the case anymore, but > it's still a pass where we spend a lot of time. This is being worked > on by Daniel Berlin and me) https://reviews.llvm.org/D26224 so there's > some hope that will be sorted out (or at least there's a plan for it). > 2) For clang, we spend 35% more time inside instcombine, and for game7 > instcombine seems to largely dominate the amount of time we spend > optimizing IR. I tried to bisect (which is not easy...

...things to notice: > 1) GVN used to be the top pass in the middle-end, in some cases, and > pretty much always in the top-3. This is not the case anymore, but > it's still a pass where we spend a lot of time. This is being worked > on by Daniel Berlin and me) https://reviews.llvm.org/D26224 so there's > some hope that will be sorted out (or at least there's a plan for it). > 2) For clang, we spend 35% more time inside instcombine, and for game7 > instcombine seems to largely dominate the amount of time we spend > optimizing IR. I tried to bisect (which is not easy...

...things to notice: > 1) GVN used to be the top pass in the middle-end, in some cases, and > pretty much always in the top-3. This is not the case anymore, but > it's still a pass where we spend a lot of time. This is being worked > on by Daniel Berlin and me) https://reviews.llvm.org/D26224 so there's > some hope that will be sorted out (or at least there's a plan for it). > 2) For clang, we spend 35% more time inside instcombine, and for game7 > instcombine seems to largely dominate the amount of time we spend > optimizing IR. I tried to bisect (which is not easy...

...things to notice: > 1) GVN used to be the top pass in the middle-end, in some cases, and > pretty much always in the top-3. This is not the case anymore, but > it's still a pass where we spend a lot of time. This is being worked > on by Daniel Berlin and me) https://reviews.llvm.org/D26224 so there's > some hope that will be sorted out (or at least there's a plan for it). > 2) For clang, we spend 35% more time inside instcombine, and for game7 > instcombine seems to largely dominate the amount of time we spend > optimizing IR. I tried to bisect (which is not easy...

...things to notice: > 1) GVN used to be the top pass in the middle-end, in some cases, and > pretty much always in the top-3. This is not the case anymore, but > it's still a pass where we spend a lot of time. This is being worked > on by Daniel Berlin and me) https://reviews.llvm.org/D26224 so there's > some hope that will be sorted out (or at least there's a plan for it). > 2) For clang, we spend 35% more time inside instcombine, and for game7 > instcombine seems to largely dominate the amount of time we spend > optimizing IR. I tried to bisect (which is not easy...