[PATCH] D84478: AMDGPU: Optimize copies to exec with other insts after exec def
Matt Arsenault via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Thu Jul 23 16:49:13 PDT 2020
arsenm created this revision.
arsenm added reviewers: nhaehnle, rampitec, foad, kerbowa.
Herald added subscribers: hiraditya, t-tye, tpr, dstuttard, yaxunl, wdng, jvesely, kzhuravl.
Herald added a project: LLVM.
It's possible to have terminator instructions after a write to exec,
so skip over them to find it.
https://reviews.llvm.org/D84478
Files:
llvm/lib/Target/AMDGPU/SIOptimizeExecMasking.cpp
llvm/test/CodeGen/AMDGPU/optimize-exec-copies-extra-insts-after-copy.mir
Index: llvm/test/CodeGen/AMDGPU/optimize-exec-copies-extra-insts-after-copy.mir
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AMDGPU/optimize-exec-copies-extra-insts-after-copy.mir
@@ -0,0 +1,52 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs -run-pass=si-optimize-exec-masking -o - %s | FileCheck %s
+
+# Make sure we can still optimize writes to exec when there are
+# additional terminators after the exec write. This can happen with
+# phi users of control flow intrinsics.
+
+---
+name: instructions_after_copy_to_exec
+tracksRegLiveness: true
+body: |
+ ; CHECK-LABEL: name: instructions_after_copy_to_exec
+ ; CHECK: bb.0:
+ ; CHECK: successors: %bb.2(0x40000000), %bb.1(0x40000000)
+ ; CHECK: liveins: $vgpr0
+ ; CHECK: renamable $vgpr1 = V_MOV_B32_e32 0, implicit $exec
+ ; CHECK: renamable $vcc = V_CMP_EQ_U32_e64 0, killed $vgpr0, implicit $exec
+ ; CHECK: $sgpr0_sgpr1 = S_AND_SAVEEXEC_B64 $vcc, implicit-def $exec, implicit-def $scc, implicit $exec
+ ; CHECK: renamable $sgpr0_sgpr1 = S_XOR_B64 $exec, killed renamable $sgpr0_sgpr1, implicit-def dead $scc
+ ; CHECK: renamable $sgpr0_sgpr1 = COPY killed renamable $sgpr0_sgpr1, implicit $exec
+ ; CHECK: S_CBRANCH_EXECZ %bb.2, implicit $exec
+ ; CHECK: bb.1:
+ ; CHECK: successors: %bb.2(0x80000000)
+ ; CHECK: liveins: $sgpr0_sgpr1
+ ; CHECK: S_NOP 0, implicit $sgpr0_sgpr1
+ ; CHECK: bb.2:
+ ; CHECK: liveins: $sgpr0_sgpr1
+ ; CHECK: S_NOP 0, implicit $sgpr0_sgpr1
+ bb.0:
+ liveins: $vgpr0
+
+ renamable $vgpr1 = V_MOV_B32_e32 0, implicit $exec
+ renamable $vcc = V_CMP_EQ_U32_e64 0, killed $vgpr0, implicit $exec
+ renamable $sgpr0_sgpr1 = COPY $exec, implicit-def $exec
+ renamable $sgpr2_sgpr3 = S_AND_B64 renamable $sgpr0_sgpr1, renamable $vcc, implicit-def dead $scc
+ renamable $sgpr0_sgpr1 = S_XOR_B64 renamable $sgpr2_sgpr3, killed renamable $sgpr0_sgpr1, implicit-def dead $scc
+ $exec = S_MOV_B64_term killed renamable $sgpr2_sgpr3
+ renamable $sgpr0_sgpr1 = S_MOV_B64_term killed renamable $sgpr0_sgpr1, implicit $exec
+ S_CBRANCH_EXECZ %bb.2, implicit $exec
+
+ bb.1:
+ liveins: $sgpr0_sgpr1
+
+ S_NOP 0, implicit $sgpr0_sgpr1
+
+ bb.2:
+ liveins: $sgpr0_sgpr1
+
+ S_NOP 0, implicit $sgpr0_sgpr1
+
+...
+
Index: llvm/lib/Target/AMDGPU/SIOptimizeExecMasking.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIOptimizeExecMasking.cpp
+++ llvm/lib/Target/AMDGPU/SIOptimizeExecMasking.cpp
@@ -298,9 +298,20 @@
if (I == E)
continue;
- // TODO: It's possible to see other terminator copies after the exec copy.
- Register CopyToExec = isCopyToExec(*I, ST);
- if (!CopyToExec.isValid())
+ // It's possible to see other terminator copies after the exec copy. This
+ // can happen if control flow pseudos had their outputs used by phis.
+ Register CopyToExec;
+
+ unsigned SearchCount = 0;
+ const unsigned SearchLimit = 5;
+ while (I != E && SearchCount++ < SearchLimit) {
+ CopyToExec = isCopyToExec(*I, ST);
+ if (CopyToExec)
+ break;
+ ++I;
+ }
+
+ if (!CopyToExec)
continue;
// Scan backwards to find the def.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D84478.280294.patch
Type: text/x-patch
Size: 3365 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200723/a06e2276/attachment.bin>
More information about the llvm-commits
mailing list