[PATCH] D33319: AMDGPU: M0 operands to spill/restore opcodes are dead

[Nicolai Hähnle via Phabricator via llvm-commits]

nhaehnle created this revision.
Herald added subscribers: t-tye, tpr, dstuttard, yaxunl, wdng, kzhuravl, qcolombet.

With scalar stores, M0 is clobbered and therefore marked as implicitly
defined. However, it is also dead.

This fixes an assertion when the Greedy Register Allocator decides to
optimize a spill/restore pair away again (via tryHintsRecoloring).


https://reviews.llvm.org/D33319

Files:
  lib/Target/AMDGPU/SIInstrInfo.cpp
  test/CodeGen/AMDGPU/spill-to-smem-m0.ll


Index: test/CodeGen/AMDGPU/spill-to-smem-m0.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/spill-to-smem-m0.ll
@@ -0,0 +1,22 @@
+; RUN: llc -O0 -march=amdgcn -mcpu=fiji -amdgpu-spill-sgpr-to-smem=1 -verify-machineinstrs -stop-before=prologepilog < %s
+
+; Spill to SMEM clobbers M0. Check that the implicit-def dead operand is present
+; in the pseudo instructions.
+
+; CHECK-LABEL: {{^}}spill_sgpr:
+; CHECK: SI_SPILL_S32_SAVE {{.*}}, implicit-def dead %m0
+; CHECK: SI_SPILL_S32_RESTORE {{.*}}, implicit-def dead %m0
+define amdgpu_kernel void @spill_sgpr(i32 addrspace(1)* %out, i32 %in) #0 {
+  %sgpr = call i32  asm sideeffect "; def $0", "=s" () #0
+  %cmp = icmp eq i32 %in, 0
+  br i1 %cmp, label %bb0, label %ret
+
+bb0:
+  call void asm sideeffect "; use $0", "s"(i32 %sgpr) #0
+  br label %ret
+
+ret:
+  ret void
+}
+
+attributes #0 = { nounwind }
Index: lib/Target/AMDGPU/SIInstrInfo.cpp
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.cpp
+++ lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -772,7 +772,7 @@
 
     if (ST.hasScalarStores()) {
       // m0 is used for offset to scalar stores if used to spill.
-      Spill.addReg(AMDGPU::M0, RegState::ImplicitDefine);
+      Spill.addReg(AMDGPU::M0, RegState::ImplicitDefine | RegState::Dead);
     }
 
     return;
@@ -873,7 +873,7 @@
 
     if (ST.hasScalarStores()) {
       // m0 is used for offset to scalar stores if used to spill.
-      Spill.addReg(AMDGPU::M0, RegState::ImplicitDefine);
+      Spill.addReg(AMDGPU::M0, RegState::ImplicitDefine | RegState::Dead);
     }
 
     return;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D33319.99438.patch
Type: text/x-patch
Size: 1666 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170518/d11c294d/attachment.bin>