[AMDGPU] Fix isReallyTriviallyReMaterializable for V_MOV_*

D57708 changed SIInstrInfo::isReallyTriviallyReMaterializable to reject
V_MOVs with extra implicit operands, but it accidentally rejected all
V_MOVs because of their implicit use of exec. Fix it but avoid adding a
moderately expensive call to MI.getDesc().getNumImplicitUses().

In real graphics shaders this changes quite a few vgpr copies into move-
immediates, which is good for avoiding stalls on GFX10.

Differential Revision: https://reviews.llvm.org/D98347

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index f7991d5..b8c1487 100644 (file)
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -116,8 +116,11 @@ bool SIInstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
   case AMDGPU::V_MOV_B64_PSEUDO:
   case AMDGPU::V_ACCVGPR_READ_B32_e64:
   case AMDGPU::V_ACCVGPR_WRITE_B32_e64:
-    // No implicit operands.
-    return MI.getNumOperands() == MI.getDesc().getNumOperands();
+    // No non-standard implicit operands.
+    assert(MI.getDesc().getNumOperands() == 2);
+    assert(MI.getDesc().getNumImplicitDefs() == 0);
+    assert(MI.getDesc().getNumImplicitUses() == 1);
+    return MI.getNumOperands() == 3;
   default:
     return false;
   }

diff --git a/llvm/test/CodeGen/AMDGPU/vgpr-remat.mir b/llvm/test/CodeGen/AMDGPU/vgpr-remat.mir
new file mode 100644 (file)
index 0000000..d43ce1e
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/vgpr-remat.mir
@@ -0,0 +1,44 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=gfx900 -run-pass=simple-register-coalescing -o - %s | FileCheck %s
+
+# Check that we get two move-immediates into %1 and %2, instead of a copy from
+# %1 to %2, because that would introduce a dependency and maybe a stall.
+---
+name: f
+tracksRegLiveness: true
+body: |
+  ; CHECK-LABEL: name: f
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.2(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   liveins: $sgpr0
+  ; CHECK:   undef %4.sub0:vreg_96 = V_MOV_B32_e32 0, implicit $exec
+  ; CHECK:   %4.sub1:vreg_96 = V_MOV_B32_e32 0, implicit $exec
+  ; CHECK:   [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0
+  ; CHECK:   $exec = S_MOV_B64_term [[COPY]]
+  ; CHECK:   S_CBRANCH_EXECZ %bb.2, implicit $exec
+  ; CHECK:   S_BRANCH %bb.1
+  ; CHECK: bb.1:
+  ; CHECK:   successors: %bb.2(0x80000000)
+  ; CHECK:   %4.sub0:vreg_96 = V_MUL_F32_e32 %4.sub0, %4.sub0, implicit $mode, implicit $exec
+  ; CHECK:   %4.sub1:vreg_96 = V_MUL_F32_e32 %4.sub1, %4.sub1, implicit $mode, implicit $exec
+  ; CHECK: bb.2:
+  ; CHECK:   S_ENDPGM 0, implicit %4
+  bb.0:
+    liveins: $sgpr0
+    %0:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    %1:vgpr_32 = COPY %0:vgpr_32
+    %2:vgpr_32 = COPY %0:vgpr_32
+    %3:sreg_64 = COPY $sgpr0
+    $exec = S_MOV_B64_term %3:sreg_64
+    S_CBRANCH_EXECZ %bb.2, implicit $exec
+    S_BRANCH %bb.1
+
+  bb.1:
+    %1:vgpr_32 = V_MUL_F32_e32 %1:vgpr_32, %1:vgpr_32, implicit $mode, implicit $exec
+    %2:vgpr_32 = V_MUL_F32_e32 %2:vgpr_32, %2:vgpr_32, implicit $mode, implicit $exec
+
+  bb.2:
+    undef %4.sub0:vreg_96 = COPY %1:vgpr_32
+    %4.sub1:vreg_96 = COPY %2:vgpr_32
+    S_ENDPGM 0, implicit %4
+...