[llvm] 804397d - AMDGPU: Do not bundle inline asm

[Matt Arsenault via llvm-commits]

Author: Matt Arsenault
Date: 2020-06-14T13:24:50-04:00
New Revision: 804397dde69a44ed23b6f9671757567b067a59d2

URL: https://github.com/llvm/llvm-project/commit/804397dde69a44ed23b6f9671757567b067a59d2
DIFF: https://github.com/llvm/llvm-project/commit/804397dde69a44ed23b6f9671757567b067a59d2.diff

LOG: AMDGPU: Do not bundle inline asm

Fixes bug 46285

Added: 
    llvm/test/CodeGen/AMDGPU/no-bundle-asm.ll

Modified: 
    llvm/lib/Target/AMDGPU/SIPostRABundler.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/SIPostRABundler.cpp b/llvm/lib/Target/AMDGPU/SIPostRABundler.cpp
index 3a64d519d105..4c72fa235975 100644
--- a/llvm/lib/Target/AMDGPU/SIPostRABundler.cpp
+++ b/llvm/lib/Target/AMDGPU/SIPostRABundler.cpp
@@ -90,7 +90,7 @@ bool SIPostRABundler::runOnMachineFunction(MachineFunction &MF) {
 
   TRI = MF.getSubtarget<GCNSubtarget>().getRegisterInfo();
   bool Changed = false;
-  const unsigned MemFlags = SIInstrFlags::MTBUF | SIInstrFlags::MUBUF |
+  const uint64_t MemFlags = SIInstrFlags::MTBUF | SIInstrFlags::MUBUF |
                             SIInstrFlags::SMRD | SIInstrFlags::DS |
                             SIInstrFlags::FLAT | SIInstrFlags::MIMG;
 
@@ -101,10 +101,11 @@ bool SIPostRABundler::runOnMachineFunction(MachineFunction &MF) {
     for (auto I = B; I != E; I = Next) {
       Next = std::next(I);
 
-      if (I->isBundled() || !I->mayLoadOrStore() ||
+      const uint64_t IMemFlags = I->getDesc().TSFlags & MemFlags;
+
+      if (IMemFlags == 0 || I->isBundled() || !I->mayLoadOrStore() ||
           B->mayLoad() != I->mayLoad() || B->mayStore() != I->mayStore() ||
-          (B->getDesc().TSFlags & MemFlags) !=
-          (I->getDesc().TSFlags & MemFlags) ||
+          ((B->getDesc().TSFlags & MemFlags) != IMemFlags) ||
           isDependentLoad(*I)) {
 
         if (B != I) {

diff  --git a/llvm/test/CodeGen/AMDGPU/no-bundle-asm.ll b/llvm/test/CodeGen/AMDGPU/no-bundle-asm.ll
new file mode 100644
index 000000000000..8e68c462960e
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/no-bundle-asm.ll
@@ -0,0 +1,35 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck %s
+
+; SIPostRABundler used to incorrectly try to form a bundle containing
+; inline asm.
+
+define amdgpu_kernel void @no_bundle_asm_sideeffect() {
+; CHECK-LABEL: no_bundle_asm_sideeffect:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    ;;#ASMSTART
+; CHECK-NEXT:    s_mov_b32 s0, 42
+; CHECK-NEXT:    ;;#ASMEND
+; CHECK-NEXT:    ;;#ASMSTART
+; CHECK-NEXT:    s_mov_b64 s[0:1], 42
+; CHECK-NEXT:    ;;#ASMEND
+; CHECK-NEXT:    s_endpgm
+  call void asm sideeffect "s_mov_b32 s0, $0", "i,~{s0}"(i32 42)
+  call void asm sideeffect "s_mov_b64 s[0:1], $0", "i,~{s[0:1]}"(i64 42)
+  ret void
+}
+
+define amdgpu_kernel void @no_bundle_asm() {
+; CHECK-LABEL: no_bundle_asm:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    ;;#ASMSTART
+; CHECK-NEXT:    s_mov_b32 s0, 42
+; CHECK-NEXT:    ;;#ASMEND
+; CHECK-NEXT:    ;;#ASMSTART
+; CHECK-NEXT:    s_mov_b64 s[0:1], 42
+; CHECK-NEXT:    ;;#ASMEND
+; CHECK-NEXT:    s_endpgm
+  call void asm "s_mov_b32 s0, $0", "i,~{s0}"(i32 42)
+  call void asm "s_mov_b64 s[0:1], $0", "i,~{s[0:1]}"(i64 42)
+  ret void
+}