[llvm] c39fba2 - [AMDGPU] S_SET_GPR_IDX_ON can be passed an immediate index (#125086)

[via llvm-commits]

Author: Jon Chesterfield
Date: 2025-01-30T16:40:12Z
New Revision: c39fba209ce655cf5997878b0bf794da269f008a

URL: https://github.com/llvm/llvm-project/commit/c39fba209ce655cf5997878b0bf794da269f008a
DIFF: https://github.com/llvm/llvm-project/commit/c39fba209ce655cf5997878b0bf794da269f008a.diff

LOG: [AMDGPU] S_SET_GPR_IDX_ON can be passed an immediate index (#125086)

Oversight found by ISel fuzz effort. Assuming the argument is a
register, in some cases it can be an immediate. Tablegen's type for the
instruction is SSrc_b32, i.e. register or immediate fine. Added the
repro from the bug reporter as a test case - prior to this patch llvm
will assert in getReg.

Fixes SWDEV-508589

Added: 
    llvm/test/CodeGen/AMDGPU/copy-to-reg-frameindex.ll

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

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 5727d14ec49e8a..a7ac2a3a4c1944 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -2366,11 +2366,11 @@ bool SIInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
     assert(ST.useVGPRIndexMode());
     Register VecReg = MI.getOperand(0).getReg();
     bool IsUndef = MI.getOperand(1).isUndef();
-    Register Idx = MI.getOperand(3).getReg();
+    MachineOperand Idx = MI.getOperand(3);
     Register SubReg = MI.getOperand(4).getImm();
 
     MachineInstr *SetOn = BuildMI(MBB, MI, DL, get(AMDGPU::S_SET_GPR_IDX_ON))
-                              .addReg(Idx)
+                              .add(Idx)
                               .addImm(AMDGPU::VGPRIndexMode::DST_ENABLE);
     SetOn->getOperand(3).setIsUndef();
 

diff  --git a/llvm/test/CodeGen/AMDGPU/copy-to-reg-frameindex.ll b/llvm/test/CodeGen/AMDGPU/copy-to-reg-frameindex.ll
new file mode 100644
index 00000000000000..d86f497aa5e13d
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/copy-to-reg-frameindex.ll
@@ -0,0 +1,38 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx90a -verify-machineinstrs | FileCheck %s
+
+define amdgpu_kernel void @copy_to_reg_frameindex(ptr addrspace(1) %out, i32 %a, i32 %b, i32 %c) {
+; CHECK-LABEL: copy_to_reg_frameindex:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    ; implicit-def: $vgpr0
+; CHECK-NEXT:  .LBB0_1: ; %loop
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    s_cmp_lt_u32 0, 16
+; CHECK-NEXT:    s_set_gpr_idx_on 0, gpr_idx(DST)
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_set_gpr_idx_off
+; CHECK-NEXT:    s_cbranch_scc1 .LBB0_1
+; CHECK-NEXT:  ; %bb.2: ; %done
+; CHECK-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    global_store_dword v1, v0, s[0:1]
+; CHECK-NEXT:    s_endpgm
+entry:
+  %B = srem i32 %c, -1
+  %alloca = alloca [16 x i32], align 4, addrspace(5)
+  br label %loop
+
+loop:
+  %inc = phi i32 [ 0, %entry ], [ %inc.i, %loop ]
+  %ptr = getelementptr [16 x i32], ptr addrspace(5) %alloca, i32 0, i32 %inc
+  store i32 %inc, ptr addrspace(5) %ptr, align 4
+  %inc.i = add i32 %inc, %B
+  %cnd = icmp uge i32 %inc.i, 16
+  br i1 %cnd, label %done, label %loop
+
+done:
+  %tmp1 = load i32, ptr addrspace(5) %alloca, align 4
+  store i32 %tmp1, ptr addrspace(1) %out, align 4
+  ret void
+}