[AMDGPU] Verify asm VGPR alignment on gfx1250

[rampitec]

Co-authored-by: Shilei Tian [email protected]

[rampitec]

• [AMDGPU] Prohibit load/store merge if scale_offset is set on gfx1250 #149895
• [AMDGPU] Prevent folding of FI with scale_offset on gfx1250 #149894
• [AMDGPU] MC support for gfx1250 scale_offset modifier #149881
• [AMDGPU] Verify asm VGPR alignment on gfx1250 #149880 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-mc

@llvm/pr-subscribers-backend-amdgpu

Author: Stanislav Mekhanoshin (rampitec)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/149880.diff

3 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp (+34-2)
• (added) llvm/test/MC/AMDGPU/gfx1250_asm_vflat_err.s (+4)
• (modified) llvm/test/MC/AMDGPU/gfx1250_asm_vop2_err.s (+3)

diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index dc8323013bbd4..5bb0e3648d2ec 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -5139,13 +5139,45 @@ bool AMDGPUAsmParser::validateAGPRLdSt(const MCInst &Inst) const {
 
 bool AMDGPUAsmParser::validateVGPRAlign(const MCInst &Inst) const {
   auto FB = getFeatureBits();
+  if (!FB[AMDGPU::FeatureGFX90AInsts] && !FB[AMDGPU::FeatureGFX1250Insts])
+    return true;
+
   unsigned Opc = Inst.getOpcode();
+  const MCRegisterInfo *MRI = getMRI();
   // DS_READ_B96_TR_B6 is the only DS instruction in GFX950, that allows
   // unaligned VGPR. All others only allow even aligned VGPRs.
-  if (!(FB[AMDGPU::FeatureGFX90AInsts]) || Opc == AMDGPU::DS_READ_B96_TR_B6_vi)
+  if (FB[AMDGPU::FeatureGFX90AInsts] && Opc == AMDGPU::DS_READ_B96_TR_B6_vi)
     return true;
 
-  const MCRegisterInfo *MRI = getMRI();
+  if (FB[AMDGPU::FeatureGFX1250Insts]) {
+    switch (Opc) {
+    default:
+      break;
+    case AMDGPU::DS_LOAD_TR6_B96:
+    case AMDGPU::DS_LOAD_TR6_B96_gfx12:
+      // DS_LOAD_TR6_B96 is the only DS instruction in GFX1250, that
+      // allows unaligned VGPR. All others only allow even aligned VGPRs.
+      return true;
+    case AMDGPU::GLOBAL_LOAD_TR6_B96:
+    case AMDGPU::GLOBAL_LOAD_TR6_B96_gfx1250: {
+      // GLOBAL_LOAD_TR6_B96 is the only GLOBAL instruction in GFX1250, that
+      // allows unaligned VGPR for vdst, but other operands still only allow
+      // even aligned VGPRs.
+      int VAddrIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vaddr);
+      if (VAddrIdx != -1) {
+        const MCOperand &Op = Inst.getOperand(VAddrIdx);
+        MCRegister Sub = MRI->getSubReg(Op.getReg(), AMDGPU::sub0);
+        if ((Sub - AMDGPU::VGPR0) & 1)
+          return false;
+      }
+      return true;
+    }
+    case AMDGPU::GLOBAL_LOAD_TR6_B96_SADDR:
+    case AMDGPU::GLOBAL_LOAD_TR6_B96_SADDR_gfx1250:
+      return true;
+    }
+  }
+
   const MCRegisterClass &VGPR32 = MRI->getRegClass(AMDGPU::VGPR_32RegClassID);
   const MCRegisterClass &AGPR32 = MRI->getRegClass(AMDGPU::AGPR_32RegClassID);
   for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
diff --git a/llvm/test/MC/AMDGPU/gfx1250_asm_vflat_err.s b/llvm/test/MC/AMDGPU/gfx1250_asm_vflat_err.s
new file mode 100644
index 0000000000000..86a63bcb4be89
--- /dev/null
+++ b/llvm/test/MC/AMDGPU/gfx1250_asm_vflat_err.s
@@ -0,0 +1,4 @@
+// RUN: not llvm-mc -triple=amdgcn -mcpu=gfx1250 -show-encoding %s 2>&1 | FileCheck --check-prefix=GFX1250-ERR --implicit-check-not=error: --strict-whitespace %s
+
+global_load_b96 v[1:3], v[0:1], off
+// GFX1250-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid register class: vgpr tuples must be 64 bit aligned
diff --git a/llvm/test/MC/AMDGPU/gfx1250_asm_vop2_err.s b/llvm/test/MC/AMDGPU/gfx1250_asm_vop2_err.s
index b68306d60cf8c..f67ad88b5ae83 100644
--- a/llvm/test/MC/AMDGPU/gfx1250_asm_vop2_err.s
+++ b/llvm/test/MC/AMDGPU/gfx1250_asm_vop2_err.s
@@ -1,5 +1,8 @@
 // RUN: not llvm-mc -triple=amdgcn -mcpu=gfx1250 -show-encoding %s 2>&1 | FileCheck --check-prefix=GFX1250-ERR --implicit-check-not=error: --strict-whitespace %s
 
+v_add_f64 v[1:2], v[1:2], v[1:2]
+// GFX1250-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid register class: vgpr tuples must be 64 bit aligned
+
 v_fmaak_f32 v4, v2, v6, 3 row_share:1
 // GFX1250-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: not a valid operand.
 // GFX1250-ERR-NEXT:{{^}}v_fmaak_f32 v4, v2, v6, 3 row_share:1