AMDGPU: Handle invariant when lowering global loads

[arsenm]

Global with invariant should be treated identically to
constant.

[arsenm]

• AMDGPU: Use ConstantPool as source value for DAG lowered kernarg loads #168917 : 2 dependent PRs (#168980 , #170309 )
• AMDGPU: Handle invariant when lowering global loads #168914 👈 (View in Graphite)
• AMDGPU: Add baseline test for split/widen invariant loads #168913
• AMDGPU: Handle invariant loads when considering if a load can be scalar #168787
• Reapply "DAG: Allow select ptr combine for non-0 address spaces" (#168292) #168786
• AMDGPU: Fix treating divergent loads as uniform #168785
• main

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

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

Changes

Global with invariant should be treated identically to
constant.

---

Full diff: https://github.com/llvm/llvm-project/pull/168914.diff

2 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+1-1)
• (modified) llvm/test/CodeGen/AMDGPU/load-global-invariant.ll (+15-8)

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 875278a3b4f97..c681d12ba7499 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -11944,7 +11944,7 @@ SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
       AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT ||
       (AS == AMDGPUAS::GLOBAL_ADDRESS &&
        Subtarget->getScalarizeGlobalBehavior() && Load->isSimple() &&
-       isMemOpHasNoClobberedMemOperand(Load))) {
+       (Load->isInvariant() || isMemOpHasNoClobberedMemOperand(Load)))) {
     if ((!Op->isDivergent() || AMDGPU::isUniformMMO(MMO)) &&
         Alignment >= Align(4) && NumElements < 32) {
       if (MemVT.isPow2VectorType() ||
diff --git a/llvm/test/CodeGen/AMDGPU/load-global-invariant.ll b/llvm/test/CodeGen/AMDGPU/load-global-invariant.ll
index b881edde0f448..6cdadc5bab5fb 100644
--- a/llvm/test/CodeGen/AMDGPU/load-global-invariant.ll
+++ b/llvm/test/CodeGen/AMDGPU/load-global-invariant.ll
@@ -50,15 +50,22 @@ define amdgpu_kernel void @load_global_v3i64(ptr addrspace(1) %dst, ptr addrspac
 define amdgpu_kernel void @load_global_v3i64_invariant(ptr addrspace(1) %dst, ptr addrspace(1) %src) #0 {
 ; CHECK-LABEL: load_global_v3i64_invariant:
 ; CHECK:       ; %bb.0:
-; CHECK-NEXT:    v_mov_b32_e32 v6, 0
-; CHECK-NEXT:    s_load_dwordx2 s[0:1], s[8:9], 0x0
-; CHECK-NEXT:    s_load_dwordx2 s[2:3], s[8:9], 0x8
+; CHECK-NEXT:    v_mov_b32_e32 v4, 0
+; CHECK-NEXT:    s_load_dwordx2 s[4:5], s[8:9], 0x0
+; CHECK-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x8
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    global_load_dwordx4 v[0:3], v6, s[2:3]
-; CHECK-NEXT:    global_load_dwordx2 v[4:5], v6, s[2:3] offset:16
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    global_store_dwordx2 v6, v[4:5], s[0:1] offset:16
-; CHECK-NEXT:    global_store_dwordx4 v6, v[0:3], s[0:1]
+; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[6:7], 0x0
+; CHECK-NEXT:    s_nop 0
+; CHECK-NEXT:    s_load_dwordx2 s[6:7], s[6:7], 0x10
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_mov_b32_e32 v0, s6
+; CHECK-NEXT:    v_mov_b32_e32 v1, s7
+; CHECK-NEXT:    global_store_dwordx2 v4, v[0:1], s[4:5] offset:16
+; CHECK-NEXT:    v_mov_b32_e32 v0, s0
+; CHECK-NEXT:    v_mov_b32_e32 v1, s1
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    v_mov_b32_e32 v3, s3
+; CHECK-NEXT:    global_store_dwordx4 v4, v[0:3], s[4:5]
 ; CHECK-NEXT:    s_endpgm
   %ld = load <3 x i64>, ptr addrspace(1) %src, align 32, !invariant.load !0
   store <3 x i64> %ld, ptr addrspace(1) %dst, align 32

[github-actions]

🐧 Linux x64 Test Results

• 186432 tests passed
• 4864 tests skipped

[shiltian]

and there is no test change with this?

[arsenm]

No, there was in the test from #168913 but it turned out to be restoring a regression from one of the later patches