[AMDGPU] Disallow negative s_load offsets in isLegalAddressingMode (llvm#91327)

Author: jayfoad

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -1610,6 +1610,16 @@ bool SITargetLowering::isLegalAddressingMode(const DataLayout &DL,
         return false;
     }
 
+    if ((AS == AMDGPUAS::CONSTANT_ADDRESS ||
+         AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT) &&
+        AM.BaseOffs < 0) {
+      // Scalar (non-buffer) loads can only use a negative offset if
+      // soffset+offset is non-negative. Since the compiler can only prove that
+      // in a few special cases, it is safer to claim that negative offsets are
+      // not supported.
+      return false;
+    }
+
     if (AM.Scale == 0) // r + i or just i, depending on HasBaseReg.
       return true;
 

llvm/test/CodeGen/AMDGPU/cgp-addressing-modes-smem.ll

@@ -279,33 +279,19 @@ end:
 }
 
 define amdgpu_cs void @test_sink_smem_offset_neg400(ptr addrspace(4) inreg %ptr, i32 inreg %val) {
-; GFX678-LABEL: test_sink_smem_offset_neg400:
-; GFX678:       ; %bb.0: ; %entry
-; GFX678-NEXT:    s_add_u32 s0, s0, 0xfffffe70
-; GFX678-NEXT:    s_addc_u32 s1, s1, -1
-; GFX678-NEXT:  .LBB5_1: ; %loop
-; GFX678-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX678-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX678-NEXT:    s_load_dword s3, s[0:1], 0x0
-; GFX678-NEXT:    s_add_i32 s2, s2, -1
-; GFX678-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX678-NEXT:    s_cbranch_scc1 .LBB5_1
-; GFX678-NEXT:  ; %bb.2: ; %end
-; GFX678-NEXT:    s_endpgm
-;
-; GFX9-LABEL: test_sink_smem_offset_neg400:
-; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:  .LBB5_1: ; %loop
-; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX9-NEXT:    s_add_i32 s2, s2, -1
-; GFX9-NEXT:    s_add_u32 s4, s0, 0xfffffe70
-; GFX9-NEXT:    s_addc_u32 s5, s1, -1
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0
-; GFX9-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX9-NEXT:    s_cbranch_scc1 .LBB5_1
-; GFX9-NEXT:  ; %bb.2: ; %end
-; GFX9-NEXT:    s_endpgm
+; GFX6789-LABEL: test_sink_smem_offset_neg400:
+; GFX6789:       ; %bb.0: ; %entry
+; GFX6789-NEXT:    s_add_u32 s0, s0, 0xfffffe70
+; GFX6789-NEXT:    s_addc_u32 s1, s1, -1
+; GFX6789-NEXT:  .LBB5_1: ; %loop
+; GFX6789-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX6789-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6789-NEXT:    s_load_dword s3, s[0:1], 0x0
+; GFX6789-NEXT:    s_add_i32 s2, s2, -1
+; GFX6789-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX6789-NEXT:    s_cbranch_scc1 .LBB5_1
+; GFX6789-NEXT:  ; %bb.2: ; %end
+; GFX6789-NEXT:    s_endpgm
 ;
 ; GFX12-LABEL: test_sink_smem_offset_neg400:
 ; GFX12:       ; %bb.0: ; %entry
@@ -337,3 +323,52 @@ loop:
 end:
   ret void
 }
+
+; Same for address space 6, constant 32-bit.
+define amdgpu_cs void @test_sink_smem_offset_neg400_32bit(ptr addrspace(6) inreg %ptr, i32 inreg %val) {
+; GFX6789-LABEL: test_sink_smem_offset_neg400_32bit:
+; GFX6789:       ; %bb.0: ; %entry
+; GFX6789-NEXT:    s_add_i32 s2, s0, 0xfffffe70
+; GFX6789-NEXT:    s_mov_b32 s3, 0
+; GFX6789-NEXT:  .LBB6_1: ; %loop
+; GFX6789-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX6789-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6789-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX6789-NEXT:    s_add_i32 s1, s1, -1
+; GFX6789-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX6789-NEXT:    s_cbranch_scc1 .LBB6_1
+; GFX6789-NEXT:  ; %bb.2: ; %end
+; GFX6789-NEXT:    s_endpgm
+;
+; GFX12-LABEL: test_sink_smem_offset_neg400_32bit:
+; GFX12:       ; %bb.0: ; %entry
+; GFX12-NEXT:    s_add_co_i32 s2, s0, 0xfffffe70
+; GFX12-NEXT:    s_mov_b32 s3, 0
+; GFX12-NEXT:  .LBB6_1: ; %loop
+; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT:    s_wait_kmcnt 0x0
+; GFX12-NEXT:    s_load_b32 s0, s[2:3], 0x0
+; GFX12-NEXT:    s_add_co_i32 s1, s1, -1
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX12-NEXT:    s_cbranch_scc1 .LBB6_1
+; GFX12-NEXT:  ; %bb.2: ; %end
+; GFX12-NEXT:    s_endpgm
+entry:
+  %gep = getelementptr i8, ptr addrspace(6) %ptr, i64 -400
+  br label %loop
+
+loop:
+  %count = phi i32 [ %dec, %loop ], [ %val, %entry ]
+  %dec = sub i32 %count, 1
+  %load = load volatile i32, ptr addrspace(6) %gep
+  %cond = icmp eq i32 %dec, 0
+  br i1 %cond, label %end, label %loop
+
+end:
+  ret void
+}
+
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; GFX678: {{.*}}
+; GFX9: {{.*}}