[llvm] [AMDGPU] Fix unaligned SMEM dword immediate offsets (PR #216269)
Carl Ritson via llvm-commits
llvm-commits at lists.llvm.org
Sun Aug 16 19:58:00 PDT 2026
================
@@ -0,0 +1,56 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -disable-separate-const-offset-from-gep=1 < %s | FileCheck --check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -disable-separate-const-offset-from-gep=1 < %s | FileCheck --check-prefix=GFX12 %s
+
+; Model one unrolled f16 component load from an 11-component vector. The vector
+; stride is 22 bytes and the fixed component offset is 2 bytes. Because the load
+; is known to be dword aligned, widenLoad promotes it to an i32 SMEM load.
+; GFX11+ must keep the complete address in sbase rather than encode the 2-byte
+; displacement as an immediate offset of s_load_b32.
+define amdgpu_kernel void @widen_f16_load_with_unaligned_imm_offset(
+; GFX11-LABEL: widen_f16_load_with_unaligned_imm_offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: s_load_b32 s6, s[4:5], 0x10
+; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_mul_i32 s6, s6, 22
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_ashr_i32 s4, s6, 31
+; GFX11-NEXT: s_add_u32 s2, s2, s6
+; GFX11-NEXT: s_addc_u32 s3, s3, s4
+; GFX11-NEXT: s_add_u32 s2, s2, 2
+; GFX11-NEXT: s_addc_u32 s3, s3, 0
+; GFX11-NEXT: s_load_b32 s2, s[2:3], 0x0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
+; GFX11-NEXT: global_store_b16 v0, v1, s[0:1]
+; GFX11-NEXT: s_endpgm
+;
+; GFX12-LABEL: widen_f16_load_with_unaligned_imm_offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_clause 0x1
+; GFX12-NEXT: s_load_b32 s6, s[4:5], 0x10
+; GFX12-NEXT: s_load_b128 s[0:3], s[4:5], 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_mul_i32 s4, s6, 22
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_ashr_i32 s5, s4, 31
+; GFX12-NEXT: s_add_nc_u64 s[2:3], s[2:3], s[4:5]
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_add_nc_u64 s[2:3], s[2:3], 2
+; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
+; GFX12-NEXT: global_store_b16 v0, v1, s[0:1]
+; GFX12-NEXT: s_endpgm
+ ptr addrspace(1) inreg %out, ptr addrspace(4) inreg %base,
+ i32 inreg %index) {
+ %dynamic.offset = mul i32 %index, 22
+ %vector.base = getelementptr i8, ptr addrspace(4) %base, i32 %dynamic.offset
+ %component.ptr = getelementptr i8, ptr addrspace(4) %vector.base, i32 2
+ %value = load half, ptr addrspace(4) %component.ptr, align 4
----------------
perlfu wrote:
`align 4` is very suspicious here, and the whole test hinges on this.
Why would we be doing an `align 4` b16 load, particularly when we know the gep is offset by 2?
https://github.com/llvm/llvm-project/pull/216269
More information about the llvm-commits
mailing list