AMDGPU: Fix truncstore from v6f32 to v6f16 (llvm#171212)

arsenm · Honey Goyal · commit 3c9c4c64efb8 · 2025-12-09T08:09:50.000Z
The v6bf16 cases work, but that's likely because v6bf16 isn't
currently an MVT.

Fixes: SWDEV-570985
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -338,6 +338,7 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
   setTruncStoreAction(MVT::v3f32, MVT::v3f16, Expand);
   setTruncStoreAction(MVT::v4f32, MVT::v4bf16, Expand);
   setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand);
+  setTruncStoreAction(MVT::v6f32, MVT::v6f16, Expand);
   setTruncStoreAction(MVT::v8f32, MVT::v8bf16, Expand);
   setTruncStoreAction(MVT::v8f32, MVT::v8f16, Expand);
   setTruncStoreAction(MVT::v16f32, MVT::v16bf16, Expand);
diff --git a/llvm/test/CodeGen/AMDGPU/fp_trunc_store_fp32_to_bf16.ll b/llvm/test/CodeGen/AMDGPU/fp_trunc_store_fp32_to_bf16.ll
@@ -119,6 +119,54 @@ entry:
   ret void
 }
 
+define void @v6(<6 x float> %num, ptr addrspace(1) %p) {
+; CHECK-LABEL: v6:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_bfe_u32 v8, v4, 16, 1
+; CHECK-NEXT:    s_movk_i32 s4, 0x7fff
+; CHECK-NEXT:    v_add3_u32 v8, v8, v4, s4
+; CHECK-NEXT:    v_or_b32_e32 v9, 0x400000, v4
+; CHECK-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
+; CHECK-NEXT:    v_cndmask_b32_e32 v4, v8, v9, vcc
+; CHECK-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; CHECK-NEXT:    v_add3_u32 v8, v8, v5, s4
+; CHECK-NEXT:    v_or_b32_e32 v9, 0x400000, v5
+; CHECK-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
+; CHECK-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
+; CHECK-NEXT:    s_mov_b32 s5, 0x7060302
+; CHECK-NEXT:    v_perm_b32 v4, v5, v4, s5
+; CHECK-NEXT:    v_bfe_u32 v5, v2, 16, 1
+; CHECK-NEXT:    v_add3_u32 v5, v5, v2, s4
+; CHECK-NEXT:    v_or_b32_e32 v8, 0x400000, v2
+; CHECK-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
+; CHECK-NEXT:    v_cndmask_b32_e32 v2, v5, v8, vcc
+; CHECK-NEXT:    v_bfe_u32 v5, v3, 16, 1
+; CHECK-NEXT:    v_add3_u32 v5, v5, v3, s4
+; CHECK-NEXT:    v_or_b32_e32 v8, 0x400000, v3
+; CHECK-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
+; CHECK-NEXT:    v_cndmask_b32_e32 v3, v5, v8, vcc
+; CHECK-NEXT:    v_perm_b32 v3, v3, v2, s5
+; CHECK-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; CHECK-NEXT:    v_add3_u32 v2, v2, v0, s4
+; CHECK-NEXT:    v_or_b32_e32 v5, 0x400000, v0
+; CHECK-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
+; CHECK-NEXT:    v_cndmask_b32_e32 v0, v2, v5, vcc
+; CHECK-NEXT:    v_bfe_u32 v2, v1, 16, 1
+; CHECK-NEXT:    v_add3_u32 v2, v2, v1, s4
+; CHECK-NEXT:    v_or_b32_e32 v5, 0x400000, v1
+; CHECK-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
+; CHECK-NEXT:    v_cndmask_b32_e32 v1, v2, v5, vcc
+; CHECK-NEXT:    v_perm_b32 v2, v1, v0, s5
+; CHECK-NEXT:    global_store_dwordx3 v[6:7], v[2:4], off
+; CHECK-NEXT:    s_waitcnt vmcnt(0)
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+entry:
+  %conv = fptrunc <6 x float> %num to <6 x bfloat>
+  store <6 x bfloat> %conv, ptr addrspace(1) %p, align 16
+  ret void
+}
+
 define void @v8(<8 x float> %num, ptr addrspace(1) %p) {
 ; CHECK-LABEL: v8:
 ; CHECK:       ; %bb.0: ; %entry
diff --git a/llvm/test/CodeGen/AMDGPU/trunc-store-f64-to-f16.ll b/llvm/test/CodeGen/AMDGPU/trunc-store-f64-to-f16.ll
@@ -53,3 +53,17 @@ define amdgpu_kernel void @global_truncstore_v16f64_to_v16f16(ptr addrspace(1) %
   store <16 x half> %cvt, ptr addrspace(1) %out
   ret void
 }
+
+; GCN-LABEL: {{^}}global_truncstore_v6f64_to_v6f16:
+define void @global_truncstore_v6f64_to_v6f16(ptr addrspace(1) %ptr, <6 x double> %src) {
+  %trunc = fptrunc <6 x double> %src to <6 x half>
+  store <6 x half> %trunc, ptr addrspace(1) %ptr
+  ret void
+}
+
+; GCN-LABEL: {{^}}global_truncstore_v6f64_to_v6bf16:
+define void @global_truncstore_v6f64_to_v6bf16(ptr addrspace(1) %ptr, <6 x double> %src) {
+  %trunc = fptrunc <6 x double> %src to <6 x bfloat>
+  store <6 x bfloat> %trunc, ptr addrspace(1) %ptr
+  ret void
+}
diff --git a/llvm/test/CodeGen/AMDGPU/trunc-store.ll b/llvm/test/CodeGen/AMDGPU/trunc-store.ll
@@ -442,5 +442,130 @@ define void @truncstore_v6i32_to_v6i16(ptr addrspace(1) %out, <6 x i32> %val) {
   ret void
 }
 
+define void @global_fp_truncstore_v6f32_to_v6bf16(ptr addrspace(1) %ptr, <6 x float> %src) {
+; SI-LABEL: global_fp_truncstore_v6f32_to_v6bf16:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v5
+; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v4
+; SI-NEXT:    v_alignbit_b32 v3, v3, v2, 16
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v7
+; SI-NEXT:    s_mov_b32 s6, 0
+; SI-NEXT:    v_alignbit_b32 v4, v5, v4, 16
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v6
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_mov_b32 s4, s6
+; SI-NEXT:    s_mov_b32 s5, s6
+; SI-NEXT:    v_alignbit_b32 v2, v2, v5, 16
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64 offset:8
+; SI-NEXT:    buffer_store_dwordx2 v[3:4], v[0:1], s[4:7], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
+; SI-NEXT:    s_setpc_b64 s[30:31]
+;
+; VI-LABEL: global_fp_truncstore_v6f32_to_v6bf16:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_bfe_u32 v9, v6, 16, 1
+; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v6
+; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
+; VI-NEXT:    v_or_b32_e32 v8, 0x400000, v6
+; VI-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
+; VI-NEXT:    v_cndmask_b32_e32 v6, v9, v8, vcc
+; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
+; VI-NEXT:    s_movk_i32 s4, 0x7fff
+; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
+; VI-NEXT:    v_add_u32_e32 v9, vcc, s4, v9
+; VI-NEXT:    v_or_b32_e32 v8, 0x400000, v7
+; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
+; VI-NEXT:    v_cndmask_b32_e32 v7, v9, v8, vcc
+; VI-NEXT:    v_bfe_u32 v8, v4, 16, 1
+; VI-NEXT:    v_add_u32_e32 v8, vcc, v8, v4
+; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
+; VI-NEXT:    v_add_u32_e32 v8, vcc, s4, v8
+; VI-NEXT:    v_alignbit_b32 v6, v7, v6, 16
+; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v4
+; VI-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
+; VI-NEXT:    v_cndmask_b32_e32 v4, v8, v7, vcc
+; VI-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; VI-NEXT:    v_add_u32_e32 v8, vcc, v8, v5
+; VI-NEXT:    v_add_u32_e32 v8, vcc, s4, v8
+; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v5
+; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
+; VI-NEXT:    v_cndmask_b32_e32 v5, v8, v7, vcc
+; VI-NEXT:    v_bfe_u32 v7, v2, 16, 1
+; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v2
+; VI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
+; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
+; VI-NEXT:    v_alignbit_b32 v5, v5, v4, 16
+; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v2
+; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
+; VI-NEXT:    v_cndmask_b32_e32 v2, v7, v4, vcc
+; VI-NEXT:    v_bfe_u32 v7, v3, 16, 1
+; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v3
+; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
+; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v3
+; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
+; VI-NEXT:    v_cndmask_b32_e32 v3, v7, v4, vcc
+; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
+; VI-NEXT:    v_alignbit_b32 v4, v3, v2, 16
+; VI-NEXT:    flat_store_dwordx3 v[0:1], v[4:6]
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_setpc_b64 s[30:31]
+  %trunc = fptrunc <6 x float> %src to <6 x bfloat>
+  store <6 x bfloat> %trunc, ptr addrspace(1) %ptr
+  ret void
+}
+
+
+define void @global_fp_truncstore_v6f32_to_v6f16(ptr addrspace(1) %ptr, <6 x float> %src) {
+; SI-LABEL: global_fp_truncstore_v6f32_to_v6f16:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; SI-NEXT:    v_or_b32_e32 v4, v4, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v5, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; SI-NEXT:    s_mov_b32 s6, 0
+; SI-NEXT:    v_or_b32_e32 v3, v2, v3
+; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_mov_b32 s4, s6
+; SI-NEXT:    s_mov_b32 s5, s6
+; SI-NEXT:    v_or_b32_e32 v2, v6, v2
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64 offset:8
+; SI-NEXT:    buffer_store_dwordx2 v[3:4], v[0:1], s[4:7], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
+; SI-NEXT:    s_setpc_b64 s[30:31]
+;
+; VI-LABEL: global_fp_truncstore_v6f32_to_v6f16:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_cvt_f16_f32_sdwa v7, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD
+; VI-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; VI-NEXT:    v_cvt_f16_f32_sdwa v5, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD
+; VI-NEXT:    v_cvt_f16_f32_e32 v8, v4
+; VI-NEXT:    v_cvt_f16_f32_sdwa v9, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD
+; VI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; VI-NEXT:    v_or_b32_e32 v4, v6, v7
+; VI-NEXT:    v_or_b32_e32 v3, v8, v5
+; VI-NEXT:    v_or_b32_e32 v2, v2, v9
+; VI-NEXT:    flat_store_dwordx3 v[0:1], v[2:4]
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    s_setpc_b64 s[30:31]
+  %trunc = fptrunc <6 x float> %src to <6 x half>
+  store <6 x half> %trunc, ptr addrspace(1) %ptr
+  ret void
+}
+
 ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
 ; GCN: {{.*}}
