[llvm] AMDGPU: Fix gfx950 ashr pk i8/u8 patterns (PR #130040)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

<details>
<summary>Changes</summary>

The correct version of the pattern got lost in a merge somewhere

---
Full diff: https://github.com/llvm/llvm-project/pull/130040.diff


2 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/VOP3Instructions.td (+4-4) 
- (modified) llvm/test/CodeGen/AMDGPU/v_ashr_pk.ll (+8-15) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/VOP3Instructions.td b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
index b80ad366b2796..a03a6f61ce0cb 100644
--- a/llvm/lib/Target/AMDGPU/VOP3Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
@@ -1547,14 +1547,14 @@ let SubtargetPredicate = HasAshrPkInsts, isReMaterializable = 1 in {
 } // End SubtargetPredicate = HasAshrPkInsts, isReMaterializable = 1
 
 class AshrPkI8Pat<VOP3_Pseudo inst, int lo, int hi>: GCNPat<
-    (i16 (or (i16 (shl (i16 (trunc (i32 (AMDGPUsmed3 (i32 (sra i32:$src1, i32:$src2)), (i32 lo), (i32 hi))))), (i16 8))),
-             (i16 (and (i16 (trunc (i32 (AMDGPUsmed3 (i32 (sra i32:$src0, i32:$src2)), (i32 lo), (i32 hi))))), (i16 255))))),
+    (i32 (or (i32 (shl (i32 (AMDGPUsmed3 (i32 (sra i32:$src1, i32:$src2)), (i32 lo), (i32 hi))), (i32 8))),
+             (i32 (and (i32 (AMDGPUsmed3 (i32 (sra i32:$src0, i32:$src2)), (i32 lo), (i32 hi))), (i32 255))))),
     (inst 0, VSrc_b32:$src0, 0,  VSrc_b32:$src1, 0, VSrc_b32:$src2, 0 )
 >;
 
 class AshrPkU8Pat<VOP3_Pseudo inst, int lo, int hi>: GCNPat<
-    (i16 (or (i16 (shl (i16 (trunc (i32 (AMDGPUsmed3 (i32 (sra i32:$src1, i32:$src2)), (i32 lo), (i32 hi))))), (i16 8))),
-             (i16 (trunc (i32 (AMDGPUsmed3 (i32 (sra i32:$src0, i32:$src2)), (i32 lo), (i32 hi))))))),
+    (i32 (or (i32 (shl (i32 (AMDGPUsmed3 (i32 (sra i32:$src1, i32:$src2)), (i32 lo), (i32 hi))), (i32 8))),
+             (i32 (AMDGPUsmed3 (i32 (sra i32:$src0, i32:$src2)), (i32 lo), (i32 hi))))),
     (inst 0, VSrc_b32:$src0, 0,  VSrc_b32:$src1, 0, VSrc_b32:$src2, 0 )
 >;
 
diff --git a/llvm/test/CodeGen/AMDGPU/v_ashr_pk.ll b/llvm/test/CodeGen/AMDGPU/v_ashr_pk.ll
index b5f43f9f68936..684ab8046563d 100644
--- a/llvm/test/CodeGen/AMDGPU/v_ashr_pk.ll
+++ b/llvm/test/CodeGen/AMDGPU/v_ashr_pk.ll
@@ -5,16 +5,12 @@ define amdgpu_kernel void @v_ashr_pk_i8_i32(ptr addrspace(1) %out, i32 %src0, i3
 ; GFX950:       ; %bb.0:
 ; GFX950-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
 ; GFX950-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x24
-; GFX950-NEXT:    v_mov_b32_e32 v1, 0xffffff80
-; GFX950-NEXT:    v_mov_b32_e32 v2, 0x7f
 ; GFX950-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX950-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX950-NEXT:    s_ashr_i32 s1, s1, s2
-; GFX950-NEXT:    s_ashr_i32 s0, s0, s2
-; GFX950-NEXT:    v_med3_i32 v3, s0, v1, v2
-; GFX950-NEXT:    v_med3_i32 v1, s1, v1, v2
-; GFX950-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; GFX950-NEXT:    v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX950-NEXT:    s_and_b32 s2, s2, 31
+; GFX950-NEXT:    v_mov_b32_e32 v1, s1
+; GFX950-NEXT:    v_mov_b32_e32 v2, s2
+; GFX950-NEXT:    v_ashr_pk_i8_i32 v1, s0, v1, v2
 ; GFX950-NEXT:    global_store_short v0, v1, s[6:7]
 ; GFX950-NEXT:    s_endpgm
   %insert.0 = insertelement <2 x i32> poison, i32 %src0, i64 0
@@ -36,15 +32,12 @@ define amdgpu_kernel void @v_ashr_pk_u8_i32(ptr addrspace(1) %out, i32 %src0, i3
 ; GFX950:       ; %bb.0:
 ; GFX950-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
 ; GFX950-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x24
-; GFX950-NEXT:    v_mov_b32_e32 v1, 0xff
 ; GFX950-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX950-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX950-NEXT:    s_ashr_i32 s1, s1, s2
-; GFX950-NEXT:    s_ashr_i32 s0, s0, s2
-; GFX950-NEXT:    v_med3_i32 v2, s0, 0, v1
-; GFX950-NEXT:    v_med3_i32 v1, s1, 0, v1
-; GFX950-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; GFX950-NEXT:    v_or_b32_e32 v1, v2, v1
+; GFX950-NEXT:    s_and_b32 s2, s2, 31
+; GFX950-NEXT:    v_mov_b32_e32 v1, s1
+; GFX950-NEXT:    v_mov_b32_e32 v2, s2
+; GFX950-NEXT:    v_ashr_pk_u8_i32 v1, s0, v1, v2
 ; GFX950-NEXT:    global_store_short v0, v1, s[6:7]
 ; GFX950-NEXT:    s_endpgm
   %insert.0 = insertelement <2 x i32> poison, i32 %src0, i64 0

``````````

</details>


https://github.com/llvm/llvm-project/pull/130040