[llvm] 726446a - AMDGPU: Fix splitting wide f32 s.buffer.load intrinsics

[Matt Arsenault via llvm-commits]

Author: Matt Arsenault
Date: 2020-02-03T12:28:08-08:00
New Revision: 726446a00964f0be5521a5ad2b21752b8571cb9a

URL: https://github.com/llvm/llvm-project/commit/726446a00964f0be5521a5ad2b21752b8571cb9a
DIFF: https://github.com/llvm/llvm-project/commit/726446a00964f0be5521a5ad2b21752b8571cb9a.diff

LOG: AMDGPU: Fix splitting wide f32 s.buffer.load intrinsics

This would witch f32 to i32, and produce an invald concat_vectors from
i32 pieces to an f32 vector.

Added: 
    

Modified: 
    llvm/lib/Target/AMDGPU/SIISelLowering.cpp
    llvm/test/CodeGen/AMDGPU/smrd.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index ee38bfc71c82..ea40c89ce639 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -5669,7 +5669,7 @@ SDValue SITargetLowering::lowerSBuffer(EVT VT, SDLoc DL, SDValue Rsrc,
 
   if (NumElts == 8 || NumElts == 16) {
     NumLoads = NumElts / 4;
-    LoadVT = MVT::v4i32;
+    LoadVT = MVT::getVectorVT(LoadVT.getScalarType(), 4);
   }
 
   SDVTList VTList = DAG.getVTList({LoadVT, MVT::Glue});
@@ -5695,7 +5695,7 @@ SDValue SITargetLowering::lowerSBuffer(EVT VT, SDLoc DL, SDValue Rsrc,
                                         LoadVT, MMO, DAG));
   }
 
-  if (VT == MVT::v8i32 || VT == MVT::v16i32)
+  if (NumElts == 8 || NumElts == 16)
     return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Loads);
 
   return Loads[0];

diff  --git a/llvm/test/CodeGen/AMDGPU/smrd.ll b/llvm/test/CodeGen/AMDGPU/smrd.ll
index 31c459406ccd..834d53e6d7f7 100644
--- a/llvm/test/CodeGen/AMDGPU/smrd.ll
+++ b/llvm/test/CodeGen/AMDGPU/smrd.ll
@@ -256,6 +256,22 @@ main_body:
   ret void
 }
 
+; dwordx8 s.buffer.load
+; GCN-LABEL: {{^}}s_buffer_load_dwordx8_v8f32:
+; VIGFX9_10: s_buffer_load_dwordx8 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x80
+; SICI: s_buffer_load_dwordx8 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x20
+define amdgpu_ps void @s_buffer_load_dwordx8_v8f32(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
+main_body:
+  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
+  %s.buffer = call <8 x float> @llvm.amdgcn.s.buffer.load.v8f32(<4 x i32> %tmp22, i32 128, i32 0)
+  %s.buffer.0 = extractelement <8 x float> %s.buffer, i32 0
+  %s.buffer.1 = extractelement <8 x float> %s.buffer, i32 2
+  %s.buffer.2 = extractelement <8 x float> %s.buffer, i32 5
+  %s.buffer.3 = extractelement <8 x float> %s.buffer, i32 7
+  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %s.buffer.0, float %s.buffer.1, float %s.buffer.2, float %s.buffer.3, i1 true, i1 true) #0
+  ret void
+}
+
 ; dwordx16 s.buffer.load
 ; GCN-LABEL: {{^}}s_buffer_load_dwordx16:
 ; VIGFX9_10: s_buffer_load_dwordx16 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x80
@@ -276,6 +292,21 @@ main_body:
   ret void
 }
 
+; GCN-LABEL: {{^}}s_buffer_load_dwordx16_v16f32:
+; VIGFX9_10: s_buffer_load_dwordx16 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x80
+; SICI: s_buffer_load_dwordx16 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x20
+define amdgpu_ps void @s_buffer_load_dwordx16_v16f32(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
+main_body:
+  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
+  %s.buffer = call <16 x float> @llvm.amdgcn.s.buffer.load.v16f32(<4 x i32> %tmp22, i32 128, i32 0)
+  %s.buffer.0 = extractelement <16 x float> %s.buffer, i32 0
+  %s.buffer.1 = extractelement <16 x float> %s.buffer, i32 3
+  %s.buffer.2 = extractelement <16 x float> %s.buffer, i32 12
+  %s.buffer.3 = extractelement <16 x float> %s.buffer, i32 15
+  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %s.buffer.0, float %s.buffer.1, float %s.buffer.2, float %s.buffer.3, i1 true, i1 true) #0
+  ret void
+}
+
 ; GCN-LABEL: {{^}}smrd_sgpr_offset:
 ; GCN: s_buffer_load_dword s{{[0-9]}}, s[0:3], s4
 define amdgpu_ps float @smrd_sgpr_offset(<4 x i32> inreg %desc, i32 inreg %offset) #0 {