[AMDGPU] Increased vector length for global/constant loads.

Summary: GCN ISA supports instructions that can read 16 consecutive dwords from memory through the scalar data cache;
         loadstoreVectorizer should take advantage of the wider vector length and pack 16/8 elements of dwords/quadwords.

Author: FarhanaAleen

Reviewed By: rampitec

Subscribers: llvm-commits, AMDGPU

Differential Revision: https://reviews.llvm.org/D44179

llvm-svn: 326910

3 files changed, 71 insertions, 1 deletions

diff --git a/llvm/test/CodeGen/AMDGPU/load-constant-f32.ll b/llvm/test/CodeGen/AMDGPU/load-constant-f32.ll
new file mode 100644
index 00000000000..4a07038fc90
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/load-constant-f32.ll
@@ -0,0 +1,37 @@
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
+
+; Tests whether a load chain of 8 constants gets vectorized into a wider load.
+; FUNC-LABEL: {{^}}constant_load_v8f32:
+; GCN: s_load_dwordx8
+; EG: VTX_READ_128
+; EG: VTX_READ_128
+define amdgpu_kernel void @constant_load_v8f32(float addrspace(4)* noalias nocapture readonly %weights, float addrspace(1)* noalias nocapture %out_ptr) {
+entry:
+  %out_ptr.promoted = load float, float addrspace(1)* %out_ptr, align 4
+  %tmp = load float, float addrspace(4)* %weights, align 4
+  %add = fadd float %tmp, %out_ptr.promoted
+  %arrayidx.1 = getelementptr inbounds float, float addrspace(4)* %weights, i64 1
+  %tmp1 = load float, float addrspace(4)* %arrayidx.1, align 4
+  %add.1 = fadd float %tmp1, %add
+  %arrayidx.2 = getelementptr inbounds float, float addrspace(4)* %weights, i64 2
+  %tmp2 = load float, float addrspace(4)* %arrayidx.2, align 4
+  %add.2 = fadd float %tmp2, %add.1
+  %arrayidx.3 = getelementptr inbounds float, float addrspace(4)* %weights, i64 3
+  %tmp3 = load float, float addrspace(4)* %arrayidx.3, align 4
+  %add.3 = fadd float %tmp3, %add.2
+  %arrayidx.4 = getelementptr inbounds float, float addrspace(4)* %weights, i64 4
+  %tmp4 = load float, float addrspace(4)* %arrayidx.4, align 4
+  %add.4 = fadd float %tmp4, %add.3
+  %arrayidx.5 = getelementptr inbounds float, float addrspace(4)* %weights, i64 5
+  %tmp5 = load float, float addrspace(4)* %arrayidx.5, align 4
+  %add.5 = fadd float %tmp5, %add.4
+  %arrayidx.6 = getelementptr inbounds float, float addrspace(4)* %weights, i64 6
+  %tmp6 = load float, float addrspace(4)* %arrayidx.6, align 4
+  %add.6 = fadd float %tmp6, %add.5
+  %arrayidx.7 = getelementptr inbounds float, float addrspace(4)* %weights, i64 7
+  %tmp7 = load float, float addrspace(4)* %arrayidx.7, align 4
+  %add.7 = fadd float %tmp7, %add.6
+  store float %add.7, float addrspace(1)* %out_ptr, align 4
+  ret void
+}
\ No newline at end of file
diff --git a/llvm/test/CodeGen/AMDGPU/load-constant-f64.ll b/llvm/test/CodeGen/AMDGPU/load-constant-f64.ll
index 56e21bc9bd6..dcd420870c2 100644
--- a/llvm/test/CodeGen/AMDGPU/load-constant-f64.ll
+++ b/llvm/test/CodeGen/AMDGPU/load-constant-f64.ll
@@ -13,3 +13,36 @@ define amdgpu_kernel void @constant_load_f64(double addrspace(1)* %out, double a
 }
 
 attributes #0 = { nounwind }
+
+; Tests whether a load-chain of 8 constants of 64bit each gets vectorized into a wider load.
+; FUNC-LABEL: {{^}}constant_load_2v4f64:
+; GCN: s_load_dwordx16
+define amdgpu_kernel void @constant_load_2v4f64(double addrspace(4)* noalias nocapture readonly %weights, double addrspace(1)* noalias nocapture %out_ptr) {
+entry:
+  %out_ptr.promoted = load double, double addrspace(1)* %out_ptr, align 4
+  %tmp = load double, double addrspace(4)* %weights, align 4
+  %add = fadd double %tmp, %out_ptr.promoted
+  %arrayidx.1 = getelementptr inbounds double, double addrspace(4)* %weights, i64 1
+  %tmp1 = load double, double addrspace(4)* %arrayidx.1, align 4
+  %add.1 = fadd double %tmp1, %add
+  %arrayidx.2 = getelementptr inbounds double, double addrspace(4)* %weights, i64 2
+  %tmp2 = load double, double addrspace(4)* %arrayidx.2, align 4
+  %add.2 = fadd double %tmp2, %add.1
+  %arrayidx.3 = getelementptr inbounds double, double addrspace(4)* %weights, i64 3
+  %tmp3 = load double, double addrspace(4)* %arrayidx.3, align 4
+  %add.3 = fadd double %tmp3, %add.2
+  %arrayidx.4 = getelementptr inbounds double, double addrspace(4)* %weights, i64 4
+  %tmp4 = load double, double addrspace(4)* %arrayidx.4, align 4
+  %add.4 = fadd double %tmp4, %add.3
+  %arrayidx.5 = getelementptr inbounds double, double addrspace(4)* %weights, i64 5
+  %tmp5 = load double, double addrspace(4)* %arrayidx.5, align 4
+  %add.5 = fadd double %tmp5, %add.4
+  %arrayidx.6 = getelementptr inbounds double, double addrspace(4)* %weights, i64 6
+  %tmp6 = load double, double addrspace(4)* %arrayidx.6, align 4
+  %add.6 = fadd double %tmp6, %add.5
+  %arrayidx.7 = getelementptr inbounds double, double addrspace(4)* %weights, i64 7
+  %tmp7 = load double, double addrspace(4)* %arrayidx.7, align 4
+  %add.7 = fadd double %tmp7, %add.6
+  store double %add.7, double addrspace(1)* %out_ptr, align 4
+  ret void
+}
diff --git a/llvm/test/CodeGen/AMDGPU/waitcnt-looptest.ll b/llvm/test/CodeGen/AMDGPU/waitcnt-looptest.ll
index 0af9bbcc25e..a941e5fb1f7 100644
--- a/llvm/test/CodeGen/AMDGPU/waitcnt-looptest.ll
+++ b/llvm/test/CodeGen/AMDGPU/waitcnt-looptest.ll
@@ -1,4 +1,4 @@
-; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=fiji -mattr=-flat-for-global | FileCheck --check-prefix=GCN %s
+; RUN: llc < %s -mtriple=amdgcn--amdhsa -mcpu=fiji -mattr=-flat-for-global -amdgpu-load-store-vectorizer=0 | FileCheck --check-prefix=GCN %s
 
 ; Check that the waitcnt insertion algorithm correctly propagates wait counts
 ; from before a loop to the loop header.