SeparateConstOffsetFromGEP: Pass address space to isLegalAddressingMode

[arsenm]

Summary

Ping since this seems to not have actually sent to the list when originally posted

[jingyue]

The change itself looks good, but is it based on some WIP? TTI::isLegalAddressingMode doesn't take an address space yet.

test/Transforms/SeparateConstOffsetFromGEP/R600/split-gep-and-gvn-addrspace-addressing-modes.ll

95

Please fix.

[arsenm]

In D10292#184760, @jingyue wrote:

The change itself looks good, but is it based on some WIP? TTI::isLegalAddressingMode doesn't take an address space yet.

Yes, I have another patch that removes the default value from TargetLowering::isLegalAddressingMode and moves it to TTI until the other uses are removed.

[jingyue]

LGTM with that patch submitted.

[arsenm]

r239262

lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp

bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
   // If LowerGEP is enabled, even the extracted constant offset can not match
   // the addressing mode, we can still do optimizations to other lowered parts
   // of variable indices. Therefore, we don't check for addressing modes in that
   // case.
   if (!LowerGEP) {
     TargetTransformInfo &TTI =
         getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
             *GEP->getParent()->getParent());
+    unsigned AddrSpace = GEP->getPointerAddressSpace();
     if (!TTI.isLegalAddressingMode(GEP->getType()->getElementType(),
                                    /*BaseGV=*/nullptr, AccumulativeByteOffset,
-                                   /*HasBaseReg=*/true, /*Scale=*/0)) {
+                                   /*HasBaseReg=*/true, /*Scale=*/0,
+                                   AddrSpace)) {
       return Changed;
     }
   }
 
   // Remove the constant offset in each sequential index. The resultant GEP
   // computes the variadic base.
   // Notice that we don't remove struct field indices here. If LowerGEP is
   // disabled, a structure index is not accumulated and we still use the old

test/Transforms/SeparateConstOffsetFromGEP/R600/lit.local.cfg

+if not 'R600' in config.root.targets:
+    config.unsupported = True
+

test/Transforms/SeparateConstOffsetFromGEP/R600/split-gep-and-gvn-addrspace-addressing-modes.ll

+; RUN: opt -mtriple=amdgcn-- -S -separate-const-offset-from-gep -reassociate-geps-verify-no-dead-code -gvn < %s | FileCheck -check-prefix=IR %s
+
+target datalayout = "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32-p24:64:64-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64"
+
+@array = internal addrspace(2) constant [4096 x [32 x float]] zeroinitializer, align 4
+
+; IR-LABEL: @sum_of_array(
+; IR: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr inbounds [4096 x [32 x float]], [4096 x [32 x float]] addrspace(2)* @array, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
+; IR: getelementptr float, float addrspace(2)* [[BASE_PTR]], i64 1
+; IR: getelementptr float, float addrspace(2)* [[BASE_PTR]], i64 32
+; IR: getelementptr float, float addrspace(2)* [[BASE_PTR]], i64 33
+define void @sum_of_array(i32 %x, i32 %y, float addrspace(1)* nocapture %output) {
+  %tmp = sext i32 %y to i64
+  %tmp1 = sext i32 %x to i64
+  %tmp2 = getelementptr inbounds [4096 x [32 x float]], [4096 x [32 x float]] addrspace(2)* @array, i64 0, i64 %tmp1, i64 %tmp
+  %tmp4 = load float, float addrspace(2)* %tmp2, align 4
+  %tmp5 = fadd float %tmp4, 0.000000e+00
+  %tmp6 = add i32 %y, 1
+  %tmp7 = sext i32 %tmp6 to i64
+  %tmp8 = getelementptr inbounds [4096 x [32 x float]], [4096 x [32 x float]] addrspace(2)* @array, i64 0, i64 %tmp1, i64 %tmp7
+  %tmp10 = load float, float addrspace(2)* %tmp8, align 4
+  %tmp11 = fadd float %tmp5, %tmp10
+  %tmp12 = add i32 %x, 1
+  %tmp13 = sext i32 %tmp12 to i64
+  %tmp14 = getelementptr inbounds [4096 x [32 x float]], [4096 x [32 x float]] addrspace(2)* @array, i64 0, i64 %tmp13, i64 %tmp
+  %tmp16 = load float, float addrspace(2)* %tmp14, align 4
+  %tmp17 = fadd float %tmp11, %tmp16
+  %tmp18 = getelementptr inbounds [4096 x [32 x float]], [4096 x [32 x float]] addrspace(2)* @array, i64 0, i64 %tmp13, i64 %tmp7
+  %tmp20 = load float, float addrspace(2)* %tmp18, align 4
+  %tmp21 = fadd float %tmp17, %tmp20
+  store float %tmp21, float addrspace(1)* %output, align 4
+  ret void
+}
+
+@array2 = internal addrspace(2) constant [4096 x [4 x float]] zeroinitializer, align 4
+
+; Some of the indices go over the maximum mubuf offset, so don't split them.
+
+; IR-LABEL: @sum_of_array_over_max_mubuf_offset(
+; IR: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
+; IR: getelementptr float, float addrspace(2)* [[BASE_PTR]], i64 255
+; IR: add i32 %x, 256
+; IR: getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
+; IR: getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
+define void @sum_of_array_over_max_mubuf_offset(i32 %x, i32 %y, float addrspace(1)* nocapture %output) {
+  %tmp = sext i32 %y to i64
+  %tmp1 = sext i32 %x to i64
+  %tmp2 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %tmp1, i64 %tmp
+  %tmp4 = load float, float addrspace(2)* %tmp2, align 4
+  %tmp5 = fadd float %tmp4, 0.000000e+00
+  %tmp6 = add i32 %y, 255
+  %tmp7 = sext i32 %tmp6 to i64
+  %tmp8 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %tmp1, i64 %tmp7
+  %tmp10 = load float, float addrspace(2)* %tmp8, align 4
+  %tmp11 = fadd float %tmp5, %tmp10
+  %tmp12 = add i32 %x, 256
+  %tmp13 = sext i32 %tmp12 to i64
+  %tmp14 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %tmp13, i64 %tmp
+  %tmp16 = load float, float addrspace(2)* %tmp14, align 4
+  %tmp17 = fadd float %tmp11, %tmp16
+  %tmp18 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(2)* @array2, i64 0, i64 %tmp13, i64 %tmp7
+  %tmp20 = load float, float addrspace(2)* %tmp18, align 4
+  %tmp21 = fadd float %tmp17, %tmp20
+  store float %tmp21, float addrspace(1)* %output, align 4
+  ret void
+}
+
+
+@lds_array = internal addrspace(3) global [4096 x [4 x float]] undef, align 4
+
+; DS instructions have a larger immediate offset, so make sure these are OK.
+; IR-LABEL: @sum_of_lds_array_over_max_mubuf_offset(
+; IR: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(3)* @lds_array, i32 0, i32 %{{[a-zA-Z0-9]+}}, i32 %{{[a-zA-Z0-9]+}}
+; IR: getelementptr float, float addrspace(3)* [[BASE_PTR]], i32 255
+; IR: getelementptr float, float addrspace(3)* [[BASE_PTR]], i32 16128
+; IR: getelementptr float, float addrspace(3)* [[BASE_PTR]], i32 16383
+define void @sum_of_lds_array_over_max_mubuf_offset(i32 %x, i32 %y, float addrspace(1)* nocapture %output) {
+  %tmp2 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(3)* @lds_array, i32 0, i32 %x, i32 %y
+  %tmp4 = load float, float addrspace(3)* %tmp2, align 4
+  %tmp5 = fadd float %tmp4, 0.000000e+00
+  %tmp6 = add i32 %y, 255
+  %tmp8 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(3)* @lds_array, i32 0, i32 %x, i32 %tmp6
+  %tmp10 = load float, float addrspace(3)* %tmp8, align 4
+  %tmp11 = fadd float %tmp5, %tmp10
+  %tmp12 = add i32 %x, 4032
+  %tmp14 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(3)* @lds_array, i32 0, i32 %tmp12, i32 %y
+  %tmp16 = load float, float addrspace(3)* %tmp14, align 4
+  %tmp17 = fadd float %tmp11, %tmp16
+  %tmp18 = getelementptr inbounds [4096 x [4 x float]], [4096 x [4 x float]] addrspace(3)* @lds_array, i32 0, i32 %tmp12, i32 %tmp6
+  %tmp20 = load float, float addrspace(3)* %tmp18, align 4
+  %tmp21 = fadd float %tmp17, %tmp20
+  store float %tmp21, float addrspace(1)* %output, align 4
+  ret void
+}
 No newline at end of file