[GloblalISel] Support lowering <3 x i8> arguments in multiple parts.

Differential Revision: https://reviews.llvm.org/D97086

diff --git a/llvm/lib/CodeGen/GlobalISel/CallLowering.cpp b/llvm/lib/CodeGen/GlobalISel/CallLowering.cpp
index 4ef66fc801ee79390cccd9ce4964eb620dbcc080..c6bfa24519364b3741029434bf9a536fe50bdcb3 100644 (file)
--- a/llvm/lib/CodeGen/GlobalISel/CallLowering.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/CallLowering.cpp
@@ -483,9 +483,6 @@ bool CallLowering::handleAssignments(CCState &CCInfo,
       continue;
     }
 
-    // FIXME: Pack registers if we have more than one.
-    Register ArgReg = Args[i].Regs[0];
-
     EVT OrigVT = EVT::getEVT(Args[i].Ty);
     EVT VAVT = VA.getValVT();
     const LLT OrigTy = getLLTForType(*Args[i].Ty, DL);
@@ -494,10 +491,12 @@ bool CallLowering::handleAssignments(CCState &CCInfo,
     // Expected to be multiple regs for a single incoming arg.
     // There should be Regs.size() ArgLocs per argument.
     unsigned NumArgRegs = Args[i].Regs.size();
-
+    MachineRegisterInfo &MRI = MF.getRegInfo();
     assert((j + (NumArgRegs - 1)) < ArgLocs.size() &&
            "Too many regs for number of args");
     for (unsigned Part = 0; Part < NumArgRegs; ++Part) {
+      Register ArgReg = Args[i].Regs[Part];
+      LLT ArgRegTy = MRI.getType(ArgReg);
       // There should be Regs.size() ArgLocs per argument.
       VA = ArgLocs[j + Part];
       if (VA.isMemLoc()) {
@@ -538,8 +537,7 @@ bool CallLowering::handleAssignments(CCState &CCInfo,
       }
 
       // This ArgLoc covers multiple pieces, so we need to split it.
-      Register NewReg =
-        MIRBuilder.getMRI()->createGenericVirtualRegister(VATy);
+      Register NewReg = MRI.createGenericVirtualRegister(VATy);
       Handler.assignValueToReg(NewReg, VA.getLocReg(), VA);
       // If it's a vector type, we either need to truncate the elements
       // or do an unmerge to get the lower block of elements.
@@ -560,8 +558,10 @@ bool CallLowering::handleAssignments(CCState &CCInfo,
           DstRegs.push_back(
               MIRBuilder.getMRI()->createGenericVirtualRegister(OrigTy));
         MIRBuilder.buildUnmerge(DstRegs, {NewReg});
-      } else {
+      } else if (VATy.getScalarSizeInBits() > ArgRegTy.getScalarSizeInBits()) {
         MIRBuilder.buildTrunc(ArgReg, {NewReg}).getReg(0);
+      } else {
+        MIRBuilder.buildCopy(ArgReg, NewReg);
       }
     }
 

diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/call-lowering-vectors.ll b/llvm/test/CodeGen/AArch64/GlobalISel/call-lowering-vectors.ll
index 2ade4ce1d7df12190610b9737c566bc3ca521e74..17e7995cb9fd231abf5c000e5b6548e20352b1dd 100644 (file)
--- a/llvm/test/CodeGen/AArch64/GlobalISel/call-lowering-vectors.ll
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/call-lowering-vectors.ll
@@ -1,7 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
 ; RUN: llc -mtriple=aarch64-linux-gnu -O0 -stop-after=irtranslator -global-isel -verify-machineinstrs %s -o - 2>&1 | FileCheck %s
 
-define i8 @v1s8_add(<1 x i8> %a0) nounwind {
+define i8 @v1s8_add(<1 x i8> %a0) {
   ; CHECK-LABEL: name: v1s8_add
   ; CHECK: bb.1 (%ir-block.0):
   ; CHECK:   liveins: $d0
@@ -13,3 +13,24 @@ define i8 @v1s8_add(<1 x i8> %a0) nounwind {
   %res = bitcast <1 x i8> %a0 to i8
   ret i8 %res
 }
+
+define i24 @test_v3i8(<3 x i8> %a) {
+  ; CHECK-LABEL: name: test_v3i8
+  ; CHECK: bb.1 (%ir-block.0):
+  ; CHECK:   liveins: $w0, $w1, $w2
+  ; CHECK:   [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+  ; CHECK:   [[COPY1:%[0-9]+]]:_(s32) = COPY [[COPY]](s32)
+  ; CHECK:   [[COPY2:%[0-9]+]]:_(s32) = COPY $w1
+  ; CHECK:   [[COPY3:%[0-9]+]]:_(s32) = COPY [[COPY2]](s32)
+  ; CHECK:   [[COPY4:%[0-9]+]]:_(s32) = COPY $w2
+  ; CHECK:   [[COPY5:%[0-9]+]]:_(s32) = COPY [[COPY4]](s32)
+  ; CHECK:   [[BUILD_VECTOR:%[0-9]+]]:_(<3 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY3]](s32), [[COPY5]](s32)
+  ; CHECK:   [[TRUNC:%[0-9]+]]:_(<3 x s8>) = G_TRUNC [[BUILD_VECTOR]](<3 x s32>)
+  ; CHECK:   [[BITCAST:%[0-9]+]]:_(s24) = G_BITCAST [[TRUNC]](<3 x s8>)
+  ; CHECK:   [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[BITCAST]](s24)
+  ; CHECK:   $w0 = COPY [[ANYEXT]](s32)
+  ; CHECK:   RET_ReallyLR implicit $w0
+  %res = bitcast <3 x i8> %a to i24
+  ret i24 %res
+}
+