AArch64 GIsel: legalize lshr operands, even if it is poison
Previously, this caused GlobalISel to emit invalid IR (a gpr32 to gpr64
copy) and fail during verification.
While this shift is not defined (returns poison), it should not crash
codegen, as it may appear inside dead code (for example, a select
instruction), and it is legal IR input, as long as the value is unused.
Discovered while trying to build Julia with LLVM v13:
https://github.com/JuliaLang/julia/pull/42602.
Reviewed By: aemerson
Differential Revision: https://reviews.llvm.org/D114389
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp b/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
index 34c990b..563f2b7 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp
@@ -1920,35 +1920,6 @@
MachineRegisterInfo &MRI = MF.getRegInfo();
switch (I.getOpcode()) {
- case TargetOpcode::G_SHL:
- case TargetOpcode::G_ASHR:
- case TargetOpcode::G_LSHR: {
- // These shifts are legalized to have 64 bit shift amounts because we want
- // to take advantage of the existing imported selection patterns that assume
- // the immediates are s64s. However, if the shifted type is 32 bits and for
- // some reason we receive input GMIR that has an s64 shift amount that's not
- // a G_CONSTANT, insert a truncate so that we can still select the s32
- // register-register variant.
- Register SrcReg = I.getOperand(1).getReg();
- Register ShiftReg = I.getOperand(2).getReg();
- const LLT ShiftTy = MRI.getType(ShiftReg);
- const LLT SrcTy = MRI.getType(SrcReg);
- if (SrcTy.isVector())
- return false;
- assert(!ShiftTy.isVector() && "unexpected vector shift ty");
- if (SrcTy.getSizeInBits() != 32 || ShiftTy.getSizeInBits() != 64)
- return false;
- auto *AmtMI = MRI.getVRegDef(ShiftReg);
- assert(AmtMI && "could not find a vreg definition for shift amount");
- if (AmtMI->getOpcode() != TargetOpcode::G_CONSTANT) {
- // Insert a subregister copy to implement a 64->32 trunc
- auto Trunc = MIB.buildInstr(TargetOpcode::COPY, {SrcTy}, {})
- .addReg(ShiftReg, 0, AArch64::sub_32);
- MRI.setRegBank(Trunc.getReg(0), RBI.getRegBank(AArch64::GPRRegBankID));
- I.getOperand(2).setReg(Trunc.getReg(0));
- }
- return true;
- }
case TargetOpcode::G_STORE: {
bool Changed = contractCrossBankCopyIntoStore(I, MRI);
MachineOperand &SrcOp = I.getOperand(0);
@@ -2950,6 +2921,28 @@
if (Opcode == TargetOpcode::G_SHL &&
MRI.getType(I.getOperand(0).getReg()).isVector())
return selectVectorSHL(I, MRI);
+
+ // These shifts were legalized to have 64 bit shift amounts because we
+ // want to take advantage of the selection patterns that assume the
+ // immediates are s64s, however, selectBinaryOp will assume both operands
+ // will have the same bit size.
+ {
+ Register SrcReg = I.getOperand(1).getReg();
+ Register ShiftReg = I.getOperand(2).getReg();
+ const LLT ShiftTy = MRI.getType(ShiftReg);
+ const LLT SrcTy = MRI.getType(SrcReg);
+ if (!SrcTy.isVector() && SrcTy.getSizeInBits() == 32 &&
+ ShiftTy.getSizeInBits() == 64) {
+ assert(!ShiftTy.isVector() && "unexpected vector shift ty");
+ auto *AmtMI = MRI.getVRegDef(ShiftReg);
+ assert(AmtMI && "could not find a vreg definition for shift amount");
+ // Insert a subregister copy to implement a 64->32 trunc
+ auto Trunc = MIB.buildInstr(TargetOpcode::COPY, {SrcTy}, {})
+ .addReg(ShiftReg, 0, AArch64::sub_32);
+ MRI.setRegBank(Trunc.getReg(0), RBI.getRegBank(AArch64::GPRRegBankID));
+ I.getOperand(2).setReg(Trunc.getReg(0));
+ }
+ }
LLVM_FALLTHROUGH;
case TargetOpcode::G_FADD:
case TargetOpcode::G_FSUB:
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/select-binop.mir b/llvm/test/CodeGen/AArch64/GlobalISel/select-binop.mir
index da05ea0..e56eb58 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/select-binop.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/select-binop.mir
@@ -492,6 +492,32 @@
...
---
+name: shl_s32_64_gpr
+legalized: true
+regBankSelected: true
+
+registers:
+ - { id: 0, class: gpr }
+ - { id: 1, class: gpr }
+ - { id: 2, class: gpr }
+
+body: |
+ bb.0:
+ liveins: $w0, $x1
+
+ ; CHECK-LABEL: name: shl_s32_64_gpr
+ ; CHECK: [[COPY:%[0-9]+]]:gpr32 = COPY $w0
+ ; CHECK: [[COPY1:%[0-9]+]]:gpr64all = COPY $x1
+ ; CHECK: [[COPY2:%[0-9]+]]:gpr32 = COPY [[COPY1]].sub_32
+ ; CHECK: [[LSLVWr:%[0-9]+]]:gpr32 = LSLVWr [[COPY]], [[COPY2]]
+ ; CHECK: $w0 = COPY [[LSLVWr]]
+ %0(s32) = COPY $w0
+ %1(s64) = COPY $x1
+ %2(s32) = G_SHL %0, %1
+ $w0 = COPY %2(s32)
+...
+
+---
# Same as add_s64_gpr, for G_SHL operations.
name: shl_s64_gpr
legalized: true
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/select-scalar-shift-imm.mir b/llvm/test/CodeGen/AArch64/GlobalISel/select-scalar-shift-imm.mir
index 76c3fd5..f6a7f36 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/select-scalar-shift-imm.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/select-scalar-shift-imm.mir
@@ -101,6 +101,50 @@
...
---
+name: ashr_cimm_32_64
+legalized: true
+regBankSelected: true
+body: |
+ bb.1:
+ liveins: $w0
+
+ ; CHECK-LABEL: name: ashr_cimm_32_64
+ ; CHECK: [[COPY:%[0-9]+]]:gpr32 = COPY $w0
+ ; CHECK: [[MOVi64imm:%[0-9]+]]:gpr64 = MOVi64imm -8
+ ; CHECK: [[COPY1:%[0-9]+]]:gpr32 = COPY [[MOVi64imm]]
+ ; CHECK: [[ASRVWr:%[0-9]+]]:gpr32 = ASRVWr [[COPY]], [[COPY1]]
+ ; CHECK: $w0 = COPY [[ASRVWr]]
+ ; CHECK: RET_ReallyLR implicit $w0
+ %0:gpr(s32) = COPY $w0
+ %3:gpr(s64) = G_CONSTANT i64 -8
+ %2:gpr(s32) = G_ASHR %0, %3(s64)
+ $w0 = COPY %2(s32)
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: lshr_cimm_32_64
+legalized: true
+regBankSelected: true
+body: |
+ bb.1:
+ liveins: $w0
+
+ ; CHECK-LABEL: name: lshr_cimm_32_64
+ ; CHECK: [[COPY:%[0-9]+]]:gpr32 = COPY $w0
+ ; CHECK: [[MOVi64imm:%[0-9]+]]:gpr64 = MOVi64imm -8
+ ; CHECK: [[COPY1:%[0-9]+]]:gpr32 = COPY [[MOVi64imm]]
+ ; CHECK: [[LSRVWr:%[0-9]+]]:gpr32 = LSRVWr [[COPY]], [[COPY1]]
+ ; CHECK: $w0 = COPY [[LSRVWr]]
+ ; CHECK: RET_ReallyLR implicit $w0
+ %0:gpr(s32) = COPY $w0
+ %3:gpr(s64) = G_CONSTANT i64 -8
+ %2:gpr(s32) = G_LSHR %0, %3(s64)
+ $w0 = COPY %2(s32)
+ RET_ReallyLR implicit $w0
+
+...
+---
name: ashr_cimm_64
legalized: true
regBankSelected: true
