[llvm] [RISCV][GISel] Add a special case to selectCopy for FPR32<->GPR on RV64. (PR #70526)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Oct 27 17:16:28 PDT 2023
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-risc-v
Author: Craig Topper (topperc)
<details>
<summary>Changes</summary>
copyPhysReg can't copy a FPR32 to a 64-bit GPR since the sizes don't
match. Replace with FMV_W_X/FMV_X_W during GISel while we still have the
size information.
This is stacked on #<!-- -->70525
---
Full diff: https://github.com/llvm/llvm-project/pull/70526.diff
4 Files Affected:
- (modified) llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp (+29-2)
- (modified) llvm/lib/Target/RISCV/RISCVInstrInfo.cpp (+28)
- (added) llvm/test/CodeGen/RISCV/GlobalISel/fpr-gpr-copy-rv32.ll (+19)
- (added) llvm/test/CodeGen/RISCV/GlobalISel/fpr-gpr-copy-rv64.ll (+35)
``````````diff
diff --git a/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp b/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
index a8ce01f703f33fa..a21936ced31e4ec 100644
--- a/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
+++ b/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
@@ -468,8 +468,10 @@ bool RISCVInstructionSelector::selectCopy(MachineInstr &MI,
if (DstReg.isPhysical())
return true;
- const TargetRegisterClass *DstRC = getRegClassForTypeOnBank(
- MRI.getType(DstReg), *RBI.getRegBank(DstReg, MRI, TRI));
+ const RegisterBank &DstRegBank = *RBI.getRegBank(DstReg, MRI, TRI);
+
+ const TargetRegisterClass *DstRC =
+ getRegClassForTypeOnBank(MRI.getType(DstReg), DstRegBank);
assert(DstRC &&
"Register class not available for LLT, register bank combination");
@@ -482,6 +484,31 @@ bool RISCVInstructionSelector::selectCopy(MachineInstr &MI,
return false;
}
+ // RV64 requires special handling for copies between GPR and FPR32.
+ // copyPhysReg considers GPR to be 64 bits.
+ // FIXME: Should we remove the XLen check from copyPhysReg?
+ if (MI.isCopy() && Subtarget->is64Bit()) {
+ Register SrcReg = MI.getOperand(1).getReg();
+ const RegisterBank &SrcRegBank = *RBI.getRegBank(SrcReg, MRI, TRI);
+
+ unsigned DstSize = RBI.getSizeInBits(DstReg, MRI, TRI);
+ unsigned SrcSize = RBI.getSizeInBits(SrcReg, MRI, TRI);
+
+ if (DstSize == 32 && SrcSize == 32 &&
+ SrcRegBank.getID() == RISCV::GPRRegBankID &&
+ DstRegBank.getID() == RISCV::FPRRegBankID) {
+ MI.setDesc(TII.get(RISCV::FMV_W_X));
+ return true;
+ }
+
+ if (DstSize == 32 && SrcSize == 32 &&
+ SrcRegBank.getID() == RISCV::FPRRegBankID &&
+ DstRegBank.getID() == RISCV::GPRRegBankID) {
+ MI.setDesc(TII.get(RISCV::FMV_X_W));
+ return true;
+ }
+ }
+
MI.setDesc(TII.get(RISCV::COPY));
return true;
}
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
index bfe43bae7cb12a5..91575bd0dd174b0 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
@@ -471,6 +471,34 @@ void RISCVInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
return;
}
+ if (STI.getXLen() == 32 && RISCV::FPR32RegClass.contains(DstReg) &&
+ RISCV::GPRRegClass.contains(SrcReg)) {
+ BuildMI(MBB, MBBI, DL, get(RISCV::FMV_W_X), DstReg)
+ .addReg(SrcReg, getKillRegState(KillSrc));
+ return;
+ }
+
+ if (STI.getXLen() == 32 && RISCV::GPRRegClass.contains(DstReg) &&
+ RISCV::FPR32RegClass.contains(SrcReg)) {
+ BuildMI(MBB, MBBI, DL, get(RISCV::FMV_X_W), DstReg)
+ .addReg(SrcReg, getKillRegState(KillSrc));
+ return;
+ }
+
+ if (STI.getXLen() == 64 && RISCV::FPR64RegClass.contains(DstReg) &&
+ RISCV::GPRRegClass.contains(SrcReg)) {
+ BuildMI(MBB, MBBI, DL, get(RISCV::FMV_D_X), DstReg)
+ .addReg(SrcReg, getKillRegState(KillSrc));
+ return;
+ }
+
+ if (STI.getXLen() == 64 && RISCV::GPRRegClass.contains(DstReg) &&
+ RISCV::FPR64RegClass.contains(SrcReg)) {
+ BuildMI(MBB, MBBI, DL, get(RISCV::FMV_X_D), DstReg)
+ .addReg(SrcReg, getKillRegState(KillSrc));
+ return;
+ }
+
// VR->VR copies.
if (RISCV::VRRegClass.contains(DstReg, SrcReg)) {
copyPhysRegVector(MBB, MBBI, DL, DstReg, SrcReg, KillSrc, RISCV::VMV1R_V);
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/fpr-gpr-copy-rv32.ll b/llvm/test/CodeGen/RISCV/GlobalISel/fpr-gpr-copy-rv32.ll
new file mode 100644
index 000000000000000..1757e5550f81aeb
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/fpr-gpr-copy-rv32.ll
@@ -0,0 +1,19 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc -mtriple=riscv32 -global-isel -mattr=+f -target-abi=ilp32 \
+; RUN: -verify-machineinstrs < %s | FileCheck -check-prefix=RV32I %s
+
+; Test copying between FPR32 and GPR on RV32.
+; FIXME: This test should be replaced with a more general calling convention
+; test once we have more FP implemented.
+
+define float @fadd(float %x, float %y) {
+; RV32I-LABEL: fadd:
+; RV32I: # %bb.0:
+; RV32I-NEXT: fmv.w.x fa5, a0
+; RV32I-NEXT: fmv.w.x fa4, a1
+; RV32I-NEXT: fadd.s fa5, fa5, fa4
+; RV32I-NEXT: fmv.x.w a0, fa5
+; RV32I-NEXT: ret
+ %a = fadd float %x, %y
+ ret float %a
+}
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/fpr-gpr-copy-rv64.ll b/llvm/test/CodeGen/RISCV/GlobalISel/fpr-gpr-copy-rv64.ll
new file mode 100644
index 000000000000000..c0b674fcbd3d247
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/fpr-gpr-copy-rv64.ll
@@ -0,0 +1,35 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc -mtriple=riscv64 -global-isel -mattr=+d -target-abi=lp64 \
+; RUN: -verify-machineinstrs < %s | FileCheck -check-prefix=RV32I %s
+
+; Test copying between FPR64 and GPR on RV64.
+; FIXME: This test should be replaced with a more general calling convention
+; test once we have more FP implemented.
+
+define double @fadd(double %x, double %y) {
+; RV32I-LABEL: fadd:
+; RV32I: # %bb.0:
+; RV32I-NEXT: fmv.d.x fa5, a0
+; RV32I-NEXT: fmv.d.x fa4, a1
+; RV32I-NEXT: fadd.d fa5, fa5, fa4
+; RV32I-NEXT: fmv.x.d a0, fa5
+; RV32I-NEXT: ret
+ %a = fadd double %x, %y
+ ret double %a
+}
+
+; Test copying between FPR32 and GPR on RV64.
+; FIXME: This test should be replaced with a more general calling convention
+; test once we have more FP implemented.
+
+define float @fadd_f32(float %x, float %y) {
+; RV32I-LABEL: fadd:
+; RV32I: # %bb.0:
+; RV32I-NEXT: fmv.d.x fa5, a0
+; RV32I-NEXT: fmv.d.x fa4, a1
+; RV32I-NEXT: fadd.d fa5, fa5, fa4
+; RV32I-NEXT: fmv.x.d a0, fa5
+; RV32I-NEXT: ret
+ %a = fadd float %x, %y
+ ret float %a
+}
``````````
</details>
https://github.com/llvm/llvm-project/pull/70526
More information about the llvm-commits
mailing list