[RISCV] Add a command line option to disable cost per use for compressed registers.

[topperc]

I've seen cases where the cost per use increase the number of spills. Disabling improves the codegen for #79918.

I propose adding this option to allow easier experimentation.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Craig Topper (topperc)

Changes

I've seen cases where the cost per use increase the number of spills. Disabling improves the codegen for #79918.

I propose adding this option to allow easier experimentation.

---

Full diff: https://github.com/llvm/llvm-project/pull/83320.diff

1 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVRegisterInfo.cpp (+6-1)

diff --git a/llvm/lib/Target/RISCV/RISCVRegisterInfo.cpp b/llvm/lib/Target/RISCV/RISCVRegisterInfo.cpp
index 9d1f01dffaaf47..a68674b221d38e 100644
--- a/llvm/lib/Target/RISCV/RISCVRegisterInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVRegisterInfo.cpp
@@ -30,6 +30,8 @@
 
 using namespace llvm;
 
+static cl::opt<bool> DisableCostPerUse("riscv-disable-cost-per-use",
+                                       cl::init(false), cl::Hidden);
 static cl::opt<bool>
     DisableRegAllocHints("riscv-disable-regalloc-hints", cl::Hidden,
                          cl::init(false),
@@ -712,7 +714,10 @@ void RISCVRegisterInfo::getOffsetOpcodes(const StackOffset &Offset,
 
 unsigned
 RISCVRegisterInfo::getRegisterCostTableIndex(const MachineFunction &MF) const {
-  return MF.getSubtarget<RISCVSubtarget>().hasStdExtCOrZca() ? 1 : 0;
+  return MF.getSubtarget<RISCVSubtarget>().hasStdExtCOrZca() &&
+                 !DisableCostPerUse
+             ? 1
+             : 0;
 }
 
 // Add two address hints to improve chances of being able to use a compressed

[preames]

LGTM

[dtcxzyw]

@topperc Could you please provide some performance data with the flag enabled?
Looks like enabling the flag leads to ~16% execution time regression and ~20% code size regression.

[topperc]

@topperc Could you please provide some performance data with the flag enabled? Looks like enabling the flag leads to ~16% execution time regression and ~20% code size regression.

I can try to get some numbers. I only had some specific examples where it seemed to cause something bad. Here's an example where enabling it reduces the number of spills https://godbolt.org/z/M4YTad6cf though it does increase the code size.