[llvm] [RISCV] Enable rematerialization for scalar loads (PR #166774)
Luke Lau via llvm-commits
llvm-commits at lists.llvm.org
Thu Nov 6 05:43:24 PST 2025
https://github.com/lukel97 created https://github.com/llvm/llvm-project/pull/166774
In some workloads we see an argument passed on the stack where it is loaded, only for it to be immediately spilled to a different slot on the stack and then reloaded from that spill slot.
We can avoid the unnecessary spill by marking loads as rematerializable. TargetTransformInfo::isReMaterializableImpl checks to make sure that any loads are `MI.isDereferenceableInvariantLoad()`, so we should be able to move the load down to the remat site.
This gives a 14.8% reduction in spills in 544.nab_r on rva23u64 -O3, and a few other smaller reductions on llvm-test-suite. I didn't find any benchmarks where the number of spills/reloads increased.
Related: #165761
>From c1d912993c651091cd481605fd0d99335d6bde82 Mon Sep 17 00:00:00 2001
From: Luke Lau <luke at igalia.com>
Date: Mon, 20 Oct 2025 18:06:11 +0800
Subject: [PATCH 1/2] Precommit tests
---
llvm/test/CodeGen/RISCV/remat.ll | 94 ++++++++++++++++++++++++++++++++
1 file changed, 94 insertions(+)
diff --git a/llvm/test/CodeGen/RISCV/remat.ll b/llvm/test/CodeGen/RISCV/remat.ll
index 92ae85f560cd4..654ca21713e8c 100644
--- a/llvm/test/CodeGen/RISCV/remat.ll
+++ b/llvm/test/CodeGen/RISCV/remat.ll
@@ -201,3 +201,97 @@ for.end: ; preds = %for.inc, %entry
}
declare i32 @foo(i32, i32, i32, i32, i32, i32)
+
+define i32 @remat_ld(i32 %0, i32 %1, i32 %2, i32 %3, i32 %4, i32 %5, i32 %6, i32 %7, i32 %stack0, i32 %stack1, ptr %p, ptr %q) {
+; RV32I-LABEL: remat_ld:
+; RV32I: # %bb.0: # %entry
+; RV32I-NEXT: addi sp, sp, -64
+; RV32I-NEXT: .cfi_def_cfa_offset 64
+; RV32I-NEXT: sw ra, 60(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s0, 56(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s1, 52(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s2, 48(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s3, 44(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s4, 40(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s5, 36(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s6, 32(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s7, 28(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s8, 24(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s9, 20(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s10, 16(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw s11, 12(sp) # 4-byte Folded Spill
+; RV32I-NEXT: .cfi_offset ra, -4
+; RV32I-NEXT: .cfi_offset s0, -8
+; RV32I-NEXT: .cfi_offset s1, -12
+; RV32I-NEXT: .cfi_offset s2, -16
+; RV32I-NEXT: .cfi_offset s3, -20
+; RV32I-NEXT: .cfi_offset s4, -24
+; RV32I-NEXT: .cfi_offset s5, -28
+; RV32I-NEXT: .cfi_offset s6, -32
+; RV32I-NEXT: .cfi_offset s7, -36
+; RV32I-NEXT: .cfi_offset s8, -40
+; RV32I-NEXT: .cfi_offset s9, -44
+; RV32I-NEXT: .cfi_offset s10, -48
+; RV32I-NEXT: .cfi_offset s11, -52
+; RV32I-NEXT: sw a0, 8(sp) # 4-byte Folded Spill
+; RV32I-NEXT: lw a2, 68(sp)
+; RV32I-NEXT: lw a3, 64(sp)
+; RV32I-NEXT: lw a0, 72(sp)
+; RV32I-NEXT: lw a1, 76(sp)
+; RV32I-NEXT: sw a3, 0(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw a3, 0(a0)
+; RV32I-NEXT: sw a2, 4(sp) # 4-byte Folded Spill
+; RV32I-NEXT: sw a2, 0(a1)
+; RV32I-NEXT: #APP
+; RV32I-NEXT: #NO_APP
+; RV32I-NEXT: lw a0, 8(sp) # 4-byte Folded Reload
+; RV32I-NEXT: beqz a0, .LBB1_2
+; RV32I-NEXT: # %bb.1: # %falsebb
+; RV32I-NEXT: li a0, 0
+; RV32I-NEXT: j .LBB1_3
+; RV32I-NEXT: .LBB1_2: # %truebb
+; RV32I-NEXT: lw a0, 4(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw a1, 0(sp) # 4-byte Folded Reload
+; RV32I-NEXT: add a0, a1, a0
+; RV32I-NEXT: .LBB1_3: # %falsebb
+; RV32I-NEXT: lw ra, 60(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s0, 56(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s1, 52(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s2, 48(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s3, 44(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s4, 40(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s5, 36(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s6, 32(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s7, 28(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s8, 24(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s9, 20(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s10, 16(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw s11, 12(sp) # 4-byte Folded Reload
+; RV32I-NEXT: .cfi_restore ra
+; RV32I-NEXT: .cfi_restore s0
+; RV32I-NEXT: .cfi_restore s1
+; RV32I-NEXT: .cfi_restore s2
+; RV32I-NEXT: .cfi_restore s3
+; RV32I-NEXT: .cfi_restore s4
+; RV32I-NEXT: .cfi_restore s5
+; RV32I-NEXT: .cfi_restore s6
+; RV32I-NEXT: .cfi_restore s7
+; RV32I-NEXT: .cfi_restore s8
+; RV32I-NEXT: .cfi_restore s9
+; RV32I-NEXT: .cfi_restore s10
+; RV32I-NEXT: .cfi_restore s11
+; RV32I-NEXT: addi sp, sp, 64
+; RV32I-NEXT: .cfi_def_cfa_offset 0
+; RV32I-NEXT: ret
+entry:
+ store i32 %stack0, ptr %p
+ store i32 %stack1, ptr %q
+ tail call void asm sideeffect "", "~{x1},~{x3},~{x4},~{x5},~{x6},~{x7},~{x8},~{x9},~{x10},~{x11},~{x12},~{x13},~{x14},~{x15},~{x16},~{x17},~{x18},~{x19},~{x20},~{x21},~{x22},~{x23},~{x24},~{x25},~{x26},~{x27},~{x28},~{x29},~{x30},~{x31}"()
+ %a = icmp eq i32 %0, 0
+ br i1 %a, label %truebb, label %falsebb
+truebb:
+ %b = add i32 %stack0, %stack1
+ ret i32 %b
+falsebb:
+ ret i32 0
+}
>From e7ebe513f83e701c0c3b4bbebdd6a1c1fefa1934 Mon Sep 17 00:00:00 2001
From: Luke Lau <luke at igalia.com>
Date: Mon, 20 Oct 2025 19:04:28 +0800
Subject: [PATCH 2/2] [RISCV] Enable rematerialization for scalar loads
---
llvm/lib/Target/RISCV/RISCVInstrInfo.td | 4 ++--
llvm/lib/Target/RISCV/RISCVInstrInfoD.td | 2 +-
llvm/lib/Target/RISCV/RISCVInstrInfoF.td | 2 +-
llvm/lib/Target/RISCV/RISCVInstrInfoZfh.td | 2 +-
llvm/test/CodeGen/RISCV/remat.ll | 6 ++----
llvm/test/CodeGen/RISCV/rvv/pr95865.ll | 4 +---
6 files changed, 8 insertions(+), 12 deletions(-)
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.td b/llvm/lib/Target/RISCV/RISCVInstrInfo.td
index 9cb53fb27a2d2..84b962b2a8607 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfo.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfo.td
@@ -768,7 +768,7 @@ def BGE : BranchCC_rri<0b101, "bge">;
def BLTU : BranchCC_rri<0b110, "bltu">;
def BGEU : BranchCC_rri<0b111, "bgeu">;
-let IsSignExtendingOpW = 1, canFoldAsLoad = 1 in {
+let IsSignExtendingOpW = 1, canFoldAsLoad = 1, isReMaterializable = 1 in {
def LB : Load_ri<0b000, "lb">, Sched<[WriteLDB, ReadMemBase]>;
def LH : Load_ri<0b001, "lh">, Sched<[WriteLDH, ReadMemBase]>;
def LW : Load_ri<0b010, "lw">, Sched<[WriteLDW, ReadMemBase]>;
@@ -889,7 +889,7 @@ def CSRRCI : CSR_ii<0b111, "csrrci">;
/// RV64I instructions
let Predicates = [IsRV64] in {
-let canFoldAsLoad = 1 in {
+let canFoldAsLoad = 1, isReMaterializable = 1 in {
def LWU : Load_ri<0b110, "lwu">, Sched<[WriteLDW, ReadMemBase]>;
def LD : Load_ri<0b011, "ld">, Sched<[WriteLDD, ReadMemBase]>;
}
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoD.td b/llvm/lib/Target/RISCV/RISCVInstrInfoD.td
index 4ffe3e62ac501..deacd41e6469a 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoD.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoD.td
@@ -71,7 +71,7 @@ defvar DExtsRV64 = [DExt, ZdinxExt];
//===----------------------------------------------------------------------===//
let Predicates = [HasStdExtD] in {
-let canFoldAsLoad = 1 in
+let canFoldAsLoad = 1, isReMaterializable = 1 in
def FLD : FPLoad_r<0b011, "fld", FPR64, WriteFLD64>;
// Operands for stores are in the order srcreg, base, offset rather than
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoF.td b/llvm/lib/Target/RISCV/RISCVInstrInfoF.td
index b30f8ec820c15..bd191001b75ec 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoF.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoF.td
@@ -330,7 +330,7 @@ class PseudoFROUND<DAGOperand Ty, ValueType vt, ValueType intvt = XLenVT>
//===----------------------------------------------------------------------===//
let Predicates = [HasStdExtF] in {
-let canFoldAsLoad = 1 in
+let canFoldAsLoad = 1, isReMaterializable = 1 in
def FLW : FPLoad_r<0b010, "flw", FPR32, WriteFLD32>;
// Operands for stores are in the order srcreg, base, offset rather than
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoZfh.td b/llvm/lib/Target/RISCV/RISCVInstrInfoZfh.td
index 1c6a5afcda49b..c172d1739ba61 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoZfh.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoZfh.td
@@ -90,7 +90,7 @@ defvar ZfhminDExts = [ZfhminDExt, ZhinxminZdinxExt, ZhinxminZdinx32Ext];
//===----------------------------------------------------------------------===//
let Predicates = [HasHalfFPLoadStoreMove] in {
-let canFoldAsLoad = 1 in
+let canFoldAsLoad = 1, isReMaterializable = 1 in
def FLH : FPLoad_r<0b001, "flh", FPR16, WriteFLD16>;
// Operands for stores are in the order srcreg, base, offset rather than
diff --git a/llvm/test/CodeGen/RISCV/remat.ll b/llvm/test/CodeGen/RISCV/remat.ll
index 654ca21713e8c..88240474f30bf 100644
--- a/llvm/test/CodeGen/RISCV/remat.ll
+++ b/llvm/test/CodeGen/RISCV/remat.ll
@@ -238,9 +238,7 @@ define i32 @remat_ld(i32 %0, i32 %1, i32 %2, i32 %3, i32 %4, i32 %5, i32 %6, i32
; RV32I-NEXT: lw a3, 64(sp)
; RV32I-NEXT: lw a0, 72(sp)
; RV32I-NEXT: lw a1, 76(sp)
-; RV32I-NEXT: sw a3, 0(sp) # 4-byte Folded Spill
; RV32I-NEXT: sw a3, 0(a0)
-; RV32I-NEXT: sw a2, 4(sp) # 4-byte Folded Spill
; RV32I-NEXT: sw a2, 0(a1)
; RV32I-NEXT: #APP
; RV32I-NEXT: #NO_APP
@@ -250,8 +248,8 @@ define i32 @remat_ld(i32 %0, i32 %1, i32 %2, i32 %3, i32 %4, i32 %5, i32 %6, i32
; RV32I-NEXT: li a0, 0
; RV32I-NEXT: j .LBB1_3
; RV32I-NEXT: .LBB1_2: # %truebb
-; RV32I-NEXT: lw a0, 4(sp) # 4-byte Folded Reload
-; RV32I-NEXT: lw a1, 0(sp) # 4-byte Folded Reload
+; RV32I-NEXT: lw a0, 68(sp)
+; RV32I-NEXT: lw a1, 64(sp)
; RV32I-NEXT: add a0, a1, a0
; RV32I-NEXT: .LBB1_3: # %falsebb
; RV32I-NEXT: lw ra, 60(sp) # 4-byte Folded Reload
diff --git a/llvm/test/CodeGen/RISCV/rvv/pr95865.ll b/llvm/test/CodeGen/RISCV/rvv/pr95865.ll
index ab9849631663c..01d66b344ec2e 100644
--- a/llvm/test/CodeGen/RISCV/rvv/pr95865.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/pr95865.ll
@@ -40,8 +40,6 @@ define i32 @main(i1 %arg.1, i64 %arg.2, i1 %arg.3, i64 %arg.4, i1 %arg.5, <vscal
; CHECK-NEXT: li t0, 12
; CHECK-NEXT: li s0, 4
; CHECK-NEXT: li t1, 20
-; CHECK-NEXT: ld a1, 112(sp)
-; CHECK-NEXT: sd a1, 0(sp) # 8-byte Folded Spill
; CHECK-NEXT: vsetivli zero, 8, e32, m2, ta, ma
; CHECK-NEXT: vmv.v.i v8, 0
; CHECK-NEXT: andi t3, a4, 1
@@ -142,7 +140,7 @@ define i32 @main(i1 %arg.1, i64 %arg.2, i1 %arg.3, i64 %arg.4, i1 %arg.5, <vscal
; CHECK-NEXT: j .LBB0_11
; CHECK-NEXT: .LBB0_12: # %for.body7.us.19
; CHECK-NEXT: vsetvli a0, zero, e32, m8, ta, ma
-; CHECK-NEXT: ld a0, 0(sp) # 8-byte Folded Reload
+; CHECK-NEXT: ld a0, 112(sp)
; CHECK-NEXT: vmv.s.x v16, a0
; CHECK-NEXT: vmv.v.i v8, 0
; CHECK-NEXT: vsetivli zero, 2, e32, m1, tu, ma
More information about the llvm-commits
mailing list