[RISCV] Don't expand zero stride vp.strided.load if SEW>XLEN

[lukel97]

A splat of a on RV32 will get lowered as a zero strided load anyway (and won't match any .vx splat patterns), so don't expand it to a scalar load + splat to avoid writing it to the stack.

[llvmbot]

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

Author: Luke Lau (lukel97)

Changes

A splat of a <n x i64> on RV32 will get lowered as a zero strided load anyway (and won't match any .vx splat patterns), so don't expand it to a scalar load + splat to avoid writing it to the stack.

---

Full diff: https://github.com/llvm/llvm-project/pull/98924.diff

3 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVCodeGenPrepare.cpp (+5)
• (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-vpload.ll (-8)
• (modified) llvm/test/CodeGen/RISCV/rvv/strided-vpload.ll (-8)

diff --git a/llvm/lib/Target/RISCV/RISCVCodeGenPrepare.cpp b/llvm/lib/Target/RISCV/RISCVCodeGenPrepare.cpp
index b3f3dc6e2256c..0e84eda0c9d07 100644
--- a/llvm/lib/Target/RISCV/RISCVCodeGenPrepare.cpp
+++ b/llvm/lib/Target/RISCV/RISCVCodeGenPrepare.cpp
@@ -174,6 +174,11 @@ bool RISCVCodeGenPrepare::expandVPStrideLoad(IntrinsicInst &II) {
                       m_Value(BasePtr), m_Zero(), m_AllOnes(), m_Value(VL))))
     return false;
 
+  // If SEW>XLEN then a splat will get lowered as a zero strided load anyway, so
+  // avoid expanding here.
+  if (II.getType()->getScalarSizeInBits() > ST->getXLen())
+    return false;
+
   if (!isKnownNonZero(VL, {*DL, DT, nullptr, &II}))
     return false;
 
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-vpload.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-vpload.ll
index 95f853b77f18b..b8c7037580c46 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-vpload.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-strided-vpload.ll
@@ -674,17 +674,9 @@ define <4 x half> @zero_strided_unmasked_vpload_4f16(ptr %ptr) {
 define <4 x i64> @zero_strided_vadd.vx(<4 x i64> %v, ptr %ptr) {
 ; CHECK-RV32-LABEL: zero_strided_vadd.vx:
 ; CHECK-RV32:       # %bb.0:
-; CHECK-RV32-NEXT:    addi sp, sp, -16
-; CHECK-RV32-NEXT:    .cfi_def_cfa_offset 16
-; CHECK-RV32-NEXT:    lw a1, 4(a0)
-; CHECK-RV32-NEXT:    lw a0, 0(a0)
-; CHECK-RV32-NEXT:    sw a1, 12(sp)
-; CHECK-RV32-NEXT:    sw a0, 8(sp)
-; CHECK-RV32-NEXT:    addi a0, sp, 8
 ; CHECK-RV32-NEXT:    vsetivli zero, 4, e64, m2, ta, ma
 ; CHECK-RV32-NEXT:    vlse64.v v10, (a0), zero
 ; CHECK-RV32-NEXT:    vadd.vv v8, v8, v10
-; CHECK-RV32-NEXT:    addi sp, sp, 16
 ; CHECK-RV32-NEXT:    ret
 ;
 ; CHECK-RV64-LABEL: zero_strided_vadd.vx:
diff --git a/llvm/test/CodeGen/RISCV/rvv/strided-vpload.ll b/llvm/test/CodeGen/RISCV/rvv/strided-vpload.ll
index 563da270272c2..0010f64a93fd6 100644
--- a/llvm/test/CodeGen/RISCV/rvv/strided-vpload.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/strided-vpload.ll
@@ -826,17 +826,9 @@ define <vscale x 1 x half> @zero_strided_unmasked_vpload_nxv1f16(ptr %ptr) {
 define <vscale x 1 x i64> @zero_strided_vadd.vx(<vscale x 1 x i64> %v, ptr %ptr) {
 ; CHECK-RV32-LABEL: zero_strided_vadd.vx:
 ; CHECK-RV32:       # %bb.0:
-; CHECK-RV32-NEXT:    addi sp, sp, -16
-; CHECK-RV32-NEXT:    .cfi_def_cfa_offset 16
-; CHECK-RV32-NEXT:    lw a1, 4(a0)
-; CHECK-RV32-NEXT:    lw a0, 0(a0)
-; CHECK-RV32-NEXT:    sw a1, 12(sp)
-; CHECK-RV32-NEXT:    sw a0, 8(sp)
-; CHECK-RV32-NEXT:    addi a0, sp, 8
 ; CHECK-RV32-NEXT:    vsetvli a1, zero, e64, m1, ta, ma
 ; CHECK-RV32-NEXT:    vlse64.v v9, (a0), zero
 ; CHECK-RV32-NEXT:    vadd.vv v8, v8, v9
-; CHECK-RV32-NEXT:    addi sp, sp, 16
 ; CHECK-RV32-NEXT:    ret
 ;
 ; CHECK-RV64-LABEL: zero_strided_vadd.vx:

[topperc]

LGTM