[RISCV] Sink splatted fpext operands #125800
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
We sink splatted operands in codegenprepare to help match .vx/.vf patterns. This extends it to also splat any fpext so that we can match widening vfwadd.vf/vfwadd.wf patterns too. Some instructions don't have .wf forms so there's no benefit to sinking the fpext. For simplicity this sinks them anyway and lets earlymachine-licm hoist them back out.
|
@llvm/pr-subscribers-backend-risc-v Author: Luke Lau (lukel97) ChangesWe sink splatted operands in codegenprepare to help match .vx/.vf patterns. This extends it to also splat any fpext so that we can match widening vfwadd.vf/vfwadd.wf patterns too. Some instructions don't have .wf forms so there's no benefit to sinking the fpext. For simplicity this sinks them anyway and lets earlymachine-licm hoist them back out. Full diff: https://github.com/llvm/llvm-project/pull/125800.diff 2 Files Affected:
diff --git a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
index fa7c7c57be25724..6799130bde130c7 100644
--- a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp
@@ -2712,7 +2712,12 @@ bool RISCVTTIImpl::isProfitableToSinkOperands(
return false;
}
- Ops.push_back(&Op->getOperandUse(0));
+ Use *InsertEltUse = &Op->getOperandUse(0);
+ // Sink any fpexts since they might be used in a widening fp pattern.
+ auto *InsertElt = cast<InsertElementInst>(InsertEltUse);
+ if (isa<FPExtInst>(InsertElt->getOperand(1)))
+ Ops.push_back(&InsertElt->getOperandUse(1));
+ Ops.push_back(InsertEltUse);
Ops.push_back(&OpIdx.value());
}
return true;
diff --git a/llvm/test/CodeGen/RISCV/rvv/sink-splat-operands.ll b/llvm/test/CodeGen/RISCV/rvv/sink-splat-operands.ll
index 735621aa4390e39..1948675ae9cf0a7 100644
--- a/llvm/test/CodeGen/RISCV/rvv/sink-splat-operands.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/sink-splat-operands.ll
@@ -5741,3 +5741,152 @@ vector.body:
for.cond.cleanup:
ret void
}
+
+define void @sink_splat_vfwadd_vf(ptr nocapture %a, ptr nocapture %b, float %f) {
+; CHECK-LABEL: sink_splat_vfwadd_vf:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: li a1, 0
+; CHECK-NEXT: li a2, 1020
+; CHECK-NEXT: vsetvli a3, zero, e32, m1, ta, ma
+; CHECK-NEXT: .LBB125_1: # %vector.body
+; CHECK-NEXT: # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: vl1re32.v v8, (a0)
+; CHECK-NEXT: addi a1, a1, 4
+; CHECK-NEXT: addi a2, a2, -4
+; CHECK-NEXT: vfwadd.vf v10, v8, fa0
+; CHECK-NEXT: vs2r.v v10, (a0)
+; CHECK-NEXT: addi a0, a0, 16
+; CHECK-NEXT: j .LBB125_1
+entry:
+ %f.ext = fpext float %f to double
+ %broadcast.splatinsert = insertelement <vscale x 2 x double> poison, double %f.ext, i32 0
+ %broadcast.splat = shufflevector <vscale x 2 x double> %broadcast.splatinsert, <vscale x 2 x double> poison, <vscale x 2 x i32> zeroinitializer
+ br label %vector.body
+
+vector.body: ; preds = %vector.body, %entry
+ %index = phi i64 [ 0, %entry ], [ %index.next, %vector.body ]
+ %0 = getelementptr float, ptr %a, i64 %index
+ %wide.load = load <vscale x 2 x float>, ptr %0
+ %ext = fpext <vscale x 2 x float> %wide.load to <vscale x 2 x double>
+ %1 = fadd <vscale x 2 x double> %ext, %broadcast.splat
+ %2 = getelementptr double, ptr %b, i64 %index
+ store <vscale x 2 x double> %1, ptr %0
+ %index.next = add i64 %index, 4
+ %3 = icmp eq i64 %index.next, 1024
+ br i1 32, label %for.cond.cleanup, label %vector.body
+
+for.cond.cleanup: ; preds = %vector.body
+ ret void
+}
+
+define void @sink_splat_vfwadd_wf(ptr nocapture %a, ptr nocapture %b, float %f) {
+; CHECK-LABEL: sink_splat_vfwadd_wf:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: li a1, 0
+; CHECK-NEXT: li a2, 1020
+; CHECK-NEXT: vsetvli a3, zero, e32, m1, ta, ma
+; CHECK-NEXT: .LBB126_1: # %vector.body
+; CHECK-NEXT: # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: vl2re64.v v8, (a0)
+; CHECK-NEXT: addi a1, a1, 4
+; CHECK-NEXT: addi a2, a2, -4
+; CHECK-NEXT: vfwadd.wf v8, v8, fa0
+; CHECK-NEXT: vs2r.v v8, (a0)
+; CHECK-NEXT: addi a0, a0, 32
+; CHECK-NEXT: j .LBB126_1
+entry:
+ %f.ext = fpext float %f to double
+ %broadcast.splatinsert = insertelement <vscale x 2 x double> poison, double %f.ext, i32 0
+ %broadcast.splat = shufflevector <vscale x 2 x double> %broadcast.splatinsert, <vscale x 2 x double> poison, <vscale x 2 x i32> zeroinitializer
+ br label %vector.body
+
+vector.body: ; preds = %vector.body, %entry
+ %index = phi i64 [ 0, %entry ], [ %index.next, %vector.body ]
+ %0 = getelementptr double, ptr %a, i64 %index
+ %wide.load = load <vscale x 2 x double>, ptr %0
+ %1 = fadd <vscale x 2 x double> %wide.load, %broadcast.splat
+ %2 = getelementptr double, ptr %b, i64 %index
+ store <vscale x 2 x double> %1, ptr %0
+ %index.next = add i64 %index, 4
+ %3 = icmp eq i64 %index.next, 1024
+ br i1 32, label %for.cond.cleanup, label %vector.body
+
+for.cond.cleanup: ; preds = %vector.body
+ ret void
+}
+
+define void @sink_splat_vfwmul_vf(ptr nocapture %a, ptr nocapture %b, float %f) {
+; CHECK-LABEL: sink_splat_vfwmul_vf:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: li a1, 0
+; CHECK-NEXT: li a2, 1020
+; CHECK-NEXT: vsetvli a3, zero, e32, m1, ta, ma
+; CHECK-NEXT: .LBB127_1: # %vector.body
+; CHECK-NEXT: # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: vl1re32.v v8, (a0)
+; CHECK-NEXT: addi a1, a1, 4
+; CHECK-NEXT: addi a2, a2, -4
+; CHECK-NEXT: vfwmul.vf v10, v8, fa0
+; CHECK-NEXT: vs2r.v v10, (a0)
+; CHECK-NEXT: addi a0, a0, 16
+; CHECK-NEXT: j .LBB127_1
+entry:
+ %f.ext = fpext float %f to double
+ %broadcast.splatinsert = insertelement <vscale x 2 x double> poison, double %f.ext, i32 0
+ %broadcast.splat = shufflevector <vscale x 2 x double> %broadcast.splatinsert, <vscale x 2 x double> poison, <vscale x 2 x i32> zeroinitializer
+ br label %vector.body
+
+vector.body: ; preds = %vector.body, %entry
+ %index = phi i64 [ 0, %entry ], [ %index.next, %vector.body ]
+ %0 = getelementptr float, ptr %a, i64 %index
+ %wide.load = load <vscale x 2 x float>, ptr %0
+ %ext = fpext <vscale x 2 x float> %wide.load to <vscale x 2 x double>
+ %1 = fmul <vscale x 2 x double> %ext, %broadcast.splat
+ %2 = getelementptr double, ptr %b, i64 %index
+ store <vscale x 2 x double> %1, ptr %0
+ %index.next = add i64 %index, 4
+ %3 = icmp eq i64 %index.next, 1024
+ br i1 32, label %for.cond.cleanup, label %vector.body
+
+for.cond.cleanup: ; preds = %vector.body
+ ret void
+}
+
+; Even though there's no vfwmul.wf we'll sink the fcvt.d.s. Make sure
+; early-machinelicm undos the sink after isel.
+define void @sink_splat_vfwmul_wf(ptr nocapture %a, ptr nocapture %b, float %f) {
+; CHECK-LABEL: sink_splat_vfwmul_wf:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: li a1, 0
+; CHECK-NEXT: li a2, 1020
+; CHECK-NEXT: fcvt.d.s fa5, fa0
+; CHECK-NEXT: vsetvli a3, zero, e64, m2, ta, ma
+; CHECK-NEXT: .LBB128_1: # %vector.body
+; CHECK-NEXT: # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: vl2re64.v v8, (a0)
+; CHECK-NEXT: addi a1, a1, 4
+; CHECK-NEXT: addi a2, a2, -4
+; CHECK-NEXT: vfmul.vf v8, v8, fa5
+; CHECK-NEXT: vs2r.v v8, (a0)
+; CHECK-NEXT: addi a0, a0, 16
+; CHECK-NEXT: j .LBB128_1
+entry:
+ %f.ext = fpext float %f to double
+ %broadcast.splatinsert = insertelement <vscale x 2 x double> poison, double %f.ext, i32 0
+ %broadcast.splat = shufflevector <vscale x 2 x double> %broadcast.splatinsert, <vscale x 2 x double> poison, <vscale x 2 x i32> zeroinitializer
+ br label %vector.body
+
+vector.body: ; preds = %vector.body, %entry
+ %index = phi i64 [ 0, %entry ], [ %index.next, %vector.body ]
+ %0 = getelementptr float, ptr %a, i64 %index
+ %wide.load = load <vscale x 2 x double>, ptr %0
+ %1 = fmul <vscale x 2 x double> %wide.load, %broadcast.splat
+ %2 = getelementptr double, ptr %b, i64 %index
+ store <vscale x 2 x double> %1, ptr %0
+ %index.next = add i64 %index, 4
+ %3 = icmp eq i64 %index.next, 1024
+ br i1 32, label %for.cond.cleanup, label %vector.body
+
+for.cond.cleanup: ; preds = %vector.body
+ ret void
+}
|
Icohedron
pushed a commit
to Icohedron/llvm-project
that referenced
this pull request
Feb 11, 2025
We sink splatted operands in codegenprepare to help match .vx/.vf patterns. This extends it to also splat any fpext so that we can match widening vfwadd.vf/vfwadd.wf patterns too. Some instructions don't have .wf forms so there's no benefit to sinking the fpext. For simplicity this sinks them anyway and lets earlymachine-licm hoist them back out.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
We sink splatted operands in codegenprepare to help match .vx/.vf patterns. This extends it to also splat any fpext so that we can match widening vfwadd.vf/vfwadd.wf patterns too.
Some instructions don't have .wf forms so there's no benefit to sinking the fpext. For simplicity this sinks them anyway and lets earlymachine-licm hoist them back out.