[llvm] [RISCV][VLOPT] Add support for integer multiply-add instructions (PR #112216)
Michael Maitland via llvm-commits
llvm-commits at lists.llvm.org
Wed Oct 16 06:46:04 PDT 2024
https://github.com/michaelmaitland updated https://github.com/llvm/llvm-project/pull/112216
>From 24d86e8274b53c2f563da13c0e9756aa21d79b12 Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Mon, 14 Oct 2024 07:25:20 -0700
Subject: [PATCH 1/2] [RISCV][VLOPT] Add support for integer multiply-add
instructions
---
llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp | 9 +-
llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll | 176 +++++++++++++++++++
2 files changed, 184 insertions(+), 1 deletion(-)
diff --git a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
index 53373b7a0f1575..4ab693eaf16f5a 100644
--- a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
@@ -565,7 +565,14 @@ static bool isSupportedInstr(const MachineInstr &MI) {
// Vector Widening Integer Multiply Instructions
// FIXME: Add support
// Vector Single-Width Integer Multiply-Add Instructions
- // FIXME: Add support
+ case RISCV::VMACC_VV:
+ case RISCV::VMACC_VX:
+ case RISCV::VNMSAC_VV:
+ case RISCV::VNMSAC_VX:
+ case RISCV::VMADD_VV:
+ case RISCV::VMADD_VX:
+ case RISCV::VNMSUB_VV:
+ case RISCV::VNMSUB_VX:
// Vector Widening Integer Multiply-Add Instructions
// FIXME: Add support
case RISCV::VWMACC_VX:
diff --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
index a360ae1998f77a..5dd174fe4c4ae2 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -1122,6 +1122,182 @@ define <vscale x 4 x i32> @vrem_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
ret <vscale x 4 x i32> %2
}
+define <vscale x 4 x i32> @vmacc_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vmacc_vv:
+; NOVLOPT: # %bb.0:
+; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, tu, ma
+; NOVLOPT-NEXT: vmacc.vv v8, v8, v10
+; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT: vmul.vv v8, v8, v10
+; NOVLOPT-NEXT: ret
+;
+; VLOPT-LABEL: vmacc_vv:
+; VLOPT: # %bb.0:
+; VLOPT-NEXT: vsetvli zero, a0, e32, m2, tu, ma
+; VLOPT-NEXT: vmacc.vv v8, v8, v10
+; VLOPT-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT: vmul.vv v8, v8, v10
+; VLOPT-NEXT: ret
+ %1 = call <vscale x 4 x i32> @llvm.riscv.vmacc.nxv4i32.nxv4i32(<vscale x 4 x i32> %a, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1, iXLen 0)
+ %2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
+ ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vmacc_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
+; NOVLOPT-LABEL: vmacc_vx:
+; NOVLOPT: # %bb.0:
+; NOVLOPT-NEXT: vmv2r.v v10, v8
+; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, tu, ma
+; NOVLOPT-NEXT: vmacc.vx v10, a0, v8
+; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT: vmul.vv v8, v10, v8
+; NOVLOPT-NEXT: ret
+;
+; VLOPT-LABEL: vmacc_vx:
+; VLOPT: # %bb.0:
+; VLOPT-NEXT: vmv2r.v v10, v8
+; VLOPT-NEXT: vsetvli zero, a1, e32, m2, tu, ma
+; VLOPT-NEXT: vmacc.vx v10, a0, v8
+; VLOPT-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT: vmul.vv v8, v10, v8
+; VLOPT-NEXT: ret
+ %1 = call <vscale x 4 x i32> @llvm.riscv.vmacc.nxv4i32.nxv4i32(<vscale x 4 x i32> %a, i32 %b, <vscale x 4 x i32> %a, iXLen -1, iXLen 0)
+ %2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+ ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vmadd_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vmadd_vv:
+; NOVLOPT: # %bb.0:
+; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, tu, ma
+; NOVLOPT-NEXT: vmadd.vv v8, v8, v10
+; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT: vmul.vv v8, v8, v10
+; NOVLOPT-NEXT: ret
+;
+; VLOPT-LABEL: vmadd_vv:
+; VLOPT: # %bb.0:
+; VLOPT-NEXT: vsetvli zero, a0, e32, m2, tu, ma
+; VLOPT-NEXT: vmadd.vv v8, v8, v10
+; VLOPT-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT: vmul.vv v8, v8, v10
+; VLOPT-NEXT: ret
+ %1 = call <vscale x 4 x i32> @llvm.riscv.vmadd.nxv4i32.nxv4i32(<vscale x 4 x i32> %a, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1, iXLen 0)
+ %2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
+ ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vmadd_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
+; NOVLOPT-LABEL: vmadd_vx:
+; NOVLOPT: # %bb.0:
+; NOVLOPT-NEXT: vmv2r.v v10, v8
+; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, tu, ma
+; NOVLOPT-NEXT: vmadd.vx v10, a0, v8
+; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT: vmul.vv v8, v10, v8
+; NOVLOPT-NEXT: ret
+;
+; VLOPT-LABEL: vmadd_vx:
+; VLOPT: # %bb.0:
+; VLOPT-NEXT: vmv2r.v v10, v8
+; VLOPT-NEXT: vsetvli zero, a1, e32, m2, tu, ma
+; VLOPT-NEXT: vmadd.vx v10, a0, v8
+; VLOPT-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT: vmul.vv v8, v10, v8
+; VLOPT-NEXT: ret
+ %1 = call <vscale x 4 x i32> @llvm.riscv.vmadd.nxv4i32.nxv4i32(<vscale x 4 x i32> %a, i32 %b, <vscale x 4 x i32> %a, iXLen -1, iXLen 0)
+ %2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+ ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vnmsac_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vnmsac_vv:
+; NOVLOPT: # %bb.0:
+; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, tu, ma
+; NOVLOPT-NEXT: vnmsac.vv v8, v8, v10
+; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT: vmul.vv v8, v8, v10
+; NOVLOPT-NEXT: ret
+;
+; VLOPT-LABEL: vnmsac_vv:
+; VLOPT: # %bb.0:
+; VLOPT-NEXT: vsetvli zero, a0, e32, m2, tu, ma
+; VLOPT-NEXT: vnmsac.vv v8, v8, v10
+; VLOPT-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT: vmul.vv v8, v8, v10
+; VLOPT-NEXT: ret
+ %1 = call <vscale x 4 x i32> @llvm.riscv.vnmsac.nxv4i32.nxv4i32(<vscale x 4 x i32> %a, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1, iXLen 0)
+ %2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
+ ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vnmsac_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
+; NOVLOPT-LABEL: vnmsac_vx:
+; NOVLOPT: # %bb.0:
+; NOVLOPT-NEXT: vmv2r.v v10, v8
+; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, tu, ma
+; NOVLOPT-NEXT: vnmsac.vx v10, a0, v8
+; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT: vmul.vv v8, v10, v8
+; NOVLOPT-NEXT: ret
+;
+; VLOPT-LABEL: vnmsac_vx:
+; VLOPT: # %bb.0:
+; VLOPT-NEXT: vmv2r.v v10, v8
+; VLOPT-NEXT: vsetvli zero, a1, e32, m2, tu, ma
+; VLOPT-NEXT: vnmsac.vx v10, a0, v8
+; VLOPT-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT: vmul.vv v8, v10, v8
+; VLOPT-NEXT: ret
+ %1 = call <vscale x 4 x i32> @llvm.riscv.vnmsac.nxv4i32.nxv4i32(<vscale x 4 x i32> %a, i32 %b, <vscale x 4 x i32> %a, iXLen -1, iXLen 0)
+ %2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+ ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vnmsub_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vnmsub_vv:
+; NOVLOPT: # %bb.0:
+; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, tu, ma
+; NOVLOPT-NEXT: vnmsub.vv v8, v8, v10
+; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT: vmul.vv v8, v8, v10
+; NOVLOPT-NEXT: ret
+;
+; VLOPT-LABEL: vnmsub_vv:
+; VLOPT: # %bb.0:
+; VLOPT-NEXT: vsetvli zero, a0, e32, m2, tu, ma
+; VLOPT-NEXT: vnmsub.vv v8, v8, v10
+; VLOPT-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT: vmul.vv v8, v8, v10
+; VLOPT-NEXT: ret
+ %1 = call <vscale x 4 x i32> @llvm.riscv.vnmsub.nxv4i32.nxv4i32(<vscale x 4 x i32> %a, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1, iXLen 0)
+ %2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
+ ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vnmsub_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
+; NOVLOPT-LABEL: vnmsub_vx:
+; NOVLOPT: # %bb.0:
+; NOVLOPT-NEXT: vmv2r.v v10, v8
+; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, tu, ma
+; NOVLOPT-NEXT: vnmsub.vx v10, a0, v8
+; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT: vmul.vv v8, v10, v8
+; NOVLOPT-NEXT: ret
+;
+; VLOPT-LABEL: vnmsub_vx:
+; VLOPT: # %bb.0:
+; VLOPT-NEXT: vmv2r.v v10, v8
+; VLOPT-NEXT: vsetvli zero, a1, e32, m2, tu, ma
+; VLOPT-NEXT: vnmsub.vx v10, a0, v8
+; VLOPT-NEXT: vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT: vmul.vv v8, v10, v8
+; VLOPT-NEXT: ret
+ %1 = call <vscale x 4 x i32> @llvm.riscv.vnmsub.nxv4i32.nxv4i32(<vscale x 4 x i32> %a, i32 %b, <vscale x 4 x i32> %a, iXLen -1, iXLen 0)
+ %2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+ ret <vscale x 4 x i32> %2
+}
+
define <vscale x 4 x i32> @vwmacc_vx(<vscale x 4 x i16> %a, i16 %b, iXLen %vl) {
; NOVLOPT-LABEL: vwmacc_vx:
; NOVLOPT: # %bb.0:
>From f48ed134a0073ee5bbb18545bc29cd058d90d3c0 Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Wed, 16 Oct 2024 06:45:26 -0700
Subject: [PATCH 2/2] fixup! add test for source operand
---
llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll | 21 ++++++++++++++++++++
1 file changed, 21 insertions(+)
diff --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
index 5dd174fe4c4ae2..c20275ae232dca 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -1143,6 +1143,27 @@ define <vscale x 4 x i32> @vmacc_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b
ret <vscale x 4 x i32> %2
}
+define <vscale x 4 x i32> @vmacc_vv_use(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vmacc_vv_use:
+; NOVLOPT: # %bb.0:
+; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT: vmul.vv v12, v8, v10
+; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, tu, ma
+; NOVLOPT-NEXT: vmacc.vv v8, v12, v10
+; NOVLOPT-NEXT: ret
+;
+; VLOPT-LABEL: vmacc_vv_use:
+; VLOPT: # %bb.0:
+; VLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT: vmul.vv v12, v8, v10
+; VLOPT-NEXT: vsetvli zero, zero, e32, m2, tu, ma
+; VLOPT-NEXT: vmacc.vv v8, v12, v10
+; VLOPT-NEXT: ret
+ %1 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1)
+ %2 = call <vscale x 4 x i32> @llvm.riscv.vmacc.nxv4i32.nxv4i32(<vscale x 4 x i32> %a, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl, iXLen 0)
+ ret <vscale x 4 x i32> %2
+}
+
define <vscale x 4 x i32> @vmacc_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
; NOVLOPT-LABEL: vmacc_vx:
; NOVLOPT: # %bb.0:
More information about the llvm-commits
mailing list