[llvm] [RISCV][VLOPT] Add support for widening integer mul-add instructions (PR #112219)

Michael Maitland via llvm-commits llvm-commits at lists.llvm.org
Tue Oct 15 15:42:41 PDT 2024


https://github.com/michaelmaitland updated https://github.com/llvm/llvm-project/pull/112219

>From 53726c479520855aa4a2a33f56a70741ee0598c1 Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Mon, 14 Oct 2024 07:46:51 -0700
Subject: [PATCH 1/6] [RISCV][VLOPT] Add support for 11.14 widening integer
 mul-add instructions

---
 llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp   |  12 +-
 llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll | 137 ++++++++++++++++---
 2 files changed, 129 insertions(+), 20 deletions(-)

diff --git a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
index 088f6d62dcbe78..fbfbc724dc1061 100644
--- a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
@@ -422,8 +422,8 @@ static OperandInfo getOperandInfo(const MachineInstr &MI,
   case RISCV::VWSUB_WX:
   // Vector Widening Integer Multiply-Add Instructions
   // Destination EEW=2*SEW and EMUL=2*LMUL. Source EEW=SEW and EMUL=LMUL.
-  // Even though the add is a 2*SEW addition, the operands of the add are the
-  // Dest which is 2*SEW and the result of the multiply which is 2*SEW.
+  // A SEW-bit*SEW-bit multiply of the sources forms a 2*SEW-bit value, which
+  // is then added to the 2*SEW-bit Dest.
   case RISCV::VWMACCU_VV:
   case RISCV::VWMACCU_VX:
   case RISCV::VWMACC_VV:
@@ -567,9 +567,13 @@ static bool isSupportedInstr(const MachineInstr &MI) {
   // Vector Single-Width Integer Multiply-Add Instructions
   // FIXME: Add support
   // Vector Widening Integer Multiply-Add Instructions
-  // FIXME: Add support
-  case RISCV::VWMACC_VX:
+  case RISCV::VWMACCU_VV:
   case RISCV::VWMACCU_VX:
+  case RISCV::VWMACC_VV:
+  case RISCV::VWMACC_VX:
+  case RISCV::VWMACCSU_VV:
+  case RISCV::VWMACCSU_VX:
+  case RISCV::VWMACCUS_VX:
   // Vector Integer Merge Instructions
   // FIXME: Add support
   // Vector Integer Move Instructions
diff --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
index a360ae1998f77a..6c7abee4069c1e 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -1122,44 +1122,149 @@ 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> @vwmacc_vx(<vscale x 4 x i16> %a, i16 %b, iXLen %vl) {
+define <vscale x 4 x i32> @vwmacc_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen %vl) {
+; NOVLOPT-LABEL: vwmacc_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmacc.vv v8, v10, v11
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vwmacc_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a0, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmacc.vv v8, v10, v11
+; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT:    vadd.vv v8, v8, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.nxv4i16(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vwmacc_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
 ; NOVLOPT-LABEL: vwmacc_vx:
 ; NOVLOPT:       # %bb.0:
-; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, ta, ma
-; NOVLOPT-NEXT:    vwmacc.vx v10, a0, v8
+; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmacc.vx v8, a0, v10
 ; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
-; NOVLOPT-NEXT:    vadd.vv v8, v10, v10
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
 ; NOVLOPT-NEXT:    ret
 ;
 ; VLOPT-LABEL: vwmacc_vx:
 ; VLOPT:       # %bb.0:
-; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, ta, ma
-; VLOPT-NEXT:    vwmacc.vx v10, a0, v8
+; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmacc.vx v8, a0, v10
 ; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
-; VLOPT-NEXT:    vadd.vv v8, v10, v10
+; VLOPT-NEXT:    vadd.vv v8, v8, v8
 ; VLOPT-NEXT:    ret
-  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.i16(<vscale x 4 x i32> poison, i16 %b, <vscale x 4 x i16> %a, iXLen -1, iXLen 0)
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.i16(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
   %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
   ret <vscale x 4 x i32> %2
 }
 
-define <vscale x 4 x i32> @vwmaccu_vx(<vscale x 4 x i16> %a, i16 %b, iXLen %vl) {
+define <vscale x 4 x i32> @vwmaccu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen %vl) {
+; NOVLOPT-LABEL: vwmaccu_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmaccu.vv v8, v10, v11
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vwmaccu_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a0, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmaccu.vv v8, v10, v11
+; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT:    vadd.vv v8, v8, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.nxv4i16(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vwmaccu_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
 ; NOVLOPT-LABEL: vwmaccu_vx:
 ; NOVLOPT:       # %bb.0:
-; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, ta, ma
-; NOVLOPT-NEXT:    vwmaccu.vx v10, a0, v8
+; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmaccu.vx v8, a0, v10
 ; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
-; NOVLOPT-NEXT:    vadd.vv v8, v10, v10
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
 ; NOVLOPT-NEXT:    ret
 ;
 ; VLOPT-LABEL: vwmaccu_vx:
 ; VLOPT:       # %bb.0:
-; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, ta, ma
-; VLOPT-NEXT:    vwmaccu.vx v10, a0, v8
+; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmaccu.vx v8, a0, v10
 ; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
-; VLOPT-NEXT:    vadd.vv v8, v10, v10
+; VLOPT-NEXT:    vadd.vv v8, v8, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.i16(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vwmaccsu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen %vl) {
+; NOVLOPT-LABEL: vwmaccsu_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmaccsu.vv v8, v10, v11
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vwmaccsu_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a0, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmaccsu.vv v8, v10, v11
+; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT:    vadd.vv v8, v8, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccsu.nxv4i32.nxv4i16(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vwmaccsu_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
+; NOVLOPT-LABEL: vwmaccsu_vx:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmaccsu.vx v8, a0, v10
+; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vwmaccsu_vx:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmaccsu.vx v8, a0, v10
+; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT:    vadd.vv v8, v8, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccsu.nxv4i32.i16(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vwmaccus_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
+; NOVLOPT-LABEL: vwmaccus_vx:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmaccus.vx v8, a0, v10
+; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vwmaccus_vx:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmaccus.vx v8, a0, v10
+; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
+; VLOPT-NEXT:    vadd.vv v8, v8, v8
 ; VLOPT-NEXT:    ret
-  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.i16(<vscale x 4 x i32> poison, i16 %b, <vscale x 4 x i16> %a, iXLen -1, iXLen 0)
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccus.nxv4i32.i16(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
   %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
   ret <vscale x 4 x i32> %2
 }

>From fb347ac530c5a6b40640030ef543b7d2c72abe36 Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Mon, 14 Oct 2024 12:40:48 -0700
Subject: [PATCH 2/6] fixup! test getOperandInfo for sources

---
 llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll | 94 +++++++++++++++-----
 1 file changed, 73 insertions(+), 21 deletions(-)

diff --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
index 6c7abee4069c1e..bbaf70eae10fa7 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -1164,48 +1164,100 @@ define <vscale x 4 x i32> @vwmacc_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4
   ret <vscale x 4 x i32> %2
 }
 
-define <vscale x 4 x i32> @vwmaccu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen %vl) {
-; NOVLOPT-LABEL: vwmaccu_vv:
+define <vscale x 4 x i64> @vwmaccu_vv_nopropagate(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, iXLen %vl) {
+; NOVLOPT-LABEL: vwmaccu_vv_nopropagate:
 ; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vmv2r.v v16, v8
 ; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
-; NOVLOPT-NEXT:    vwmaccu.vv v8, v10, v11
-; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
-; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
+; NOVLOPT-NEXT:    vwmaccu.vv v16, v10, v11
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, tu, ma
+; NOVLOPT-NEXT:    vwmaccu.vv v12, v16, v8
+; NOVLOPT-NEXT:    vmv4r.v v8, v12
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vwmaccu_vv_nopropagate:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vmv2r.v v16, v8
+; VLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmaccu.vv v16, v10, v11
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, tu, ma
+; VLOPT-NEXT:    vwmaccu.vv v12, v16, v8
+; VLOPT-NEXT:    vmv4r.v v8, v12
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.nxv4i16(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
+  %2 = call <vscale x 4 x i64> @llvm.riscv.vwmaccu.nxv4i64.nxv4i32(<vscale x 4 x i64> %d, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl, iXLen 0)
+  ret <vscale x 4 x i64> %2
+}
+
+define <vscale x 4 x i64> @vwmaccu_vv_propagate(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, iXLen %vl) {
+; NOVLOPT-LABEL: vwmaccu_vv_propagate:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vmv2r.v v16, v8
+; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmaccu.vv v16, v10, v11
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, tu, ma
+; NOVLOPT-NEXT:    vwmaccu.vv v12, v8, v16
+; NOVLOPT-NEXT:    vmv4r.v v8, v12
 ; NOVLOPT-NEXT:    ret
 ;
-; VLOPT-LABEL: vwmaccu_vv:
+; VLOPT-LABEL: vwmaccu_vv_propagate:
 ; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vmv2r.v v16, v8
 ; VLOPT-NEXT:    vsetvli zero, a0, e16, m1, tu, ma
-; VLOPT-NEXT:    vwmaccu.vv v8, v10, v11
-; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
-; VLOPT-NEXT:    vadd.vv v8, v8, v8
+; VLOPT-NEXT:    vwmaccu.vv v16, v10, v11
+; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, tu, ma
+; VLOPT-NEXT:    vwmaccu.vv v12, v8, v16
+; VLOPT-NEXT:    vmv4r.v v8, v12
 ; VLOPT-NEXT:    ret
   %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.nxv4i16(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
-  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
-  ret <vscale x 4 x i32> %2
+  %2 = call <vscale x 4 x i64> @llvm.riscv.vwmaccu.nxv4i64.nxv4i32(<vscale x 4 x i64> %d, <vscale x 4 x i32> %a, <vscale x 4 x i32> %1, iXLen %vl, iXLen 0)
+  ret <vscale x 4 x i64> %2
 }
 
-define <vscale x 4 x i32> @vwmaccu_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
-; NOVLOPT-LABEL: vwmaccu_vx:
+define <vscale x 4 x i32> @vwmaccu_vx_nopropagate(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, <vscale x 4 x i16> %d, i16 %e, iXLen %vl) {
+; NOVLOPT-LABEL: vwmaccu_vx_nopropagate:
 ; NOVLOPT:       # %bb.0:
-; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vsetvli a3, zero, e16, m1, tu, ma
 ; NOVLOPT-NEXT:    vwmaccu.vx v8, a0, v10
-; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
-; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
+; NOVLOPT-NEXT:    vsetvli zero, a2, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmaccu.vx v8, a1, v11
 ; NOVLOPT-NEXT:    ret
 ;
-; VLOPT-LABEL: vwmaccu_vx:
+; VLOPT-LABEL: vwmaccu_vx_nopropagate:
 ; VLOPT:       # %bb.0:
-; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
+; VLOPT-NEXT:    vsetvli a3, zero, e16, m1, tu, ma
 ; VLOPT-NEXT:    vwmaccu.vx v8, a0, v10
-; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
-; VLOPT-NEXT:    vadd.vv v8, v8, v8
+; VLOPT-NEXT:    vsetvli zero, a2, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmaccu.vx v8, a1, v11
 ; VLOPT-NEXT:    ret
   %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.i16(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
-  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.i16(<vscale x 4 x i32> %1, i16 %e, <vscale x 4 x i16> %d, iXLen %vl, iXLen 0)
   ret <vscale x 4 x i32> %2
 }
 
+define <vscale x 4 x i64> @vwmaccu_vx_propagate(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, i32 %e, iXLen %vl) {
+; NOVLOPT-LABEL: vwmaccu_vx_propagate:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a3, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmaccu.vx v8, a0, v10
+; NOVLOPT-NEXT:    vsetvli zero, a2, e32, m2, tu, ma
+; NOVLOPT-NEXT:    vwmaccu.vx v12, a1, v8
+; NOVLOPT-NEXT:    vmv4r.v v8, v12
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vwmaccu_vx_propagate:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a2, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmaccu.vx v8, a0, v10
+; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, tu, ma
+; VLOPT-NEXT:    vwmaccu.vx v12, a1, v8
+; VLOPT-NEXT:    vmv4r.v v8, v12
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.i16(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
+  %2 = call <vscale x 4 x i64> @llvm.riscv.vwmaccu.nxv4i64.i32(<vscale x 4 x i64> %d, i32 %e, <vscale x 4 x i32> %1, iXLen %vl, iXLen 0)
+  ret <vscale x 4 x i64> %2
+}
+
 define <vscale x 4 x i32> @vwmaccsu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen %vl) {
 ; NOVLOPT-LABEL: vwmaccsu_vv:
 ; NOVLOPT:       # %bb.0:

>From 27379bb153377b4a69e9ef8c951012e318b28146 Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Tue, 15 Oct 2024 08:55:54 -0700
Subject: [PATCH 3/6] fixup! revise tests

---
 llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll | 102 +++++++++----------
 1 file changed, 51 insertions(+), 51 deletions(-)

diff --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
index bbaf70eae10fa7..e4d90c242962f4 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -1122,8 +1122,8 @@ 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> @vwmacc_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen %vl) {
-; NOVLOPT-LABEL: vwmacc_vv:
+define <vscale x 4 x i32> @vwmacc_vv_propagate(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen %vl) {
+; NOVLOPT-LABEL: vwmacc_vv_propagate:
 ; NOVLOPT:       # %bb.0:
 ; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
 ; NOVLOPT-NEXT:    vwmacc.vv v8, v10, v11
@@ -1131,7 +1131,7 @@ define <vscale x 4 x i32> @vwmacc_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %
 ; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
 ; NOVLOPT-NEXT:    ret
 ;
-; VLOPT-LABEL: vwmacc_vv:
+; VLOPT-LABEL: vwmacc_vv_propagate:
 ; VLOPT:       # %bb.0:
 ; VLOPT-NEXT:    vsetvli zero, a0, e16, m1, tu, ma
 ; VLOPT-NEXT:    vwmacc.vv v8, v10, v11
@@ -1143,8 +1143,33 @@ define <vscale x 4 x i32> @vwmacc_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %
   ret <vscale x 4 x i32> %2
 }
 
-define <vscale x 4 x i32> @vwmacc_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
-; NOVLOPT-LABEL: vwmacc_vx:
+define <vscale x 4 x i64> @vwmacc_vv_nopropagate(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, iXLen %vl) {
+; NOVLOPT-LABEL: vwmacc_vv_nopropagate:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vmv2r.v v16, v8
+; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmacc.vv v16, v10, v11
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, tu, ma
+; NOVLOPT-NEXT:    vwmacc.vv v12, v16, v8
+; NOVLOPT-NEXT:    vmv4r.v v8, v12
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vwmacc_vv_nopropagate:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vmv2r.v v16, v8
+; VLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmacc.vv v16, v10, v11
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, tu, ma
+; VLOPT-NEXT:    vwmacc.vv v12, v16, v8
+; VLOPT-NEXT:    vmv4r.v v8, v12
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.nxv4i16(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
+  %2 = call <vscale x 4 x i64> @llvm.riscv.vwmacc.nxv4i64.nxv4i32(<vscale x 4 x i64> %d, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl, iXLen 0)
+  ret <vscale x 4 x i64> %2
+}
+
+define <vscale x 4 x i32> @vwmacc_vx_propagate(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
+; NOVLOPT-LABEL: vwmacc_vx_propagate:
 ; NOVLOPT:       # %bb.0:
 ; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
 ; NOVLOPT-NEXT:    vwmacc.vx v8, a0, v10
@@ -1152,7 +1177,7 @@ define <vscale x 4 x i32> @vwmacc_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4
 ; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
 ; NOVLOPT-NEXT:    ret
 ;
-; VLOPT-LABEL: vwmacc_vx:
+; VLOPT-LABEL: vwmacc_vx_propagate:
 ; VLOPT:       # %bb.0:
 ; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
 ; VLOPT-NEXT:    vwmacc.vx v8, a0, v10
@@ -1164,33 +1189,29 @@ define <vscale x 4 x i32> @vwmacc_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4
   ret <vscale x 4 x i32> %2
 }
 
-define <vscale x 4 x i64> @vwmaccu_vv_nopropagate(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, iXLen %vl) {
-; NOVLOPT-LABEL: vwmaccu_vv_nopropagate:
+define <vscale x 4 x i32> @vwmacc_vx_nopropagate(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
+; NOVLOPT-LABEL: vwmacc_vx_nopropagate:
 ; NOVLOPT:       # %bb.0:
-; NOVLOPT-NEXT:    vmv2r.v v16, v8
-; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
-; NOVLOPT-NEXT:    vwmaccu.vv v16, v10, v11
-; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, tu, ma
-; NOVLOPT-NEXT:    vwmaccu.vv v12, v16, v8
-; NOVLOPT-NEXT:    vmv4r.v v8, v12
+; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmacc.vx v8, a0, v10
+; NOVLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmacc.vx v8, a0, v10
 ; NOVLOPT-NEXT:    ret
 ;
-; VLOPT-LABEL: vwmaccu_vv_nopropagate:
+; VLOPT-LABEL: vwmacc_vx_nopropagate:
 ; VLOPT:       # %bb.0:
-; VLOPT-NEXT:    vmv2r.v v16, v8
-; VLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
-; VLOPT-NEXT:    vwmaccu.vv v16, v10, v11
-; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, tu, ma
-; VLOPT-NEXT:    vwmaccu.vv v12, v16, v8
-; VLOPT-NEXT:    vmv4r.v v8, v12
+; VLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmacc.vx v8, a0, v10
+; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmacc.vx v8, a0, v10
 ; VLOPT-NEXT:    ret
-  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.nxv4i16(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
-  %2 = call <vscale x 4 x i64> @llvm.riscv.vwmaccu.nxv4i64.nxv4i32(<vscale x 4 x i64> %d, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl, iXLen 0)
-  ret <vscale x 4 x i64> %2
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.i16(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.i16(<vscale x 4 x i32> %1, i16 %b, <vscale x 4 x i16> %c, iXLen %vl, iXLen 0)
+  ret <vscale x 4 x i32> %2
 }
 
-define <vscale x 4 x i64> @vwmaccu_vv_propagate(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, iXLen %vl) {
-; NOVLOPT-LABEL: vwmaccu_vv_propagate:
+define <vscale x 4 x i64> @vwmaccu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, iXLen %vl) {
+; NOVLOPT-LABEL: vwmaccu_vv:
 ; NOVLOPT:       # %bb.0:
 ; NOVLOPT-NEXT:    vmv2r.v v16, v8
 ; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
@@ -1200,7 +1221,7 @@ define <vscale x 4 x i64> @vwmaccu_vv_propagate(<vscale x 4 x i32> %a, <vscale x
 ; NOVLOPT-NEXT:    vmv4r.v v8, v12
 ; NOVLOPT-NEXT:    ret
 ;
-; VLOPT-LABEL: vwmaccu_vv_propagate:
+; VLOPT-LABEL: vwmaccu_vv:
 ; VLOPT:       # %bb.0:
 ; VLOPT-NEXT:    vmv2r.v v16, v8
 ; VLOPT-NEXT:    vsetvli zero, a0, e16, m1, tu, ma
@@ -1214,29 +1235,8 @@ define <vscale x 4 x i64> @vwmaccu_vv_propagate(<vscale x 4 x i32> %a, <vscale x
   ret <vscale x 4 x i64> %2
 }
 
-define <vscale x 4 x i32> @vwmaccu_vx_nopropagate(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, <vscale x 4 x i16> %d, i16 %e, iXLen %vl) {
-; NOVLOPT-LABEL: vwmaccu_vx_nopropagate:
-; NOVLOPT:       # %bb.0:
-; NOVLOPT-NEXT:    vsetvli a3, zero, e16, m1, tu, ma
-; NOVLOPT-NEXT:    vwmaccu.vx v8, a0, v10
-; NOVLOPT-NEXT:    vsetvli zero, a2, e16, m1, tu, ma
-; NOVLOPT-NEXT:    vwmaccu.vx v8, a1, v11
-; NOVLOPT-NEXT:    ret
-;
-; VLOPT-LABEL: vwmaccu_vx_nopropagate:
-; VLOPT:       # %bb.0:
-; VLOPT-NEXT:    vsetvli a3, zero, e16, m1, tu, ma
-; VLOPT-NEXT:    vwmaccu.vx v8, a0, v10
-; VLOPT-NEXT:    vsetvli zero, a2, e16, m1, tu, ma
-; VLOPT-NEXT:    vwmaccu.vx v8, a1, v11
-; VLOPT-NEXT:    ret
-  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.i16(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
-  %2 = call <vscale x 4 x i32> @llvm.riscv.vwmaccu.nxv4i32.i16(<vscale x 4 x i32> %1, i16 %e, <vscale x 4 x i16> %d, iXLen %vl, iXLen 0)
-  ret <vscale x 4 x i32> %2
-}
-
-define <vscale x 4 x i64> @vwmaccu_vx_propagate(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, i32 %e, iXLen %vl) {
-; NOVLOPT-LABEL: vwmaccu_vx_propagate:
+define <vscale x 4 x i64> @vwmaccu_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, i32 %e, iXLen %vl) {
+; NOVLOPT-LABEL: vwmaccu_vx:
 ; NOVLOPT:       # %bb.0:
 ; NOVLOPT-NEXT:    vsetvli a3, zero, e16, m1, tu, ma
 ; NOVLOPT-NEXT:    vwmaccu.vx v8, a0, v10
@@ -1245,7 +1245,7 @@ define <vscale x 4 x i64> @vwmaccu_vx_propagate(<vscale x 4 x i32> %a, i16 %b, <
 ; NOVLOPT-NEXT:    vmv4r.v v8, v12
 ; NOVLOPT-NEXT:    ret
 ;
-; VLOPT-LABEL: vwmaccu_vx_propagate:
+; VLOPT-LABEL: vwmaccu_vx:
 ; VLOPT:       # %bb.0:
 ; VLOPT-NEXT:    vsetvli zero, a2, e16, m1, tu, ma
 ; VLOPT-NEXT:    vwmaccu.vx v8, a0, v10

>From ed89c46aa5b80a09f5d64b6e2dbd4ea1306b57e4 Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Tue, 15 Oct 2024 11:32:21 -0700
Subject: [PATCH 4/6] fixup! fix nopropagate bug

---
 llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp   |  6 +-
 llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll | 68 ++++----------------
 2 files changed, 15 insertions(+), 59 deletions(-)

diff --git a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
index fbfbc724dc1061..23eedf923f8158 100644
--- a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
@@ -423,7 +423,8 @@ static OperandInfo getOperandInfo(const MachineInstr &MI,
   // Vector Widening Integer Multiply-Add Instructions
   // Destination EEW=2*SEW and EMUL=2*LMUL. Source EEW=SEW and EMUL=LMUL.
   // A SEW-bit*SEW-bit multiply of the sources forms a 2*SEW-bit value, which
-  // is then added to the 2*SEW-bit Dest.
+  // is then added to the 2*SEW-bit Dest. These instructions never have a
+  // passthru operand.
   case RISCV::VWMACCU_VV:
   case RISCV::VWMACCU_VX:
   case RISCV::VWMACC_VV:
@@ -431,8 +432,7 @@ static OperandInfo getOperandInfo(const MachineInstr &MI,
   case RISCV::VWMACCSU_VV:
   case RISCV::VWMACCSU_VX:
   case RISCV::VWMACCUS_VX: {
-    bool IsOp1 = HasPassthru ? MO.getOperandNo() == 2 : MO.getOperandNo() == 1;
-    bool TwoTimes = IsMODef || IsOp1;
+    bool TwoTimes = IsMODef || MO.getOperandNo() == 1;
     unsigned Log2EEW = TwoTimes ? MILog2SEW + 1 : MILog2SEW;
     RISCVII::VLMUL EMUL =
         TwoTimes ? RISCVVType::twoTimesVLMUL(MIVLMul) : MIVLMul;
diff --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
index e4d90c242962f4..25cbb95aab6da5 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -1122,54 +1122,31 @@ 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> @vwmacc_vv_propagate(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen %vl) {
-; NOVLOPT-LABEL: vwmacc_vv_propagate:
+define <vscale x 4 x i64> @vwmacc_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, iXLen %vl) {
+; NOVLOPT-LABEL: vwmacc_vv:
 ; NOVLOPT:       # %bb.0:
 ; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
 ; NOVLOPT-NEXT:    vwmacc.vv v8, v10, v11
-; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
-; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
-; NOVLOPT-NEXT:    ret
-;
-; VLOPT-LABEL: vwmacc_vv_propagate:
-; VLOPT:       # %bb.0:
-; VLOPT-NEXT:    vsetvli zero, a0, e16, m1, tu, ma
-; VLOPT-NEXT:    vwmacc.vv v8, v10, v11
-; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
-; VLOPT-NEXT:    vadd.vv v8, v8, v8
-; VLOPT-NEXT:    ret
-  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.nxv4i16(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
-  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
-  ret <vscale x 4 x i32> %2
-}
-
-define <vscale x 4 x i64> @vwmacc_vv_nopropagate(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, iXLen %vl) {
-; NOVLOPT-LABEL: vwmacc_vv_nopropagate:
-; NOVLOPT:       # %bb.0:
-; NOVLOPT-NEXT:    vmv2r.v v16, v8
-; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
-; NOVLOPT-NEXT:    vwmacc.vv v16, v10, v11
 ; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, tu, ma
-; NOVLOPT-NEXT:    vwmacc.vv v12, v16, v8
+; NOVLOPT-NEXT:    vwmacc.vv v12, v8, v8
 ; NOVLOPT-NEXT:    vmv4r.v v8, v12
 ; NOVLOPT-NEXT:    ret
 ;
-; VLOPT-LABEL: vwmacc_vv_nopropagate:
+; VLOPT-LABEL: vwmacc_vv:
 ; VLOPT:       # %bb.0:
-; VLOPT-NEXT:    vmv2r.v v16, v8
-; VLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
-; VLOPT-NEXT:    vwmacc.vv v16, v10, v11
-; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, tu, ma
-; VLOPT-NEXT:    vwmacc.vv v12, v16, v8
+; VLOPT-NEXT:    vsetvli zero, a0, e16, m1, tu, ma
+; VLOPT-NEXT:    vwmacc.vv v8, v10, v11
+; VLOPT-NEXT:    vsetvli zero, zero, e32, m2, tu, ma
+; VLOPT-NEXT:    vwmacc.vv v12, v8, v8
 ; VLOPT-NEXT:    vmv4r.v v8, v12
 ; VLOPT-NEXT:    ret
   %1 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.nxv4i16(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
-  %2 = call <vscale x 4 x i64> @llvm.riscv.vwmacc.nxv4i64.nxv4i32(<vscale x 4 x i64> %d, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl, iXLen 0)
+  %2 = call <vscale x 4 x i64> @llvm.riscv.vwmacc.nxv4i32.nxv4i16(<vscale x 4 x i64> %d, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl, iXLen 0)
   ret <vscale x 4 x i64> %2
 }
 
-define <vscale x 4 x i32> @vwmacc_vx_propagate(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
-; NOVLOPT-LABEL: vwmacc_vx_propagate:
+define <vscale x 4 x i32> @vwmacc_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
+; NOVLOPT-LABEL: vwmacc_vx:
 ; NOVLOPT:       # %bb.0:
 ; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
 ; NOVLOPT-NEXT:    vwmacc.vx v8, a0, v10
@@ -1177,7 +1154,7 @@ define <vscale x 4 x i32> @vwmacc_vx_propagate(<vscale x 4 x i32> %a, i16 %b, <v
 ; NOVLOPT-NEXT:    vadd.vv v8, v8, v8
 ; NOVLOPT-NEXT:    ret
 ;
-; VLOPT-LABEL: vwmacc_vx_propagate:
+; VLOPT-LABEL: vwmacc_vx:
 ; VLOPT:       # %bb.0:
 ; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
 ; VLOPT-NEXT:    vwmacc.vx v8, a0, v10
@@ -1189,27 +1166,6 @@ define <vscale x 4 x i32> @vwmacc_vx_propagate(<vscale x 4 x i32> %a, i16 %b, <v
   ret <vscale x 4 x i32> %2
 }
 
-define <vscale x 4 x i32> @vwmacc_vx_nopropagate(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen %vl) {
-; NOVLOPT-LABEL: vwmacc_vx_nopropagate:
-; NOVLOPT:       # %bb.0:
-; NOVLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
-; NOVLOPT-NEXT:    vwmacc.vx v8, a0, v10
-; NOVLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
-; NOVLOPT-NEXT:    vwmacc.vx v8, a0, v10
-; NOVLOPT-NEXT:    ret
-;
-; VLOPT-LABEL: vwmacc_vx_nopropagate:
-; VLOPT:       # %bb.0:
-; VLOPT-NEXT:    vsetvli a2, zero, e16, m1, tu, ma
-; VLOPT-NEXT:    vwmacc.vx v8, a0, v10
-; VLOPT-NEXT:    vsetvli zero, a1, e16, m1, tu, ma
-; VLOPT-NEXT:    vwmacc.vx v8, a0, v10
-; VLOPT-NEXT:    ret
-  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.i16(<vscale x 4 x i32> %a, i16 %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
-  %2 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.i16(<vscale x 4 x i32> %1, i16 %b, <vscale x 4 x i16> %c, iXLen %vl, iXLen 0)
-  ret <vscale x 4 x i32> %2
-}
-
 define <vscale x 4 x i64> @vwmaccu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, <vscale x 4 x i64> %d, iXLen %vl) {
 ; NOVLOPT-LABEL: vwmaccu_vv:
 ; NOVLOPT:       # %bb.0:

>From 81ce6dabafc5940db6053606061245968a794dc8 Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Tue, 15 Oct 2024 11:36:00 -0700
Subject: [PATCH 5/6] fixup! add test that we bail out on tied def

---
 llvm/test/CodeGen/RISCV/rvv/vl-opt.ll | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt.ll
index 0b3e67ec895566..677f01de44143c 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt.ll
@@ -111,6 +111,19 @@ define <vscale x 4 x i32> @different_imm_vl_with_tu(<vscale x 4 x i32> %passthru
   ret <vscale x 4 x i32> %w
 }
 
+define <vscale x 4 x i32> @dont_optimize_tied_def(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen %vl) {
+; CHECK-LABEL: dont_optimize_tied_def:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
+; CHECK-NEXT:    vwmacc.vv v8, v10, v11
+; CHECK-NEXT:    vsetvli zero, a0, e16, m1, tu, ma
+; CHECK-NEXT:    vwmacc.vv v8, v10, v11
+; CHECK-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.nxv4i16(<vscale x 4 x i32> %a, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen -1, iXLen 0)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vwmacc.nxv4i32.nxv4i16(<vscale x 4 x i32> %1, <vscale x 4 x i16> %b, <vscale x 4 x i16> %c, iXLen %vl, iXLen 0)
+  ret <vscale x 4 x i32> %2
+}
+
 ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
 ; NOVLOPT: {{.*}}
 ; VLOPT: {{.*}}

>From e456af859f485bd3d938aeb426682cb8b6946170 Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Tue, 15 Oct 2024 15:42:24 -0700
Subject: [PATCH 6/6] fixup! relocate to widening section

---
 llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp | 31 +++++++++++-----------
 1 file changed, 16 insertions(+), 15 deletions(-)

diff --git a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
index 23eedf923f8158..8868e3b7395adb 100644
--- a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
@@ -404,7 +404,19 @@ static OperandInfo getOperandInfo(const MachineInstr &MI,
   case RISCV::VWMULSU_VV:
   case RISCV::VWMULSU_VX:
   case RISCV::VWMULU_VV:
-  case RISCV::VWMULU_VX: {
+  case RISCV::VWMULU_VX:
+  // Vector Widening Integer Multiply-Add Instructions
+  // Destination EEW=2*SEW and EMUL=2*LMUL. Source EEW=SEW and EMUL=LMUL.
+  // A SEW-bit*SEW-bit multiply of the sources forms a 2*SEW-bit value, which
+  // is then added to the 2*SEW-bit Dest. These instructions never have a
+  // passthru operand.
+  case RISCV::VWMACCU_VV:
+  case RISCV::VWMACCU_VX:
+  case RISCV::VWMACC_VV:
+  case RISCV::VWMACC_VX:
+  case RISCV::VWMACCSU_VV:
+  case RISCV::VWMACCSU_VX:
+  case RISCV::VWMACCUS_VX: {
     unsigned Log2EEW = IsMODef ? MILog2SEW + 1 : MILog2SEW;
     RISCVII::VLMUL EMUL =
         IsMODef ? RISCVVType::twoTimesVLMUL(MIVLMul) : MIVLMul;
@@ -419,20 +431,9 @@ static OperandInfo getOperandInfo(const MachineInstr &MI,
   case RISCV::VWADD_WV:
   case RISCV::VWADD_WX:
   case RISCV::VWSUB_WV:
-  case RISCV::VWSUB_WX:
-  // Vector Widening Integer Multiply-Add Instructions
-  // Destination EEW=2*SEW and EMUL=2*LMUL. Source EEW=SEW and EMUL=LMUL.
-  // A SEW-bit*SEW-bit multiply of the sources forms a 2*SEW-bit value, which
-  // is then added to the 2*SEW-bit Dest. These instructions never have a
-  // passthru operand.
-  case RISCV::VWMACCU_VV:
-  case RISCV::VWMACCU_VX:
-  case RISCV::VWMACC_VV:
-  case RISCV::VWMACC_VX:
-  case RISCV::VWMACCSU_VV:
-  case RISCV::VWMACCSU_VX:
-  case RISCV::VWMACCUS_VX: {
-    bool TwoTimes = IsMODef || MO.getOperandNo() == 1;
+  case RISCV::VWSUB_WX: {
+    bool IsOp1 = HasPassthru ? MO.getOperandNo() == 2 : MO.getOperandNo() == 1;
+    bool TwoTimes = IsMODef || IsOp1;
     unsigned Log2EEW = TwoTimes ? MILog2SEW + 1 : MILog2SEW;
     RISCVII::VLMUL EMUL =
         TwoTimes ? RISCVVType::twoTimesVLMUL(MIVLMul) : MIVLMul;



More information about the llvm-commits mailing list