[llvm] 131b7fe - [RISCV][VLOPT] Add support for widening integer mul-add instructions (#112219)

via llvm-commits llvm-commits at lists.llvm.org
Fri Dec 6 12:03:46 PST 2024


Author: Michael Maitland
Date: 2024-12-06T15:03:43-05:00
New Revision: 131b7fe2b108fb33efd031371f0f8a993e374860

URL: https://github.com/llvm/llvm-project/commit/131b7fe2b108fb33efd031371f0f8a993e374860
DIFF: https://github.com/llvm/llvm-project/commit/131b7fe2b108fb33efd031371f0f8a993e374860.diff

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

This adds support for these instructions and also tests getOperandInfo
for these instructions as well. I think the VL on the using add
instruction can be optimized further, once we add support for optimizing
non-vlmax.

Added: 
    

Modified: 
    llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
    llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
    llvm/test/CodeGen/RISCV/rvv/vl-opt.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
index d4c1dd7dc549bf..e75eddb423e4b9 100644
--- a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
@@ -403,7 +403,19 @@ 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;
@@ -418,18 +430,7 @@ 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.
-  // 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.
-  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: {
+  case RISCV::VWSUB_WX: {
     bool IsOp1 = HasPassthru ? MO.getOperandNo() == 2 : MO.getOperandNo() == 1;
     bool TwoTimes = IsMODef || IsOp1;
     unsigned Log2EEW = TwoTimes ? MILog2SEW + 1 : MILog2SEW;
@@ -571,9 +572,13 @@ 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 11f603b56b6e56..39cc90b812f99e 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -1248,44 +1248,149 @@ define <vscale x 4 x i64> @vwmulu_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
   ret <vscale x 4 x i64> %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, <vscale x 4 x i32> %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, v12
+; 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, 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 i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %d, 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, <vscale x 4 x i32> %d, 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, v12
+; 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, 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> %d, 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, <vscale x 4 x i32> %d, i32 %e, 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 zero, a1, e32, m2, ta, ma
-; NOVLOPT-NEXT:    vadd.vv v8, v10, v10
+; NOVLOPT-NEXT:    vsetvli a1, zero, e16, m1, tu, ma
+; NOVLOPT-NEXT:    vwmaccu.vx v8, a0, v10
+; NOVLOPT-NEXT:    vsetvli zero, a2, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v12
 ; 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, a2, 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, 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 i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %d, 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, v10, v10
+; 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
 }

diff  --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt.ll
index ff4f3e24ec17b0..1efe96c975754a 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt.ll
@@ -136,3 +136,17 @@ define <vscale x 4 x i32> @
diff erent_imm_vl_with_tu(<vscale x 4 x i32> %passthru
   %w = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> %passthru, <vscale x 4 x i32> %v, <vscale x 4 x i32> %a,iXLen 4)
   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
+}
+


        


More information about the llvm-commits mailing list