[RISCV][VLOPT] Add support for widening integer mul-add instructions #112219

michaelmaitland · 2024-10-14T14:49:10Z

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.

llvmbot · 2024-10-14T14:49:31Z

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

Author: Michael Maitland (michaelmaitland)

Changes

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.

Full diff: https://github.com/llvm/llvm-project/pull/112219.diff

2 Files Affected:

(modified) llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp (+8-4)
(modified) llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll (+121-16)

diff --git a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
index 90af9ef898d951..7c5ce23ad72fc3 100644
--- a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
@@ -421,8 +421,8 @@ static OperandInfo getOperandInfo(const MachineInstr &MI,
   case RISCV::VWSUB_WX:
   // 11.14. 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:
@@ -552,9 +552,13 @@ static bool isSupportedInstr(const MachineInstr &MI) {
   // 11.13. Vector Single-Width Integer Multiply-Add Instructions
   // FIXME: Add support for 11.13 instructions
   // 11.14. Vector Widening Integer Multiply-Add Instructions
-  // FIXME: Add support for 11.14 instructions
-  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:
   // 11.15. Vector Integer Merge Instructions
   // FIXME: Add support for 11.15 instructions
   // 11.16. 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 107252338829bd..7be3964f2ae3c6 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -804,44 +804,149 @@ define <vscale x 4 x i32> @vmulhsu_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
 }

llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll

topperc · 2024-10-14T18:50:24Z

This does not test getOperandInfo for the multiplicands. Should we have tests where these instructions are the ones providing the VL?

michaelmaitland · 2024-10-14T19:42:02Z

I have updated the tests to consider the sources.

The tests are generated stacked on #112244. This PR must be rebased on #112244 so it can pass tests.

llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll

llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp

…tions

lukel97 · 2024-10-28T09:53:19Z

llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll

+; 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)


Nit, should the user instruction for this test and the tests above also be vadd.vv for consistency with the other added tests?

…lvm#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.

michaelmaitland added the backend:RISC-V label Oct 14, 2024

michaelmaitland requested review from preames, topperc, luke957 and wangpc-pp October 14, 2024 14:49

michaelmaitland commented Oct 14, 2024

View reviewed changes

llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll Show resolved Hide resolved

michaelmaitland changed the title ~~[RISCV][VLOPT] Add support for 11.14 widening integer mul-add instructions~~ [RISCV][VLOPT] Add support for widening integer mul-add instructions Oct 15, 2024

michaelmaitland force-pushed the vlopt-wide-int-mul-add branch 2 times, most recently from 67941c2 to 27379bb Compare October 15, 2024 15:58

topperc reviewed Oct 15, 2024

View reviewed changes

llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll Outdated Show resolved Hide resolved

topperc reviewed Oct 15, 2024

View reviewed changes

llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp Outdated Show resolved Hide resolved

topperc reviewed Oct 15, 2024

View reviewed changes

llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp Show resolved Hide resolved

[RISCV][VLOPT] Add support for 11.14 widening integer mul-add instruc…

4cb75a2

…tions

michaelmaitland force-pushed the vlopt-wide-int-mul-add branch from e456af8 to 4cb75a2 Compare October 16, 2024 19:21

michaelmaitland requested review from lukel97 and topperc October 17, 2024 16:42

lukel97 approved these changes Oct 28, 2024

View reviewed changes

fixup! use vadd in tests

cbe89ca

michaelmaitland merged commit 131b7fe into llvm:main Dec 6, 2024
5 of 7 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[RISCV][VLOPT] Add support for widening integer mul-add instructions #112219

[RISCV][VLOPT] Add support for widening integer mul-add instructions #112219

michaelmaitland commented Oct 14, 2024

llvmbot commented Oct 14, 2024

topperc commented Oct 14, 2024

michaelmaitland commented Oct 14, 2024

lukel97 Oct 28, 2024

michaelmaitland Dec 6, 2024

[RISCV][VLOPT] Add support for widening integer mul-add instructions #112219

[RISCV][VLOPT] Add support for widening integer mul-add instructions #112219

Conversation

michaelmaitland commented Oct 14, 2024

llvmbot commented Oct 14, 2024

topperc commented Oct 14, 2024

michaelmaitland commented Oct 14, 2024

lukel97 Oct 28, 2024

Choose a reason for hiding this comment

michaelmaitland Dec 6, 2024

Choose a reason for hiding this comment