[RISCV] Rematerialize vfmv.v.f #108007
Merged
[RISCV] Rematerialize vfmv.v.f #108007
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
|
@llvm/pr-subscribers-backend-risc-v Author: Luke Lau (lukel97) ChangesThis is the same principle as vmv.v.x in #107993, but for floats. Full diff: https://github.com/llvm/llvm-project/pull/108007.diff 3 Files Affected:
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
index 325a50c9f48a1c..4d9bb401995fd9 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
@@ -169,6 +169,7 @@ Register RISCVInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
bool RISCVInstrInfo::isReallyTriviallyReMaterializable(
const MachineInstr &MI) const {
switch (RISCV::getRVVMCOpcode(MI.getOpcode())) {
+ case RISCV::VFMV_V_F:
case RISCV::VMV_V_I:
case RISCV::VID_V:
if (MI.getOperand(1).isUndef() &&
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
index e11f176bfe6041..572cf158e9c77f 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
@@ -6557,6 +6557,7 @@ defm PseudoVFMERGE : VPseudoVMRG_FM;
//===----------------------------------------------------------------------===//
// 13.16. Vector Floating-Point Move Instruction
//===----------------------------------------------------------------------===//
+let isReMaterializable = 1 in
defm PseudoVFMV_V : VPseudoVMV_F;
//===----------------------------------------------------------------------===//
diff --git a/llvm/test/CodeGen/RISCV/rvv/remat.ll b/llvm/test/CodeGen/RISCV/rvv/remat.ll
index 2b12249378eb1f..afaab9214fe92b 100644
--- a/llvm/test/CodeGen/RISCV/rvv/remat.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/remat.ll
@@ -171,3 +171,68 @@ define void @vmv.v.i(ptr %p) {
store volatile <vscale x 8 x i64> %vmv.v.i, ptr %p
ret void
}
+
+define void @vfmv.v.f(ptr %p, double %x) {
+; POSTRA-LABEL: vfmv.v.f:
+; POSTRA: # %bb.0:
+; POSTRA-NEXT: vsetvli a1, zero, e64, m8, ta, ma
+; POSTRA-NEXT: vfmv.v.f v8, fa0
+; POSTRA-NEXT: vs8r.v v8, (a0)
+; POSTRA-NEXT: vl8re64.v v16, (a0)
+; POSTRA-NEXT: vl8re64.v v24, (a0)
+; POSTRA-NEXT: vl8re64.v v0, (a0)
+; POSTRA-NEXT: vl8re64.v v8, (a0)
+; POSTRA-NEXT: vs8r.v v8, (a0)
+; POSTRA-NEXT: vs8r.v v0, (a0)
+; POSTRA-NEXT: vs8r.v v24, (a0)
+; POSTRA-NEXT: vs8r.v v16, (a0)
+; POSTRA-NEXT: vfmv.v.f v8, fa0
+; POSTRA-NEXT: vs8r.v v8, (a0)
+; POSTRA-NEXT: fsd fa0, 0(a0)
+; POSTRA-NEXT: ret
+;
+; PRERA-LABEL: vfmv.v.f:
+; PRERA: # %bb.0:
+; PRERA-NEXT: addi sp, sp, -16
+; PRERA-NEXT: .cfi_def_cfa_offset 16
+; PRERA-NEXT: csrr a1, vlenb
+; PRERA-NEXT: slli a1, a1, 3
+; PRERA-NEXT: sub sp, sp, a1
+; PRERA-NEXT: .cfi_escape 0x0f, 0x0d, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0x08, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 8 * vlenb
+; PRERA-NEXT: vsetvli a1, zero, e64, m8, ta, ma
+; PRERA-NEXT: vfmv.v.f v8, fa0
+; PRERA-NEXT: vs8r.v v8, (a0)
+; PRERA-NEXT: vl8re64.v v16, (a0)
+; PRERA-NEXT: addi a1, sp, 16
+; PRERA-NEXT: vs8r.v v16, (a1) # Unknown-size Folded Spill
+; PRERA-NEXT: vl8re64.v v24, (a0)
+; PRERA-NEXT: vl8re64.v v0, (a0)
+; PRERA-NEXT: vl8re64.v v16, (a0)
+; PRERA-NEXT: vs8r.v v16, (a0)
+; PRERA-NEXT: vs8r.v v0, (a0)
+; PRERA-NEXT: vs8r.v v24, (a0)
+; PRERA-NEXT: vl8r.v v16, (a1) # Unknown-size Folded Reload
+; PRERA-NEXT: vs8r.v v16, (a0)
+; PRERA-NEXT: vs8r.v v8, (a0)
+; PRERA-NEXT: fsd fa0, 0(a0)
+; PRERA-NEXT: csrr a0, vlenb
+; PRERA-NEXT: slli a0, a0, 3
+; PRERA-NEXT: add sp, sp, a0
+; PRERA-NEXT: addi sp, sp, 16
+; PRERA-NEXT: ret
+ %vfmv.v.f = call <vscale x 8 x double> @llvm.riscv.vfmv.v.f.nxv8f64(<vscale x 8 x double> poison, double %x, i64 -1)
+ store volatile <vscale x 8 x double> %vfmv.v.f, ptr %p
+
+ %a = load volatile <vscale x 8 x double>, ptr %p
+ %b = load volatile <vscale x 8 x double>, ptr %p
+ %c = load volatile <vscale x 8 x double>, ptr %p
+ %d = load volatile <vscale x 8 x double>, ptr %p
+ store volatile <vscale x 8 x double> %d, ptr %p
+ store volatile <vscale x 8 x double> %c, ptr %p
+ store volatile <vscale x 8 x double> %b, ptr %p
+ store volatile <vscale x 8 x double> %a, ptr %p
+
+ store volatile <vscale x 8 x double> %vfmv.v.f, ptr %p
+ store volatile double %x, ptr %p
+ ret void
+}
|
lukel97
added a commit
to lukel97/llvm-project
that referenced
this pull request
Sep 10, 2024
Continuing with llvm#107993 and llvm#108007, this handles the last of the main rematerializable vector instructions. Program regalloc.NumSpills regalloc.NumReloads regalloc.NumRemats lhs rhs diff lhs rhs diff lhs rhs diff 508.namd_r 6598.00 6598.00 0.0% 15509.00 15509.00 0.0% 2387.00 2387.00 0.0% 505.mcf_r 141.00 141.00 0.0% 372.00 372.00 0.0% 36.00 36.00 0.0% 641.leela_s 356.00 356.00 0.0% 525.00 525.00 0.0% 117.00 117.00 0.0% 631.deepsjeng_s 353.00 353.00 0.0% 682.00 682.00 0.0% 124.00 124.00 0.0% 623.xalancbmk_s 1548.00 1548.00 0.0% 2466.00 2466.00 0.0% 620.00 620.00 0.0% 620.omnetpp_s 946.00 946.00 0.0% 1485.00 1485.00 0.0% 1178.00 1178.00 0.0% 605.mcf_s 141.00 141.00 0.0% 372.00 372.00 0.0% 36.00 36.00 0.0% 557.xz_r 289.00 289.00 0.0% 505.00 505.00 0.0% 172.00 172.00 0.0% 541.leela_r 356.00 356.00 0.0% 525.00 525.00 0.0% 117.00 117.00 0.0% 531.deepsjeng_r 353.00 353.00 0.0% 682.00 682.00 0.0% 124.00 124.00 0.0% 520.omnetpp_r 946.00 946.00 0.0% 1485.00 1485.00 0.0% 1178.00 1178.00 0.0% 523.xalancbmk_r 1548.00 1548.00 0.0% 2466.00 2466.00 0.0% 620.00 620.00 0.0% 619.lbm_s 68.00 68.00 0.0% 70.00 70.00 0.0% 1.00 1.00 0.0% 519.lbm_r 73.00 73.00 0.0% 75.00 75.00 0.0% 1.00 1.00 0.0% 657.xz_s 289.00 289.00 0.0% 505.00 505.00 0.0% 172.00 172.00 0.0% 511.povray_r 1937.00 1936.00 -0.1% 3629.00 3628.00 -0.0% 517.00 518.00 0.2% 502.gcc_r 12450.00 12442.00 -0.1% 27328.00 27317.00 -0.0% 9409.00 9409.00 0.0% 602.gcc_s 12450.00 12442.00 -0.1% 27328.00 27317.00 -0.0% 9409.00 9409.00 0.0% 638.imagick_s 4181.00 4178.00 -0.1% 11342.00 11338.00 -0.0% 3366.00 3368.00 0.1% 538.imagick_r 4181.00 4178.00 -0.1% 11342.00 11338.00 -0.0% 3366.00 3368.00 0.1% 500.perlbench_r 4178.00 4175.00 -0.1% 9162.00 9159.00 -0.0% 2410.00 2410.00 0.0% 600.perlbench_s 4178.00 4175.00 -0.1% 9162.00 9159.00 -0.0% 2410.00 2410.00 0.0% 525.x264_r 1886.00 1884.00 -0.1% 4561.00 4559.00 -0.0% 471.00 471.00 0.0% 625.x264_s 1886.00 1884.00 -0.1% 4561.00 4559.00 -0.0% 471.00 471.00 0.0% 510.parest_r 42740.00 42689.00 -0.1% 82400.00 82252.00 -0.2% 5612.00 5620.00 0.1% 644.nab_s 753.00 752.00 -0.1% 1183.00 1182.00 -0.1% 318.00 318.00 0.0% 544.nab_r 753.00 752.00 -0.1% 1183.00 1182.00 -0.1% 318.00 318.00 0.0% 526.blender_r 13105.00 13084.00 -0.2% 26478.00 26442.00 -0.1% 18991.00 18989.00 -0.0% Geomean difference -0.0% -0.0% 0.0% There's an extra spill in one of the test cases, but it's likely noise from the spill weights and isn't an issue in practice.
This is the same principle as vmv.v.x in llvm#107993, but for floats. Program regalloc.NumSpills regalloc.NumReloads regalloc.NumRemats lhs rhs diff lhs rhs diff lhs rhs diff 519.lbm_r 73.00 73.00 0.0% 75.00 75.00 0.0% 1.00 1.00 0.0% 544.nab_r 753.00 753.00 0.0% 1183.00 1183.00 0.0% 318.00 318.00 0.0% 619.lbm_s 68.00 68.00 0.0% 70.00 70.00 0.0% 1.00 1.00 0.0% 644.nab_s 753.00 753.00 0.0% 1183.00 1183.00 0.0% 318.00 318.00 0.0% 508.namd_r 6598.00 6597.00 -0.0% 15509.00 15503.00 -0.0% 2387.00 2393.00 0.3% 526.blender_r 13105.00 13084.00 -0.2% 26478.00 26443.00 -0.1% 18991.00 18996.00 0.0% 510.parest_r 42740.00 42665.00 -0.2% 82400.00 82309.00 -0.1% 5612.00 5648.00 0.6% 511.povray_r 1937.00 1929.00 -0.4% 3629.00 3620.00 -0.2% 517.00 525.00 1.5% 538.imagick_r 4181.00 4150.00 -0.7% 11342.00 11125.00 -1.9% 3366.00 3366.00 0.0% 638.imagick_s 4181.00 4150.00 -0.7% 11342.00 11125.00 -1.9% 3366.00 3366.00 0.0% Geomean difference -0.2% -0.4% 0.2%
lukel97
force-pushed
the
vector-remat/vfmv.v.f
branch
from
287fba4 to
b19c5df
Compare
lukel97
added a commit
to lukel97/llvm-project
that referenced
this pull request
Sep 11, 2024
Continuing with llvm#107993 and llvm#108007, this handles the last of the main rematerializable vector instructions. Program regalloc.NumSpills regalloc.NumReloads regalloc.NumRemats lhs rhs diff lhs rhs diff lhs rhs diff 508.namd_r 6598.00 6598.00 0.0% 15509.00 15509.00 0.0% 2387.00 2387.00 0.0% 505.mcf_r 141.00 141.00 0.0% 372.00 372.00 0.0% 36.00 36.00 0.0% 641.leela_s 356.00 356.00 0.0% 525.00 525.00 0.0% 117.00 117.00 0.0% 631.deepsjeng_s 353.00 353.00 0.0% 682.00 682.00 0.0% 124.00 124.00 0.0% 623.xalancbmk_s 1548.00 1548.00 0.0% 2466.00 2466.00 0.0% 620.00 620.00 0.0% 620.omnetpp_s 946.00 946.00 0.0% 1485.00 1485.00 0.0% 1178.00 1178.00 0.0% 605.mcf_s 141.00 141.00 0.0% 372.00 372.00 0.0% 36.00 36.00 0.0% 557.xz_r 289.00 289.00 0.0% 505.00 505.00 0.0% 172.00 172.00 0.0% 541.leela_r 356.00 356.00 0.0% 525.00 525.00 0.0% 117.00 117.00 0.0% 531.deepsjeng_r 353.00 353.00 0.0% 682.00 682.00 0.0% 124.00 124.00 0.0% 520.omnetpp_r 946.00 946.00 0.0% 1485.00 1485.00 0.0% 1178.00 1178.00 0.0% 523.xalancbmk_r 1548.00 1548.00 0.0% 2466.00 2466.00 0.0% 620.00 620.00 0.0% 619.lbm_s 68.00 68.00 0.0% 70.00 70.00 0.0% 1.00 1.00 0.0% 519.lbm_r 73.00 73.00 0.0% 75.00 75.00 0.0% 1.00 1.00 0.0% 657.xz_s 289.00 289.00 0.0% 505.00 505.00 0.0% 172.00 172.00 0.0% 511.povray_r 1937.00 1936.00 -0.1% 3629.00 3628.00 -0.0% 517.00 518.00 0.2% 502.gcc_r 12450.00 12442.00 -0.1% 27328.00 27317.00 -0.0% 9409.00 9409.00 0.0% 602.gcc_s 12450.00 12442.00 -0.1% 27328.00 27317.00 -0.0% 9409.00 9409.00 0.0% 638.imagick_s 4181.00 4178.00 -0.1% 11342.00 11338.00 -0.0% 3366.00 3368.00 0.1% 538.imagick_r 4181.00 4178.00 -0.1% 11342.00 11338.00 -0.0% 3366.00 3368.00 0.1% 500.perlbench_r 4178.00 4175.00 -0.1% 9162.00 9159.00 -0.0% 2410.00 2410.00 0.0% 600.perlbench_s 4178.00 4175.00 -0.1% 9162.00 9159.00 -0.0% 2410.00 2410.00 0.0% 525.x264_r 1886.00 1884.00 -0.1% 4561.00 4559.00 -0.0% 471.00 471.00 0.0% 625.x264_s 1886.00 1884.00 -0.1% 4561.00 4559.00 -0.0% 471.00 471.00 0.0% 510.parest_r 42740.00 42689.00 -0.1% 82400.00 82252.00 -0.2% 5612.00 5620.00 0.1% 644.nab_s 753.00 752.00 -0.1% 1183.00 1182.00 -0.1% 318.00 318.00 0.0% 544.nab_r 753.00 752.00 -0.1% 1183.00 1182.00 -0.1% 318.00 318.00 0.0% 526.blender_r 13105.00 13084.00 -0.2% 26478.00 26442.00 -0.1% 18991.00 18989.00 -0.0% Geomean difference -0.0% -0.0% 0.0% There's an extra spill in one of the test cases, but it's likely noise from the spill weights and isn't an issue in practice.
lukel97
added a commit
that referenced
this pull request
Sep 11, 2024
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.
This is the same principle as vmv.v.x in #107993, but for floats.