[RISCV] Refine known bits for READ_VLENB

This implements known bits for READ_VALUE using any information known about minimum and maximum VLEN. There's an additional assumption that VLEN is a power of two.

The motivation here is mostly to remove the last use of getMinVLen, but while I was here, I decided to also fix the bug for VLEN < 128 and handle max from command line generically too.

Differential Revision: https://reviews.llvm.org/D128758

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 27fbc93..9772f81 100644 (file)
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -9388,14 +9388,15 @@ void RISCVTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
     break;
   }
   case RISCVISD::READ_VLENB: {
-    // If we know the minimum VLen from Zvl extensions, we can use that to
-    // determine the trailing zeros of VLENB.
-    // FIXME: Limit to 128 bit vectors until we have more testing.
-    unsigned MinVLenB = std::min(128U, Subtarget.getMinVLen()) / 8;
-    if (MinVLenB > 0)
-      Known.Zero.setLowBits(Log2_32(MinVLenB));
-    // We assume VLENB is no more than 65536 / 8 bytes.
-    Known.Zero.setBitsFrom(14);
+    // We can use the minimum and maximum VLEN values to bound VLENB.  We
+    // know VLEN must be a power of two.
+    const unsigned MinVLenB = Subtarget.getRealMinVLen() / 8;
+    const unsigned MaxVLenB = Subtarget.getRealMaxVLen() / 8;
+    assert(MinVLenB > 0 && "READ_VLENB without vector extension enabled?");
+    Known.Zero.setLowBits(Log2_32(MinVLenB));
+    Known.Zero.setBitsFrom(Log2_32(MaxVLenB)+1);
+    if (MaxVLenB == MinVLenB)
+      Known.One.setBit(Log2_32(MinVLenB));
     break;
   }
   case ISD::INTRINSIC_W_CHAIN:

diff --git a/llvm/test/CodeGen/RISCV/rvv/rvv-vscale.i64.ll b/llvm/test/CodeGen/RISCV/rvv/rvv-vscale.i64.ll
index 82f671f..5cf12ac 100644 (file)
--- a/llvm/test/CodeGen/RISCV/rvv/rvv-vscale.i64.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/rvv-vscale.i64.ll
@@ -1,8 +1,14 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple riscv64 -mattr=+m,+v < %s \
-; RUN:    | FileCheck %s -check-prefix=RV64
+; RUN:    | FileCheck %s -check-prefixes=RV64,RV64-VLENUNK
 ; RUN: llc -mtriple riscv32 -mattr=+m,+v < %s \
 ; RUN:    | FileCheck %s -check-prefix=RV32
+; RUN: llc -mtriple riscv64 -mattr=+m,+v -riscv-v-vector-bits-min=256 < %s \
+; RUN:    | FileCheck %s -check-prefixes=RV64,RV64-VLEN256MIN
+; RUN: llc -mtriple riscv64 -mattr=+m,+v -riscv-v-vector-bits-max=256 < %s \
+; RUN:    | FileCheck %s -check-prefixes=RV64,RV64-VLEN256MAX
+; RUN: llc -mtriple riscv64 -mattr=+m,+v -riscv-v-vector-bits-min=256 -riscv-v-vector-bits-max=256 < %s \
+; RUN:    | FileCheck %s -check-prefixes=RV64-VLEN256EXACT
 
 
 define i64 @vscale_zero() nounwind {
@@ -16,6 +22,11 @@ define i64 @vscale_zero() nounwind {
 ; RV32-NEXT:    li a0, 0
 ; RV32-NEXT:    li a1, 0
 ; RV32-NEXT:    ret
+;
+; RV64-VLEN256EXACT-LABEL: vscale_zero:
+; RV64-VLEN256EXACT:       # %bb.0: # %entry
+; RV64-VLEN256EXACT-NEXT:    li a0, 0
+; RV64-VLEN256EXACT-NEXT:    ret
 entry:
   %0 = call i64 @llvm.vscale.i64()
   %1 = mul i64 %0, 0
@@ -35,6 +46,11 @@ define i64 @vscale_one() nounwind {
 ; RV32-NEXT:    srli a0, a0, 3
 ; RV32-NEXT:    li a1, 0
 ; RV32-NEXT:    ret
+;
+; RV64-VLEN256EXACT-LABEL: vscale_one:
+; RV64-VLEN256EXACT:       # %bb.0: # %entry
+; RV64-VLEN256EXACT-NEXT:    li a0, 4
+; RV64-VLEN256EXACT-NEXT:    ret
 entry:
   %0 = call i64 @llvm.vscale.i64()
   %1 = mul i64 %0, 1
@@ -54,6 +70,11 @@ define i64 @vscale_uimmpow2xlen() nounwind {
 ; RV32-NEXT:    slli a0, a0, 3
 ; RV32-NEXT:    li a1, 0
 ; RV32-NEXT:    ret
+;
+; RV64-VLEN256EXACT-LABEL: vscale_uimmpow2xlen:
+; RV64-VLEN256EXACT:       # %bb.0: # %entry
+; RV64-VLEN256EXACT-NEXT:    li a0, 256
+; RV64-VLEN256EXACT-NEXT:    ret
 entry:
   %0 = call i64 @llvm.vscale.i64()
   %1 = mul i64 %0, 64
@@ -75,6 +96,11 @@ define i64 @vscale_non_pow2() nounwind {
 ; RV32-NEXT:    add a0, a1, a0
 ; RV32-NEXT:    li a1, 0
 ; RV32-NEXT:    ret
+;
+; RV64-VLEN256EXACT-LABEL: vscale_non_pow2:
+; RV64-VLEN256EXACT:       # %bb.0: # %entry
+; RV64-VLEN256EXACT-NEXT:    li a0, 96
+; RV64-VLEN256EXACT-NEXT:    ret
 entry:
   %0 = call i64 @llvm.vscale.i64()
   %1 = mul i64 %0, 24
@@ -99,6 +125,11 @@ define i64 @vscale_select(i32 %x, i32 %y) {
 ; RV32-NEXT:    srli a0, a0, 3
 ; RV32-NEXT:    li a1, 0
 ; RV32-NEXT:    ret
+;
+; RV64-VLEN256EXACT-LABEL: vscale_select:
+; RV64-VLEN256EXACT:       # %bb.0:
+; RV64-VLEN256EXACT-NEXT:    li a0, 4
+; RV64-VLEN256EXACT-NEXT:    ret
   %a = call i64 @llvm.vscale.i64()
   %b = and i64 %a, 4294967295
   %c = icmp eq i32 %x, %y
@@ -119,6 +150,11 @@ define i64 @vscale_high_bits_zero() nounwind {
 ; RV32-NEXT:    srli a0, a0, 3
 ; RV32-NEXT:    li a1, 0
 ; RV32-NEXT:    ret
+;
+; RV64-VLEN256EXACT-LABEL: vscale_high_bits_zero:
+; RV64-VLEN256EXACT:       # %bb.0: # %entry
+; RV64-VLEN256EXACT-NEXT:    li a0, 4
+; RV64-VLEN256EXACT-NEXT:    ret
 entry:
   %0 = call i64 @llvm.vscale.i64()
   %1 = and i64 %0, 2047
@@ -126,12 +162,12 @@ entry:
 }
 
 define i64 @vscale_masked() nounwind {
-; RV64-LABEL: vscale_masked:
-; RV64:       # %bb.0: # %entry
-; RV64-NEXT:    csrr a0, vlenb
-; RV64-NEXT:    srli a0, a0, 3
-; RV64-NEXT:    andi a0, a0, 510
-; RV64-NEXT:    ret
+; RV64-VLENUNK-LABEL: vscale_masked:
+; RV64-VLENUNK:       # %bb.0: # %entry
+; RV64-VLENUNK-NEXT:    csrr a0, vlenb
+; RV64-VLENUNK-NEXT:    srli a0, a0, 3
+; RV64-VLENUNK-NEXT:    andi a0, a0, 510
+; RV64-VLENUNK-NEXT:    ret
 ;
 ; RV32-LABEL: vscale_masked:
 ; RV32:       # %bb.0: # %entry
@@ -140,6 +176,24 @@ define i64 @vscale_masked() nounwind {
 ; RV32-NEXT:    andi a0, a0, 510
 ; RV32-NEXT:    li a1, 0
 ; RV32-NEXT:    ret
+;
+; RV64-VLEN256MIN-LABEL: vscale_masked:
+; RV64-VLEN256MIN:       # %bb.0: # %entry
+; RV64-VLEN256MIN-NEXT:    csrr a0, vlenb
+; RV64-VLEN256MIN-NEXT:    srli a0, a0, 3
+; RV64-VLEN256MIN-NEXT:    andi a0, a0, 508
+; RV64-VLEN256MIN-NEXT:    ret
+;
+; RV64-VLEN256MAX-LABEL: vscale_masked:
+; RV64-VLEN256MAX:       # %bb.0: # %entry
+; RV64-VLEN256MAX-NEXT:    csrr a0, vlenb
+; RV64-VLEN256MAX-NEXT:    srli a0, a0, 3
+; RV64-VLEN256MAX-NEXT:    ret
+;
+; RV64-VLEN256EXACT-LABEL: vscale_masked:
+; RV64-VLEN256EXACT:       # %bb.0: # %entry
+; RV64-VLEN256EXACT-NEXT:    li a0, 4
+; RV64-VLEN256EXACT-NEXT:    ret
 entry:
   %0 = call i64 @llvm.vscale.i64()
   %1 = and i64 %0, 511