Skip to content

[WebAssembly] Ofast pmin/pmax pattern matchers #100107

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 1 commit into from
Jul 24, 2024
Merged

[WebAssembly] Ofast pmin/pmax pattern matchers #100107

merged 1 commit into from
Jul 24, 2024

Conversation

sparker-arm
Copy link
Contributor

With fast-math, the ordered setcc nodes are converted to setcc nodes which do not care about NaNs, so add patterns that use setlt, setle, setgt and setge.

@sparker-arm sparker-arm requested review from tlively and aheejin July 23, 2024 11:30
@llvmbot
Copy link
Member

llvmbot commented Jul 23, 2024

@llvm/pr-subscribers-backend-webassembly

Author: Sam Parker (sparker-arm)

Changes

With fast-math, the ordered setcc nodes are converted to setcc nodes which do not care about NaNs, so add patterns that use setlt, setle, setgt and setge.

Patch is 25.27 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/100107.diff

3 Files Affected:

  • (modified) llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td (+10-2)
  • (modified) llvm/test/CodeGen/WebAssembly/simd-arith.ll (+408)
  • (modified) llvm/test/CodeGen/WebAssembly/vector-reduce.ll (+14-26)
diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
index 2ee430c88169d..d6c6425b10041 100644
--- a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
@@ -1270,7 +1270,11 @@ def pmin : PatFrags<(ops node:$lhs, node:$rhs), [
                     (vselect (setolt $rhs, $lhs), $rhs, $lhs),
                     (vselect (setole $rhs, $lhs), $rhs, $lhs),
                     (vselect (setogt $lhs, $rhs), $rhs, $lhs),
-                    (vselect (setoge $lhs, $rhs), $rhs, $lhs)
+                    (vselect (setoge $lhs, $rhs), $rhs, $lhs),
+                    (vselect (setlt $lhs, $rhs), $lhs, $rhs),
+                    (vselect (setle $lhs, $rhs), $lhs, $rhs),
+                    (vselect (setgt $lhs, $rhs), $rhs, $lhs),
+                    (vselect (setge $lhs, $rhs), $rhs, $lhs)
 ]>;
 defm PMIN : SIMDBinaryFP<pmin, "pmin", 234>;
 
@@ -1279,7 +1283,11 @@ def pmax : PatFrags<(ops node:$lhs, node:$rhs), [
                     (vselect (setogt $rhs, $lhs), $rhs, $lhs),
                     (vselect (setoge $rhs, $lhs), $rhs, $lhs),
                     (vselect (setolt $lhs, $rhs), $rhs, $lhs),
-                    (vselect (setole $lhs, $rhs), $rhs, $lhs)
+                    (vselect (setole $lhs, $rhs), $rhs, $lhs),
+                    (vselect (setgt $lhs, $rhs), $lhs, $rhs),
+                    (vselect (setge $lhs, $rhs), $lhs, $rhs),
+                    (vselect (setlt $lhs, $rhs), $rhs, $lhs),
+                    (vselect (setle $lhs, $rhs), $rhs, $lhs)
 ]>;
 defm PMAX : SIMDBinaryFP<pmax, "pmax", 235>;
 
diff --git a/llvm/test/CodeGen/WebAssembly/simd-arith.ll b/llvm/test/CodeGen/WebAssembly/simd-arith.ll
index 67388b688e3bb..c47762e31ec23 100644
--- a/llvm/test/CodeGen/WebAssembly/simd-arith.ll
+++ b/llvm/test/CodeGen/WebAssembly/simd-arith.ll
@@ -12499,6 +12499,210 @@ define <4 x float> @pmin_v4f32(<4 x float> %x, <4 x float> %y) {
   ret <4 x float> %a
 }
 
+define <4 x float> @pmin_v4f32_fast_olt(<4 x float> %x, <4 x float> %y) {
+; SIMD128-LABEL: pmin_v4f32_fast_olt:
+; SIMD128:         .functype pmin_v4f32_fast_olt (v128, v128) -> (v128)
+; SIMD128-NEXT:  # %bb.0:
+; SIMD128-NEXT:    f32x4.pmin $push0=, $1, $0
+; SIMD128-NEXT:    return $pop0
+;
+; SIMD128-FAST-LABEL: pmin_v4f32_fast_olt:
+; SIMD128-FAST:         .functype pmin_v4f32_fast_olt (v128, v128) -> (v128)
+; SIMD128-FAST-NEXT:  # %bb.0:
+; SIMD128-FAST-NEXT:    f32x4.pmin $push0=, $1, $0
+; SIMD128-FAST-NEXT:    return $pop0
+;
+; NO-SIMD128-LABEL: pmin_v4f32_fast_olt:
+; NO-SIMD128:         .functype pmin_v4f32_fast_olt (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-NEXT:  # %bb.0:
+; NO-SIMD128-NEXT:    f32.lt $push0=, $8, $4
+; NO-SIMD128-NEXT:    f32.select $push1=, $8, $4, $pop0
+; NO-SIMD128-NEXT:    f32.store 12($0), $pop1
+; NO-SIMD128-NEXT:    f32.lt $push2=, $7, $3
+; NO-SIMD128-NEXT:    f32.select $push3=, $7, $3, $pop2
+; NO-SIMD128-NEXT:    f32.store 8($0), $pop3
+; NO-SIMD128-NEXT:    f32.lt $push4=, $6, $2
+; NO-SIMD128-NEXT:    f32.select $push5=, $6, $2, $pop4
+; NO-SIMD128-NEXT:    f32.store 4($0), $pop5
+; NO-SIMD128-NEXT:    f32.lt $push6=, $5, $1
+; NO-SIMD128-NEXT:    f32.select $push7=, $5, $1, $pop6
+; NO-SIMD128-NEXT:    f32.store 0($0), $pop7
+; NO-SIMD128-NEXT:    return
+;
+; NO-SIMD128-FAST-LABEL: pmin_v4f32_fast_olt:
+; NO-SIMD128-FAST:         .functype pmin_v4f32_fast_olt (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-FAST-NEXT:  # %bb.0:
+; NO-SIMD128-FAST-NEXT:    f32.lt $push0=, $5, $1
+; NO-SIMD128-FAST-NEXT:    f32.select $push1=, $5, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    f32.store 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    f32.lt $push2=, $6, $2
+; NO-SIMD128-FAST-NEXT:    f32.select $push3=, $6, $2, $pop2
+; NO-SIMD128-FAST-NEXT:    f32.store 4($0), $pop3
+; NO-SIMD128-FAST-NEXT:    f32.lt $push4=, $7, $3
+; NO-SIMD128-FAST-NEXT:    f32.select $push5=, $7, $3, $pop4
+; NO-SIMD128-FAST-NEXT:    f32.store 8($0), $pop5
+; NO-SIMD128-FAST-NEXT:    f32.lt $push6=, $8, $4
+; NO-SIMD128-FAST-NEXT:    f32.select $push7=, $8, $4, $pop6
+; NO-SIMD128-FAST-NEXT:    f32.store 12($0), $pop7
+; NO-SIMD128-FAST-NEXT:    return
+  %c = fcmp fast olt <4 x float> %y, %x
+  %a = select <4 x i1> %c, <4 x float> %y, <4 x float> %x
+  ret <4 x float> %a
+}
+
+define <4 x float> @pmin_v4f32_fast_ogt(<4 x float> %x, <4 x float> %y) {
+; SIMD128-LABEL: pmin_v4f32_fast_ogt:
+; SIMD128:         .functype pmin_v4f32_fast_ogt (v128, v128) -> (v128)
+; SIMD128-NEXT:  # %bb.0:
+; SIMD128-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-NEXT:    return $pop0
+;
+; SIMD128-FAST-LABEL: pmin_v4f32_fast_ogt:
+; SIMD128-FAST:         .functype pmin_v4f32_fast_ogt (v128, v128) -> (v128)
+; SIMD128-FAST-NEXT:  # %bb.0:
+; SIMD128-FAST-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-FAST-NEXT:    return $pop0
+;
+; NO-SIMD128-LABEL: pmin_v4f32_fast_ogt:
+; NO-SIMD128:         .functype pmin_v4f32_fast_ogt (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-NEXT:  # %bb.0:
+; NO-SIMD128-NEXT:    f32.lt $push0=, $4, $8
+; NO-SIMD128-NEXT:    f32.select $push1=, $8, $4, $pop0
+; NO-SIMD128-NEXT:    f32.store 12($0), $pop1
+; NO-SIMD128-NEXT:    f32.lt $push2=, $3, $7
+; NO-SIMD128-NEXT:    f32.select $push3=, $7, $3, $pop2
+; NO-SIMD128-NEXT:    f32.store 8($0), $pop3
+; NO-SIMD128-NEXT:    f32.lt $push4=, $2, $6
+; NO-SIMD128-NEXT:    f32.select $push5=, $6, $2, $pop4
+; NO-SIMD128-NEXT:    f32.store 4($0), $pop5
+; NO-SIMD128-NEXT:    f32.lt $push6=, $1, $5
+; NO-SIMD128-NEXT:    f32.select $push7=, $5, $1, $pop6
+; NO-SIMD128-NEXT:    f32.store 0($0), $pop7
+; NO-SIMD128-NEXT:    return
+;
+; NO-SIMD128-FAST-LABEL: pmin_v4f32_fast_ogt:
+; NO-SIMD128-FAST:         .functype pmin_v4f32_fast_ogt (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-FAST-NEXT:  # %bb.0:
+; NO-SIMD128-FAST-NEXT:    f32.lt $push0=, $1, $5
+; NO-SIMD128-FAST-NEXT:    f32.select $push1=, $5, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    f32.store 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    f32.lt $push2=, $2, $6
+; NO-SIMD128-FAST-NEXT:    f32.select $push3=, $6, $2, $pop2
+; NO-SIMD128-FAST-NEXT:    f32.store 4($0), $pop3
+; NO-SIMD128-FAST-NEXT:    f32.lt $push4=, $3, $7
+; NO-SIMD128-FAST-NEXT:    f32.select $push5=, $7, $3, $pop4
+; NO-SIMD128-FAST-NEXT:    f32.store 8($0), $pop5
+; NO-SIMD128-FAST-NEXT:    f32.lt $push6=, $4, $8
+; NO-SIMD128-FAST-NEXT:    f32.select $push7=, $8, $4, $pop6
+; NO-SIMD128-FAST-NEXT:    f32.store 12($0), $pop7
+; NO-SIMD128-FAST-NEXT:    return
+  %c = fcmp fast olt <4 x float> %x, %y
+  %a = select <4 x i1> %c, <4 x float> %y, <4 x float> %x
+  ret <4 x float> %a
+}
+
+define <4 x float> @pmin_v4f32_fast_ole(<4 x float> %x, <4 x float> %y) {
+; SIMD128-LABEL: pmin_v4f32_fast_ole:
+; SIMD128:         .functype pmin_v4f32_fast_ole (v128, v128) -> (v128)
+; SIMD128-NEXT:  # %bb.0:
+; SIMD128-NEXT:    f32x4.pmin $push0=, $1, $0
+; SIMD128-NEXT:    return $pop0
+;
+; SIMD128-FAST-LABEL: pmin_v4f32_fast_ole:
+; SIMD128-FAST:         .functype pmin_v4f32_fast_ole (v128, v128) -> (v128)
+; SIMD128-FAST-NEXT:  # %bb.0:
+; SIMD128-FAST-NEXT:    f32x4.pmin $push0=, $1, $0
+; SIMD128-FAST-NEXT:    return $pop0
+;
+; NO-SIMD128-LABEL: pmin_v4f32_fast_ole:
+; NO-SIMD128:         .functype pmin_v4f32_fast_ole (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-NEXT:  # %bb.0:
+; NO-SIMD128-NEXT:    f32.le $push0=, $8, $4
+; NO-SIMD128-NEXT:    f32.select $push1=, $8, $4, $pop0
+; NO-SIMD128-NEXT:    f32.store 12($0), $pop1
+; NO-SIMD128-NEXT:    f32.le $push2=, $7, $3
+; NO-SIMD128-NEXT:    f32.select $push3=, $7, $3, $pop2
+; NO-SIMD128-NEXT:    f32.store 8($0), $pop3
+; NO-SIMD128-NEXT:    f32.le $push4=, $6, $2
+; NO-SIMD128-NEXT:    f32.select $push5=, $6, $2, $pop4
+; NO-SIMD128-NEXT:    f32.store 4($0), $pop5
+; NO-SIMD128-NEXT:    f32.le $push6=, $5, $1
+; NO-SIMD128-NEXT:    f32.select $push7=, $5, $1, $pop6
+; NO-SIMD128-NEXT:    f32.store 0($0), $pop7
+; NO-SIMD128-NEXT:    return
+;
+; NO-SIMD128-FAST-LABEL: pmin_v4f32_fast_ole:
+; NO-SIMD128-FAST:         .functype pmin_v4f32_fast_ole (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-FAST-NEXT:  # %bb.0:
+; NO-SIMD128-FAST-NEXT:    f32.le $push0=, $5, $1
+; NO-SIMD128-FAST-NEXT:    f32.select $push1=, $5, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    f32.store 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    f32.le $push2=, $6, $2
+; NO-SIMD128-FAST-NEXT:    f32.select $push3=, $6, $2, $pop2
+; NO-SIMD128-FAST-NEXT:    f32.store 4($0), $pop3
+; NO-SIMD128-FAST-NEXT:    f32.le $push4=, $7, $3
+; NO-SIMD128-FAST-NEXT:    f32.select $push5=, $7, $3, $pop4
+; NO-SIMD128-FAST-NEXT:    f32.store 8($0), $pop5
+; NO-SIMD128-FAST-NEXT:    f32.le $push6=, $8, $4
+; NO-SIMD128-FAST-NEXT:    f32.select $push7=, $8, $4, $pop6
+; NO-SIMD128-FAST-NEXT:    f32.store 12($0), $pop7
+; NO-SIMD128-FAST-NEXT:    return
+  %c = fcmp fast ole <4 x float> %y, %x
+  %a = select <4 x i1> %c, <4 x float> %y, <4 x float> %x
+  ret <4 x float> %a
+}
+
+define <4 x float> @pmin_v4f32_fast_oge(<4 x float> %x, <4 x float> %y) {
+; SIMD128-LABEL: pmin_v4f32_fast_oge:
+; SIMD128:         .functype pmin_v4f32_fast_oge (v128, v128) -> (v128)
+; SIMD128-NEXT:  # %bb.0:
+; SIMD128-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-NEXT:    return $pop0
+;
+; SIMD128-FAST-LABEL: pmin_v4f32_fast_oge:
+; SIMD128-FAST:         .functype pmin_v4f32_fast_oge (v128, v128) -> (v128)
+; SIMD128-FAST-NEXT:  # %bb.0:
+; SIMD128-FAST-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-FAST-NEXT:    return $pop0
+;
+; NO-SIMD128-LABEL: pmin_v4f32_fast_oge:
+; NO-SIMD128:         .functype pmin_v4f32_fast_oge (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-NEXT:  # %bb.0:
+; NO-SIMD128-NEXT:    f32.le $push0=, $4, $8
+; NO-SIMD128-NEXT:    f32.select $push1=, $8, $4, $pop0
+; NO-SIMD128-NEXT:    f32.store 12($0), $pop1
+; NO-SIMD128-NEXT:    f32.le $push2=, $3, $7
+; NO-SIMD128-NEXT:    f32.select $push3=, $7, $3, $pop2
+; NO-SIMD128-NEXT:    f32.store 8($0), $pop3
+; NO-SIMD128-NEXT:    f32.le $push4=, $2, $6
+; NO-SIMD128-NEXT:    f32.select $push5=, $6, $2, $pop4
+; NO-SIMD128-NEXT:    f32.store 4($0), $pop5
+; NO-SIMD128-NEXT:    f32.le $push6=, $1, $5
+; NO-SIMD128-NEXT:    f32.select $push7=, $5, $1, $pop6
+; NO-SIMD128-NEXT:    f32.store 0($0), $pop7
+; NO-SIMD128-NEXT:    return
+;
+; NO-SIMD128-FAST-LABEL: pmin_v4f32_fast_oge:
+; NO-SIMD128-FAST:         .functype pmin_v4f32_fast_oge (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-FAST-NEXT:  # %bb.0:
+; NO-SIMD128-FAST-NEXT:    f32.le $push0=, $1, $5
+; NO-SIMD128-FAST-NEXT:    f32.select $push1=, $5, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    f32.store 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    f32.le $push2=, $2, $6
+; NO-SIMD128-FAST-NEXT:    f32.select $push3=, $6, $2, $pop2
+; NO-SIMD128-FAST-NEXT:    f32.store 4($0), $pop3
+; NO-SIMD128-FAST-NEXT:    f32.le $push4=, $3, $7
+; NO-SIMD128-FAST-NEXT:    f32.select $push5=, $7, $3, $pop4
+; NO-SIMD128-FAST-NEXT:    f32.store 8($0), $pop5
+; NO-SIMD128-FAST-NEXT:    f32.le $push6=, $4, $8
+; NO-SIMD128-FAST-NEXT:    f32.select $push7=, $8, $4, $pop6
+; NO-SIMD128-FAST-NEXT:    f32.store 12($0), $pop7
+; NO-SIMD128-FAST-NEXT:    return
+  %c = fcmp fast ole <4 x float> %x, %y
+  %a = select <4 x i1> %c, <4 x float> %y, <4 x float> %x
+  ret <4 x float> %a
+}
+
 define <4 x i32> @pmin_int_v4f32(<4 x i32> %x, <4 x i32> %y) {
 ; SIMD128-LABEL: pmin_int_v4f32:
 ; SIMD128:         .functype pmin_int_v4f32 (v128, v128) -> (v128)
@@ -12619,6 +12823,210 @@ define <4 x float> @pmax_v4f32(<4 x float> %x, <4 x float> %y) {
   ret <4 x float> %a
 }
 
+define <4 x float> @pmax_v4f32_fast_ogt(<4 x float> %x, <4 x float> %y) {
+; SIMD128-LABEL: pmax_v4f32_fast_ogt:
+; SIMD128:         .functype pmax_v4f32_fast_ogt (v128, v128) -> (v128)
+; SIMD128-NEXT:  # %bb.0:
+; SIMD128-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-NEXT:    return $pop0
+;
+; SIMD128-FAST-LABEL: pmax_v4f32_fast_ogt:
+; SIMD128-FAST:         .functype pmax_v4f32_fast_ogt (v128, v128) -> (v128)
+; SIMD128-FAST-NEXT:  # %bb.0:
+; SIMD128-FAST-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-FAST-NEXT:    return $pop0
+;
+; NO-SIMD128-LABEL: pmax_v4f32_fast_ogt:
+; NO-SIMD128:         .functype pmax_v4f32_fast_ogt (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-NEXT:  # %bb.0:
+; NO-SIMD128-NEXT:    f32.gt $push0=, $4, $8
+; NO-SIMD128-NEXT:    f32.select $push1=, $4, $8, $pop0
+; NO-SIMD128-NEXT:    f32.store 12($0), $pop1
+; NO-SIMD128-NEXT:    f32.gt $push2=, $3, $7
+; NO-SIMD128-NEXT:    f32.select $push3=, $3, $7, $pop2
+; NO-SIMD128-NEXT:    f32.store 8($0), $pop3
+; NO-SIMD128-NEXT:    f32.gt $push4=, $2, $6
+; NO-SIMD128-NEXT:    f32.select $push5=, $2, $6, $pop4
+; NO-SIMD128-NEXT:    f32.store 4($0), $pop5
+; NO-SIMD128-NEXT:    f32.gt $push6=, $1, $5
+; NO-SIMD128-NEXT:    f32.select $push7=, $1, $5, $pop6
+; NO-SIMD128-NEXT:    f32.store 0($0), $pop7
+; NO-SIMD128-NEXT:    return
+;
+; NO-SIMD128-FAST-LABEL: pmax_v4f32_fast_ogt:
+; NO-SIMD128-FAST:         .functype pmax_v4f32_fast_ogt (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-FAST-NEXT:  # %bb.0:
+; NO-SIMD128-FAST-NEXT:    f32.gt $push0=, $1, $5
+; NO-SIMD128-FAST-NEXT:    f32.select $push1=, $1, $5, $pop0
+; NO-SIMD128-FAST-NEXT:    f32.store 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    f32.gt $push2=, $2, $6
+; NO-SIMD128-FAST-NEXT:    f32.select $push3=, $2, $6, $pop2
+; NO-SIMD128-FAST-NEXT:    f32.store 4($0), $pop3
+; NO-SIMD128-FAST-NEXT:    f32.gt $push4=, $3, $7
+; NO-SIMD128-FAST-NEXT:    f32.select $push5=, $3, $7, $pop4
+; NO-SIMD128-FAST-NEXT:    f32.store 8($0), $pop5
+; NO-SIMD128-FAST-NEXT:    f32.gt $push6=, $4, $8
+; NO-SIMD128-FAST-NEXT:    f32.select $push7=, $4, $8, $pop6
+; NO-SIMD128-FAST-NEXT:    f32.store 12($0), $pop7
+; NO-SIMD128-FAST-NEXT:    return
+  %c = fcmp fast ogt <4 x float> %x, %y
+  %a = select <4 x i1> %c, <4 x float> %x, <4 x float> %y
+  ret <4 x float> %a
+}
+
+define <4 x float> @pmax_v4f32_fast_olt(<4 x float> %x, <4 x float> %y) {
+; SIMD128-LABEL: pmax_v4f32_fast_olt:
+; SIMD128:         .functype pmax_v4f32_fast_olt (v128, v128) -> (v128)
+; SIMD128-NEXT:  # %bb.0:
+; SIMD128-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-NEXT:    return $pop0
+;
+; SIMD128-FAST-LABEL: pmax_v4f32_fast_olt:
+; SIMD128-FAST:         .functype pmax_v4f32_fast_olt (v128, v128) -> (v128)
+; SIMD128-FAST-NEXT:  # %bb.0:
+; SIMD128-FAST-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-FAST-NEXT:    return $pop0
+;
+; NO-SIMD128-LABEL: pmax_v4f32_fast_olt:
+; NO-SIMD128:         .functype pmax_v4f32_fast_olt (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-NEXT:  # %bb.0:
+; NO-SIMD128-NEXT:    f32.lt $push0=, $4, $8
+; NO-SIMD128-NEXT:    f32.select $push1=, $8, $4, $pop0
+; NO-SIMD128-NEXT:    f32.store 12($0), $pop1
+; NO-SIMD128-NEXT:    f32.lt $push2=, $3, $7
+; NO-SIMD128-NEXT:    f32.select $push3=, $7, $3, $pop2
+; NO-SIMD128-NEXT:    f32.store 8($0), $pop3
+; NO-SIMD128-NEXT:    f32.lt $push4=, $2, $6
+; NO-SIMD128-NEXT:    f32.select $push5=, $6, $2, $pop4
+; NO-SIMD128-NEXT:    f32.store 4($0), $pop5
+; NO-SIMD128-NEXT:    f32.lt $push6=, $1, $5
+; NO-SIMD128-NEXT:    f32.select $push7=, $5, $1, $pop6
+; NO-SIMD128-NEXT:    f32.store 0($0), $pop7
+; NO-SIMD128-NEXT:    return
+;
+; NO-SIMD128-FAST-LABEL: pmax_v4f32_fast_olt:
+; NO-SIMD128-FAST:         .functype pmax_v4f32_fast_olt (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-FAST-NEXT:  # %bb.0:
+; NO-SIMD128-FAST-NEXT:    f32.lt $push0=, $1, $5
+; NO-SIMD128-FAST-NEXT:    f32.select $push1=, $5, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    f32.store 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    f32.lt $push2=, $2, $6
+; NO-SIMD128-FAST-NEXT:    f32.select $push3=, $6, $2, $pop2
+; NO-SIMD128-FAST-NEXT:    f32.store 4($0), $pop3
+; NO-SIMD128-FAST-NEXT:    f32.lt $push4=, $3, $7
+; NO-SIMD128-FAST-NEXT:    f32.select $push5=, $7, $3, $pop4
+; NO-SIMD128-FAST-NEXT:    f32.store 8($0), $pop5
+; NO-SIMD128-FAST-NEXT:    f32.lt $push6=, $4, $8
+; NO-SIMD128-FAST-NEXT:    f32.select $push7=, $8, $4, $pop6
+; NO-SIMD128-FAST-NEXT:    f32.store 12($0), $pop7
+; NO-SIMD128-FAST-NEXT:    return
+  %c = fcmp fast olt <4 x float> %x, %y
+  %a = select <4 x i1> %c, <4 x float> %y, <4 x float> %x
+  ret <4 x float> %a
+}
+
+define <4 x float> @pmax_v4f32_fast_oge(<4 x float> %x, <4 x float> %y) {
+; SIMD128-LABEL: pmax_v4f32_fast_oge:
+; SIMD128:         .functype pmax_v4f32_fast_oge (v128, v128) -> (v128)
+; SIMD128-NEXT:  # %bb.0:
+; SIMD128-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-NEXT:    return $pop0
+;
+; SIMD128-FAST-LABEL: pmax_v4f32_fast_oge:
+; SIMD128-FAST:         .functype pmax_v4f32_fast_oge (v128, v128) -> (v128)
+; SIMD128-FAST-NEXT:  # %bb.0:
+; SIMD128-FAST-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-FAST-NEXT:    return $pop0
+;
+; NO-SIMD128-LABEL: pmax_v4f32_fast_oge:
+; NO-SIMD128:         .functype pmax_v4f32_fast_oge (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-NEXT:  # %bb.0:
+; NO-SIMD128-NEXT:    f32.ge $push0=, $4, $8
+; NO-SIMD128-NEXT:    f32.select $push1=, $4, $8, $pop0
+; NO-SIMD128-NEXT:    f32.store 12($0), $pop1
+; NO-SIMD128-NEXT:    f32.ge $push2=, $3, $7
+; NO-SIMD128-NEXT:    f32.select $push3=, $3, $7, $pop2
+; NO-SIMD128-NEXT:    f32.store 8($0), $pop3
+; NO-SIMD128-NEXT:    f32.ge $push4=, $2, $6
+; NO-SIMD128-NEXT:    f32.select $push5=, $2, $6, $pop4
+; NO-SIMD128-NEXT:    f32.store 4($0), $pop5
+; NO-SIMD128-NEXT:    f32.ge $push6=, $1, $5
+; NO-SIMD128-NEXT:    f32.select $push7=, $1, $5, $pop6
+; NO-SIMD128-NEXT:    f32.store 0($0), $pop7
+; NO-SIMD128-NEXT:    return
+;
+; NO-SIMD128-FAST-LABEL: pmax_v4f32_fast_oge:
+; NO-SIMD128-FAST:         .functype pmax_v4f32_fast_oge (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-FAST-NEXT:  # %bb.0:
+; NO-SIMD128-FAST-NEXT:    f32.ge $push0=, $1, $5
+; NO-SIMD128-FAST-NEXT:    f32.select $push1=, $1, $5, $pop0
+; NO-SIMD128-FAST-NEXT:    f32.store 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    f32.ge $push2=, $2, $6
+; NO-SIMD128-FAST-NEXT:    f32.select $push3=, $2, $6, $pop2
+; NO-SIMD128-FAST-NEXT:    f32.store 4($0), $pop3
+; NO-SIMD128-FAST-NEXT:    f32.ge $push4=, $3, $7
+; NO-SIMD128-FAST-NEXT:    f32.select $push5=, $3, $7, $pop4
+; NO-SIMD128-FAST-NEXT:    f32.store 8($0), $pop5
+; NO-SIMD128-FAST-NEXT:    f32.ge $push6=, $4, $8
+; NO-SIMD128-FAST-NEXT:    f32.select $push7=, $4, $8, $pop6
+; NO-SIMD128-FAST-NEXT:    f32.store 12($0), $pop7
+; NO-SIMD128-FAST-NEXT:    return
+  %c = fcmp fast oge <4 x float> %x, %y
+  %a = select <4 x i1> %c, <4 x float> %x, <4 x float> %y
+  ret <4 x float> %a
+}
+
+define <4 x float> @pmax_v4f32_fast_ole(<4 x float> %x, <4 x float> %y) {
+; SIMD128-LABEL: pmax_v4f32_fast_ole:
+; SIMD128:         .functype pmax_v4f32_fast_ole (v128, v128) -> (v128)
+; SIMD128-NEXT:  # %bb.0:
+; SIMD128-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-NEXT:    return $pop0
+;
+; SIMD128-FAST-LABEL: pmax_v4f32_fast_ole:
+; SIMD128-FAST:         .functype pmax_v4f32_fast_ole (v128, v128) -> (v128)
+; SIMD128-FAST-NEXT:  # %bb.0:
+; SIMD128-FAST-NEXT:    f32x4.pmax $push0=, $0, $1
+; SIMD128-FAST-NEXT:    return $pop0
+;
+; NO-SIMD128-LABEL: pmax_v4f32_fast_ole:
+; NO-SIMD128:         .functype pmax_v4f32_fast_ole (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-NEXT:  # %bb.0:
+; NO-SIMD128-NEXT:    f32.le $push0=, $4, $8
+; NO-SIMD128-NEXT:    f32.select $push1=, $8, $4, $pop0
+; NO-SIMD128-NEXT:    f32.store 12($0), $pop1
+; NO-SIMD128-NEXT:    f32.le $push2=, $3, $7
+; NO-SIMD128-NEXT:    f32.select $push3=, $7, $3, $pop2
+; NO-SIMD128-NEXT:    f32.store 8($0), $pop3
+; NO-SIMD128-NEXT:    f32.le $push4=, $2, $6
+; NO-SIMD128-NEXT:    f32.select $push5=, $6, $2, $pop4
+; NO-SIMD128-NEXT:    f32.store 4($0), $pop5
+; NO-SIMD128-NEXT:    f32.le $push6=, $1, $5
+; NO-SIMD128-NEXT:    f32.select $push7=, $5, $1, $pop6
+; NO-SIMD128-NEXT:    f32.store 0($0), $pop7
+; NO-SIMD128-NEXT:    return
+;
+; NO-SIMD128-FAST-LABEL: pmax_v4f32_fast_ole:
+; NO-SIMD128-FAST:         .functype pmax_v4f32_fast_ole (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
+; NO-SIMD128-FAST-NEXT:  # %bb.0:
+; NO-SIMD128-FAST-NEXT:    f32.le $push0=, $1, $5
+; NO-SIMD128-FAST-NEXT:    f32.sele...
[truncated]

tlively
tlively approved these changes Jul 23, 2024
View reviewed changes
Copy link
Collaborator

@tlively tlively left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nice, thanks!

@sparker-arm
[WebAssembly] Ofast pmin/pmax pattern matchers
21d2980 
With fast-math, the ordered setcc nodes are converted to setcc nodes
which do not care about NaNs, so add patterns that use setlt, setle,
setgt and setge.
@sparker-arm
Copy link
Contributor Author

Cheers, I just needed to fix a couple of typos before merging.

@sparker-arm sparker-arm merged commit a3de21c into llvm:main Jul 24, 2024
yuxuanchen1997 pushed a commit that referenced this pull request Jul 25, 2024
@sparker-arm @yuxuanchen1997
[WebAssembly] Ofast pmin/pmax pattern matchers (#100107)
9e1d538 
Summary:
With fast-math, the ordered setcc nodes are converted to setcc nodes
which do not care about NaNs, so add patterns that use setlt, setle,
setgt and setge.

Test Plan: 

Reviewers: 

Subscribers: 

Tasks: 

Tags: 


Differential Revision: https://phabricator.intern.facebook.com/D60250651
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@tlively tlively tlively approved these changes

@aheejin aheejin Awaiting requested review from aheejin

Assignees
No one assigned
Labels
backend:WebAssembly
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@sparker-arm @llvmbot @tlively