From 618d3bf62fbcfa6ca7f827ad4090143b8535b4a2 Mon Sep 17 00:00:00 2001
From: gingerBill <bill@gingerbill.org>
Date: Sat, 28 May 2022 13:42:58 +0100
Subject: [PATCH] Improve vector comparison `==` `!=` for horizontal reduction

---
 src/llvm_backend_expr.cpp | 37 ++++++++++++++++++++++++++++---------
 1 file changed, 28 insertions(+), 9 deletions(-)

diff --git a/src/llvm_backend_expr.cpp b/src/llvm_backend_expr.cpp
index 55b76b93a..1894e85f6 100644
--- a/src/llvm_backend_expr.cpp
+++ b/src/llvm_backend_expr.cpp
@@ -2585,16 +2585,35 @@ lbValue lb_emit_comp(lbProcedure *p, TokenKind op_kind, lbValue left, lbValue ri
 		}
 		GB_ASSERT_MSG(mask != nullptr, "Unhandled comparison kind %s (%s) %.*s %s (%s)", type_to_string(left.type), type_to_string(base_type(left.type)), LIT(token_strings[op_kind]), type_to_string(right.type), type_to_string(base_type(right.type)));
 
-		// TODO(bill): is this a good approach to dealing with comparisons of vectors?
-		char const *name = "llvm.vector.reduce.umax";
-		LLVMTypeRef types[1] = {LLVMTypeOf(mask)};
-		unsigned id = LLVMLookupIntrinsicID(name, gb_strlen(name));
-		GB_ASSERT_MSG(id != 0, "Unable to find %s.%s", name, LLVMPrintTypeToString(types[0]));
-		LLVMValueRef ip = LLVMGetIntrinsicDeclaration(p->module->mod, id, types, gb_count_of(types));
+		/* NOTE(bill, 2022-05-28):
+			Thanks to Per Vognsen, sign extending <N x i1> to
+			a vector of the same width as the input vector, bit casting to an integer,
+			and then comparing against zero is the better option
+			See: https://lists.llvm.org/pipermail/llvm-dev/2012-September/053046.html
 
-		LLVMValueRef args[1] = {};
-		args[0] = mask;
-		res.value = LLVMBuildCall(p->builder, ip, args, gb_count_of(args), "");
+			// Example assuming 128-bit vector
+
+			%1 = <4 x float> ...
+			%2 = <4 x float> ...
+			%3 = fcmp oeq <4 x float> %1, %2
+			%4 = sext <4 x i1> %3 to <4 x i32>
+			%5 = bitcast <4 x i32> %4 to i128
+			%6 = icmp ne i128 %5, 0
+			br i1 %6, label %true1, label %false2
+
+			This will result in 1 cmpps + 1 ptest + 1 br
+			(even without SSE4.1, contrary to what the mail list states, because of pmovmskb)
+
+		*/
+
+		unsigned count = cast(unsigned)get_array_type_count(a);
+		unsigned elem_sz = cast(unsigned)(type_size_of(elem)*8);
+		LLVMTypeRef mask_type = LLVMVectorType(LLVMIntTypeInContext(p->module->ctx, elem_sz), count);
+		mask = LLVMBuildSExtOrBitCast(p->builder, mask, mask_type, "");
+
+		LLVMTypeRef mask_int_type = LLVMIntTypeInContext(p->module->ctx, cast(unsigned)(8*type_size_of(a)));
+		LLVMValueRef mask_int = LLVMBuildBitCast(p->builder, mask, mask_int_type, "");
+		res.value = LLVMBuildICmp(p->builder, LLVMIntNE, mask_int, LLVMConstNull(LLVMTypeOf(mask_int)), "");
 		return res;
 
 	} else {