Merge pull request #156 from JuliaGPU/tb/algebraic_simplfications

Add algebraic simplifications to get rid of 1-based index IR bloat

Author: maleadt

src/compiler/codegen/passes/pipeline.jl

@@ -111,6 +111,31 @@ const FMA_RULES = RewriteRule[
 
 fma_fusion_pass!(sci::StructuredIRCode) = rewrite_patterns!(sci, FMA_RULES)
 
+#=============================================================================
+ Algebraic Simplification (rewrite)
+=============================================================================#
+
+# Cancel inverse addi/subi pairs: x+c-c → x, x-c+c → x.
+# Repeated ~c binds enforce that both operands are the same value.
+
+const ALGEBRA_RULES = RewriteRule[
+    @rewrite Intrinsics.subi(Intrinsics.addi(~x, ~c), ~c) => ~x
+    @rewrite Intrinsics.addi(Intrinsics.subi(~x, ~c), ~c) => ~x
+]
+
+algebra_pass!(sci::StructuredIRCode) = rewrite_patterns!(sci, ALGEBRA_RULES)
+
+#=============================================================================
+ Combined Rule Set
+=============================================================================#
+
+const ALL_REWRITE_RULES = RewriteRule[
+    NORMALIZE_RULES...,
+    ALGEBRA_RULES...,
+    SVE_RULES...,
+    FMA_RULES...,
+]
+
 #=============================================================================
  Pass Pipeline
 =============================================================================#
@@ -122,10 +147,8 @@ Run the full pass pipeline on a StructuredIRCode. Called for both kernel
 and subprogram compilation.
 """
 function run_passes!(sci::StructuredIRCode)
-    # Rewrite passes (order matters: normalize before optimize, SVE before FMA)
-    normalize_pass!(sci)
-    scalar_view_elim_pass!(sci)
-    fma_fusion_pass!(sci)
+    # All rewrite rules in one fixpoint pass
+    rewrite_patterns!(sci, ALL_REWRITE_RULES)
 
     # Memory ordering
     alias_result = alias_analysis_pass!(sci)